3 years ago, a question you could not answer had a price attached to it.

You searched the internal wiki and found a page last edited in 2021 by someone who had left. You asked the two people most likely to know and got a shrug and a maybe. You then had a choice: drop it, or spend a day working it out. If you worked it out, you almost always wrote it down, not because the company had a documentation culture, but because you had just paid for that answer and had no intention of paying twice.

That cycle, repeated a few thousand times a year by a few hundred people, is how an institution accumulates knowledge. It is not elegant. It is driven almost entirely by irritation.

Today, the same question takes 4 seconds and produces a confident, well-structured, largely correct answer. Nothing is searched, nobody is asked, nothing is written down, and — crucially — nobody learns that the organisation had never worked it out in the first place.

The answer was excellent. The transaction that used to make the organisation smarter did not occur.

This is the shape of the thing I want to describe: not a failure, not a loss, but a quiet cessation of accumulation, taking place inside a system that has never felt better informed in its life.

What happens to a person

Start at the individual scale, where it has actually been measured.

In June 2025, a team at the MIT Media Lab, led by Nataliya Kosmyna, published a study with the deliberately provocative title Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. 54 participants wrote essays across 4 months under 3 conditions: with an LLM, with a search engine, or with nothing but their own head. The researchers ran EEG throughout and tested recall afterwards.

The essays produced by the LLM group were fine. That matters, because this is not a story about bad output.

The recall was not fine. More than 80% of the LLM users could not quote a single sentence from an essay they had finished minutes earlier.

Asked to reproduce their own text after a delay, the LLM group managed around 17%; the unassisted group managed around 46%.

The EEG showed weaker connectivity across the networks associated with the effortful integration of new material. And the participants reported something subtler and more revealing: a diminished sense of authorship. They did not quite feel the essays were theirs.

They were right. In the only sense that matters for memory, the essays were not theirs. Memory is not a recording device that captures whatever passes in front of it. Encoding is a cost. It is paid in effort: in retrieval attempts, in failed sentences, in the specific discomfort of holding a half-formed idea while trying to make it a whole one. Remove the effort and you do not get the memory for free. You get the artefact without the memory.

The study is a preprint and has not yet been peer reviewed, and it deserves the usual caution.

But the finding it points at is not exotic. It is one of the most robust results in cognitive psychology, arriving in new clothes: we remember what we generate far better than what we are given.

Now stop thinking about one person and one essay. Think about 10,000 people, 3 years, and every decision, diagnosis, judgement and explanation those people produce.

What an organisation is actually made of

Institutions have always had a memory problem, and it has always been worse than anyone admits, because most of what an organisation knows has never been written anywhere.

In 1996, the anthropologist Julian Orr published Talking About Machines, an ethnography of Xerox photocopier technicians that has quietly influenced a generation of thinking about work. Orr rode along with technicians for months, and his central finding was awkward for everyone who had commissioned it: the knowledge that made these people good at the job was essentially absent from the service manuals they had been given. It lived in stories, told at breakfast, in car parks, over the phone. This machine in this office with this humidity. That fault which presents as a paper jam and is always a worn sensor. The customer who will describe the symptom wrongly, in a specific and predictable way.

“The circulation of stories among the technicians is the principal means by which they stay informed of the developing subtleties of machine behaviour.”

— Julian E. Orr, Talking About Machines: An Ethnography of a Modern Job, 1996

Xerox’s response has become a small legend in knowledge management: they gave the technicians two-way radios, so the stories could keep circulating without everyone having to be in the same room at the same time. The intervention was not to write the knowledge down. It was to protect the channel through which it moved.

The expensive version of the same lesson belongs to NASA. The F-1 engine that lifted Saturn V off the pad was, by any measure, one of the most thoroughly documented machines ever built; Rocketdyne ran a dedicated production knowledge retention programme that produced 20 volumes covering injector rings, valves, assembly, checkout, thermal insulation, cabling. Everything a reasonable person would think you needed.

When engineers returned to the F-1 decades later, the 20 volumes were not enough. Each engine had been effectively hand-built, with its own undocumented quirks. Welding and fabrication techniques lived in the hands of specific people, and those people had retired or died. The drawings had been produced under crushing deadline pressure without computer aids, and a drawing made that way records the shape of the thing but not the reasoning, the failed alternative, or the trick that made it work.

The blueprints survived. The knowledge did not. And nobody at NASA was careless: they had done everything the textbook says to do.

Both stories converge on the same uncomfortable point. The valuable part of institutional knowledge was never the record. It was the person who could tell you why the record was wrong.

The absence that did all the work

Here is the mechanism I think we have broken, and it is worth being precise about it, because it is easy to mistake for nostalgia.

Organisations do not learn because they are curious. They learn because they periodically run into the wall of their own ignorance, and the collision hurts enough to produce a response.

The collision had a very specific form: you asked, and nothing came back. That silence was not an inconvenience. It was a measurement. It told you, with total reliability, that this was something the organisation had never worked out — as distinct from something it had worked out and filed somewhere you had not looked. Those are entirely different states, and the old system distinguished between them for free.

Almost everything good followed from that distinction. It told you where the frontier was. It told you which questions were worth someone’s afternoon. And it created the only reliable motive for documentation ever observed in the wild, which is not diligence but resentment: I am never doing this again.

AI-mediated knowledge access removes the silence. Every question now returns a fluent, immediate, structurally impeccable answer, and the answer looks exactly the same whether it was retrieved from something the organisation genuinely knows or reconstructed on the spot from the general shape of the world.

This is not speculation. Anna Gausen, Bhaskar Mitra and Siân Lindley, writing at ACM FAccT in 2024, examined how AI-mediated enterprise knowledge access systems reshape organisational memory.

Among their findings: these systems increase the visibility of knowledge that already exists while simultaneously diminishing members’ organic awareness of knowledge gaps. Workers stop developing an accurate internal map of what the organisation does not know, because they no longer encounter its edges.

An organisation that cannot detect its own ignorance does not become ignorant dramatically. It simply stops filling anything in, and the gaps persist indefinitely, invisibly, being papered over on demand, 4 seconds at a time.

Four ways an organisation leaks

The failure is not one thing. It is four distinct leaks, and they compound, because each one hides the next.

• The Vanished Gap. Knowledge that never gets created, because the absence that would have prompted its creation is no longer perceptible. Nobody decides not to investigate; the question simply never presents itself as open. This is the deepest leak and the only one that is genuinely invisible, because its symptom is the non-occurrence of an event. You cannot audit for work that was never triggered. The tell, when it eventually surfaces, is a company that has been operating for 3 years on an assumption nobody ever tested, defended by a confident answer that has been given 400 times.

• The Ephemeral Answer. Knowledge that is created but has nowhere to land. The reasoning happens inside a conversation with a model, genuinely good reasoning, often better than what would have happened otherwise…and then the tab closes. It is not written down, because writing it down feels redundant when the thing that produced it is always available. But the model does not remember your organisation. It remembers nothing about the decision, the constraint, or the four options you rejected. The next person asks the same question and gets a different plausible answer, and neither of them will ever know they disagreed. Institutional memory is being replaced by a per-session hallucination of institutional memory.

• The Hollow Fluency. Knowledge that is produced and delivered by a person who has not encoded it. This is the MIT finding at organisational scale. The output is real and often correct; the person cannot answer a follow-up question, cannot spot when the situation has changed enough to invalidate the answer, and cannot teach it to anyone. A workforce in this state looks extremely capable in every artefact it produces and extremely thin in every unscripted conversation. It is also, quietly, unable to train its own successors, because you cannot transmit what you never held.

• The Severed Inheritance. Knowledge that used to move between people and now does not, because the channel has been rerouted. Orr’s whole point was that the stories were the transmission mechanism. When a junior colleague asks a senior one a question, the answer is the smallest part of what transfers; they also acquire the reasoning, the hesitation, the aside about the time this went wrong in 2019, and, not trivially, a relationship that makes the next question possible. Asking a model returns the answer with none of the freight. It is more efficient in precisely the way that a wire transfer is more efficient than a conversation about money.

The four compound in a specific order. The Severed Inheritance dries up the stories. The Hollow Fluency means nobody notices, because everyone still sounds competent. The Ephemeral Answer means nothing new is being written to replace what is being lost. And the Vanished Gap means no alarm ever sounds, because the organisation has lost the instrument that used to detect absence.

Where the leak becomes visible

If all of this were purely invisible, it would be hard to argue about. But there is one place the water shows on the ceiling, and it is worth looking at closely.

In July 2025, METR ran a randomised controlled trial with 16 experienced open-source developers, covering 246 real tasks in mature repositories on which the developers had an average of 5 years of prior experience. The developers expected AI tools to speed them up by 24%. Afterwards, they reported having been sped up by about 20%. Measured, they were 19% slower.

The perception gap, nearly 40 points, is the part that gets quoted. The mechanism is the part that matters here. These were extraordinarily high-context environments. The developers carried, in their heads, years of accumulated knowledge about why this module is structured strangely, which tests lie, what the maintainer will reject in review. The model had none of that and no way to get it. So the developers spent their time explaining their own tacit knowledge to a system that could not retain it, then reviewing output that did not reflect it.

The productivity finding is interesting. The structural finding is more important: the slowdown was largest exactly where human accumulated knowledge was deepest. Which tells you something uncomfortable about the incentive gradient.

In low-context work, where nobody has accumulated anything, the tools are transformative. In high-context work, they are a drag, and the way to make them stop being a drag is to lower the context. To standardise, flatten, document less idiosyncratically, and gradually stop building the deep specific knowledge that made the environment hard for the model in the first place.

Nobody will announce that decision. It will arrive as a series of sensible efficiency choices, each of which is defensible, and at the end of it the organisation will be perfectly legible to its tools and will contain nothing that took 15 years to learn.

How to hold water

The instinct at this point is to reach for documentation policy, and the instinct is wrong, or at least badly insufficient. NASA had 20 volumes. Documentation captures what was decided. The leak is in why, and in who could tell you the record is out of date.

• If you lead an organisation: restore the ability to see absence. This is the single highest-leverage intervention available, and it is mostly a matter of instrumentation and honesty. Any internal AI system that answers from your own material should be able to tell the difference between “this is grounded in something we wrote and here it is” and “this is a reasonable reconstruction,” and it should say so, visibly, every time. That distinction is the modern equivalent of the empty wiki search, and if your system does not expose it, you have bought a machine whose primary function is to conceal the state of your own knowledge from you. Then measure the right thing. Every organisation now measures output, and output has never been less informative. Ask instead: what did we learn last quarter, where is it written, and who wrote it? If nobody can answer without opening a chat history, you do not have institutional memory.

• If you manage a team: protect the channel, not the archive. Xerox’s answer to Orr was radios, and it was correct. The equivalent today is deliberately preserving the conditions in which people tell each other things: the review conversation that could have been a comment, the pairing session, the fifteen minutes at the end of an incident where somebody says what actually happened rather than what the report will say. These look like inefficiency and they are the transmission mechanism. And make the reasoning the deliverable. A decision record that says what was chosen is worth a fraction of one that says what was rejected and why, because the rejected options are the part a model will confidently reinvent for you next year.

• If you are doing the work: understand that recall has become a choice, and that it now has to be made deliberately, because the default is no longer to remember. Pick two or three domains where you intend to be genuinely deep, and refuse the shortcut there: do it the slow way, generate the answer before you check it, be willing to be slower for a year. Be fast everywhere else, without guilt. And write down not what you concluded, but what you tried that failed. That is the part nobody else has, the part no model can reconstruct, and the part that will still be worth something when the answer to every well-posed question is free.

• If you are early in your career: you are entering an environment that will let you produce a competent answer to almost any question, and remember almost none of it. That is an extraordinary offer and it is a trap, because in 10 years your value will be almost entirely the small set of things you can speak about with nothing open in front of you. Also: keep asking senior people questions you could ask a model. Not for the answer. For the 20 minutes around the answer, which is where the actual transfer happens and which is the thing quietly disappearing from professional life.

The uncomfortable truth

The prevailing story is that AI makes organisations smarter. Every measurement available supports it: questions answered, time to answer, coverage, satisfaction. On any dashboard you could build, an organisation with good AI-mediated knowledge access is dramatically better informed than the same organisation was three years ago.

That story is true and it is describing the wrong quantity. It is measuring access, not accumulation. Retrieval, not memory. And the two feel identical from inside, which is precisely why this can run for years without anyone raising it.

The distinction is not academic. A person with excellent access and no memory can answer any question and cannot notice that a question has stopped making sense. An organisation in the same state can serve every request and cannot tell that the ground beneath one of its oldest assumptions has moved, because the mechanism that used to report ground movement was the failure to find an answer, and it no longer fails.

Ruth Schwartz Cowan’s washing machines did not reduce housework because the standard rose. Jevons’ efficient engines did not reduce coal because demand rose. This one is stranger than either, because nothing rose. Something simply stopped: the small, irritating, unglamorous loop by which a company converts a moment of not-knowing into something it permanently knows.

Nobody switched it off. It just stopped being triggered, because the condition that triggered it — silence — no longer occurs anywhere in the building.

So ask your organisation what it learnt last year. Not what it produced, shipped, answered or automated. What it learnt, where that learning is written, and whose name is on it.

If the honest answer is that it knows a great deal more than it did and has learned nothing at all, that is not a paradox. That is the leak, and it has been running the whole time, at exactly the speed of your fastest answer.

In March, an operations analyst discovers that the weekly reporting pack — eleven hours of his life, every week, for six years — can be assembled in four.

He does not tell anyone. This is not deviousness; it is instinct, and the instinct is correct. He spends the first genuinely empty Friday afternoon of his professional life reading about something entirely unrelated to logistics, and feels, for about three weeks, that a small and private justice has been done.

In April, someone notices the pack has been arriving on Thursday rather than Friday. Nothing is said. In May, a colleague in another team refers, in passing, to “when the pack lands on Thursday,” and the sentence goes unchallenged, because it is now simply a description of reality. By June, Thursday is the deadline. Nobody moved it. No one held a meeting about it, no policy was amended, no manager made a decision they would recognise as a decision. The deadline moved the way a shoreline moves.

Then, because the team is now visibly capable of more, a second pack appears — the regional breakdown, discussed for years and always deemed impossible. It is not impossible any more. It is a genuine improvement to the business, everyone is pleased, and the analyst builds the template.

By July he is working eleven hours a week on reporting again.

Ask him what happened and he will describe a series of reasonable events, each of which he would defend individually. He is not overworked because of a villain. He is overworked because of a ratchet, and the defining property of a ratchet is that it turns one way, quietly, under load, with a mechanism so ordinary that nobody thinks to look at it.

What the eight-month study actually found

In February 2026, the Harvard Business Review published research by Aruna Ranganathan and Xingqi Maggie Ye who had followed roughly 200 employees at a US technology company for eight months as generative AI worked its way through their jobs. It is one of the few studies of the period that watched the same people for long enough to see the second-order effects rather than the launch-week enthusiasm.

The tools worked. That is worth stating plainly, because the finding is not a story about disappointing technology. People moved faster. They handled more kinds of task. By any reasonable measure of individual capability, the intervention succeeded.

And 83% of them said their workload had increased.

The researchers identified three distinct mechanisms, and it is worth sitting with each, because none of them look like a problem while they are happening.

Task expansion. Work that had previously been out of reach became accessible, so people reached for it. Not because they were instructed to, but because the barrier that used to make it a non-question had gone. The regional breakdown was impossible for six years; the moment it stopped being impossible, it stopped being optional, without anyone ever converting it from one to the other.

Blurred boundaries. Work began leaking into the interstitial spaces that used to be breaks. If a thing takes ninety seconds instead of twenty minutes, it can be done in a lift, in a queue, at 22:40, between two other things. Each individual instance is trivially small, which is precisely why none of them ever gets refused. The working day did not get extended by decision. It got extended by capillary action.

Multitasking. With execution faster, people ran more threads simultaneously. This is the most insidious of the three, because sustained partial attention produces a strong internal sensation of productivity while degrading the quality of thought — and the sensation is what people report, while the degradation is what shows up three months later as fatigue nobody can source.

The common feature of all three: nobody chose them. There was no moment at which an organisation resolved to spend its AI dividend on intensification rather than on quality, or margin, or rest. In the absence of a decision, the system defaulted to more. Systems generally do.

The oldest pattern in the history of technology

If this feels novel, that is only because we are inside it. It is one of the most thoroughly documented patterns in economic and social history, and it has been observed at least twice with total clarity by people nobody listened to.

In 1865, the English economist William Stanley Jevons published The Coal Question, addressing a debate about whether Britain’s coal reserves would last. The prevailing view was reassuring: engines were becoming dramatically more efficient, therefore less coal would be needed. Jevons demolished it in a sentence that has since been named after him.

“It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.”

— William Stanley Jevons, The Coal Question, 1865

His logic was simple and inescapable. Efficiency reduces the cost of using something. Reducing the cost of something increases the demand for it. Watt’s improved engine did not reduce Britain’s coal consumption; it made steam power cheap enough to put in places it had never been, and consumption rose by an order of magnitude. Efficiency was not the brake. It was the accelerator.

A century later, the historian Ruth Schwartz Cowan asked the domestic version of the same question. Her 1983 book More Work for Mother traced household technology from the open hearth to the microwave and confronted a fact that ought to have been a scandal: after washing machines, vacuum cleaners, running hot water, gas ovens, commercial flour and refrigeration, American women were spending about as many hours on housework as their great-grandmothers had.

Cowan’s explanation was not that the appliances were bad. They were transformative. Several things happened at once. Work previously done by servants, husbands and children was quietly absorbed by one person, because the machines made it a one-person job. And — the part that matters here — the standard rose to meet the new capability. Weekly laundry became daily laundry. Acceptable cleanliness became a moving target that tracked, almost exactly, the capability of the equipment available to achieve it. The washing machine did not buy time. It bought a higher definition of clean.

That is the mechanism we are living in now, running at software speed. Cowan’s ratchet took decades to turn. Ours turns in about a quarter.

Two accurate reports of the same event

The strangest artefact of this period is a pair of statistics that appear to contradict each other and do not.

The Upwork Research Institute surveyed 2,500 executives, employees and freelancers across four countries and found that 96% of C-suite leaders expected AI to raise productivity — while 77% of employees actually using the tools said those tools had added to their workload. A further 47% said they had no idea how to deliver the productivity gains being expected of them.

The standard reading of this is a communication gap: leadership doesn’t understand the coalface, someone should run a listening tour. That reading is comfortable and it is wrong, because it assumes one of the two groups is mistaken.

Neither is. They are describing the same event from opposite ends of the ratchet. At the organisational level, output per person genuinely rose — the executives are reading their instruments correctly. At the individual level, the work genuinely became heavier — the employees are reading theirs correctly too. Both measurements are accurate. The gain and the cost simply landed on different people, and there is no single vantage point inside the organisation from which both are visible at once.

The Harvard data shows exactly where each landed. Burnout was reported by 62% of associates and 61% of entry-level workers, against 38% of C-suite leaders. The people who set the pace are not the people who run at it. This is not a moral failing on the part of leadership; it is a structural feature of how the information travels. Throughput is legible upward. Strain is not. A dashboard will show you a rising line months before it shows you what the line cost, and by then the line is the baseline and the baseline is not negotiable.

What actually gets consumed

It is tempting to describe all this as a story about time, but time is not really the resource being eaten. The resource being eaten is slack.

Slack is the loose, uncommitted, faintly embarrassing space inside a working system. The gap between finishing one thing and starting the next. The Thursday afternoon with nothing due. The meeting that ended early. The conversation in the corridor that had no agenda and produced, six weeks later, the only good idea of the quarter.

On any dashboard, slack is indistinguishable from waste. That is its central tragedy. It has no output, no ticket number, and no defender in a budget meeting. Which makes it the very first thing an efficiency programme finds, and the very last thing anyone thinks to protect.

But slack is where several things happen that happen nowhere else. It is where people learn — not through training, but through the unhurried repetition and reflection that turns procedure into judgement. It is where people notice: the mis-stated assumption, the number that looks slightly wrong, the client who has gone quiet. Noticing requires spare attention, and spare attention is the first casualty of running three threads at once. It is where people recover, and a system with no recovery capacity is not efficient but brittle, in the specific engineering sense that it has no ability to absorb a shock without deforming.

And it is where people ask whether the work is the right work. This is the most valuable function of slack and the easiest to lose, because the question only occurs to people who have a moment in which nothing is due. Fill every moment and you get an organisation that executes with tremendous velocity and never once asks where it is going.

Here is the uncomfortable arithmetic. AI creates a large, visible, easily measured productivity gain. Slack is a large, invisible, entirely unmeasured buffer. When the gain arrives, the pressure to convert it into output falls first on the part of the system that cannot demonstrate its own value. The buffer is consumed before anyone realises it was an asset, and its absence is only detectable later, by the strange fact that nobody in the organisation seems to be learning anything or catching anything early.

The four positions on the ratchet

The pattern presents differently depending on where you are standing. Four positions, and almost everyone is in one of them right now.

• The Absorber. Discovers a genuine efficiency gain and quietly reinvests all of it in more output, telling no one. The motives are usually decent: pride in the work, a wish to be useful, or a well-founded suspicion that admitting to spare capacity is professionally unwise. The effect is to set a new baseline for everyone, including themselves, on the basis of a capability they never disclosed. Absorbers are rewarded exactly once, and measured against the new level permanently. The tell is that they cannot explain how their workload grew, only that it did — and that when they eventually slow down, it registers not as a return to normal but as a decline.

• The Sprawler. Responds to lowered barriers by expanding scope in every direction. Everything now feels doable, so everything gets attempted: the second report, the side analysis, the redesign nobody asked for. The work is real and often good. But scope acquired because it is possible, rather than because it is important, has no natural limit, and the Sprawler ends up carrying nine responsibilities at a level of attention appropriate to three. Their year-end review says “broad impact.” Their calendar says something else. The Sprawler’s error is treating capability as an instruction.

• The Leaker. Never took on extra work and never agreed to longer hours, yet works appreciably more than a year ago. The mechanism is the ninety-second task: small enough to do in a queue, on a sofa, at 22:40, and therefore never large enough to justify refusing. The working day did not expand by decision; it expanded by a thousand individually trivial concessions, none of which would survive being described out loud as a policy. Leakers are usually the last to recognise the problem, because at no point did they do anything unreasonable.

• The Ratchet-Setter. Sits above the work, watches throughput rise, and adjusts the expectation upward — reasonably, incrementally, and without ever having produced at the new rate personally. This is not villainy; it is the most natural act in management, and the information reaching them fully supports it. The rising line is real. What the line does not contain is the multitasking load, the evaporated slack, or the fact that the gain was one-off while the expectation is permanent. The Ratchet-Setter’s characteristic mistake is to treat a capability ceiling as a floor. Their characteristic surprise arrives around month nine, in the form of resignations from people who never once complained.

  ---

How to hold a line

The first thing to accept is that the ratchet has no author. There is nobody to persuade and nothing to appeal. It is a default that operates in the absence of a decision, which means the only counter available is an explicit decision, made early, and said out loud.

If you lead an organisation: name what the gain is for. There are four possible destinations — more output, higher quality, wider scope, or recovered capacity — and if you do not choose, the system chooses more output every single time. Choosing is not a soft gesture; it is the only lever that exists. And treat slack as infrastructure rather than waste. Some of the most durable firms run deliberate, funded unproductivity — protected time with no deliverable attached — not from generosity but because they have worked out that a system at 100% utilisation has no capacity to learn, notice, or absorb anything unexpected.

If you manage a team: make baseline changes visible. When a deadline moves earlier or a new deliverable appears, say so explicitly, in writing, as a change — because the danger is not the change itself but its silence. A baseline that everyone can see can be discussed and, if necessary, reversed. A baseline that arrived by drift cannot even be identified. And notice that your best performers are the ones setting the pace for everyone else, usually without meaning to. The Absorber on your team is quietly rewriting the standard your whole department will be held to.

If you are the one doing the work: stop treating your saved time as a private windfall to be hidden, and start treating it as a negotiating position to be spent deliberately. Say what you are doing with it — depth on the important thing, the analysis nobody has had time for, the review step you always skipped. Time claimed for a stated purpose is defensible. Time simply not visible gets reallocated by someone else. And learn to state what you did not do and why: the scope you declined, the request you pushed back a week. In an environment where everything is possible, the ability to say what you deliberately left undone is the only remaining evidence that judgement is being exercised at all.

If you are early in your career: understand that you have arrived at the top of a ratchet and will be told the current pace is normal. It is not normal. It is roughly three years old, and the people who set it did not have to sustain it while learning the job. Build your own protected slack before anyone can accuse you of having spare capacity, and be extremely careful about being the fastest person in the room. Speed is rewarded once. The baseline it creates lasts.

The uncomfortable truth

The prevailing story about AI and work is one of eventual liberation. First a difficult adoption period, then the productivity dividend, then — the four-day week, the reclaimed evening, the human hours restored. It is a decent story and it has one flaw: it assumes that the gain, once created, flows to the person who created it. Nothing in the history of technology supports that assumption.

The gain from the steam engine was real, and it was spent on more steam. The gain from the washing machine was real, and it was spent on cleaner houses. Neither was stolen. In both cases the capability was converted, invisibly and almost immediately, into a higher standard of what counted as adequate — and the higher standard, once established, was never renegotiated downward. Nobody in 1900 proposed returning to a fortnightly wash.

What makes this cycle different is only its speed. Cowan’s ratchet turned over generations. Ours turns over quarters, which means a person can now watch the entire arc happen to them personally — discover a gain in March, see it absorbed by June, and be described as tired by July — while remaining unable to point at the moment anything changed, because no such moment exists.

So the question worth asking is not whether AI will give you your time back. That framing has already conceded the argument, because it treats the outcome as something the technology decides.

The question is what your organisation decided to do with the gain, and whether anyone can remember deciding. If nobody can, that is not an oversight. That is the answer. In the absence of a decision, the ratchet decides, and the ratchet has exactly one setting.

Everyone is waiting for AI to hand back their time. It already did. It arrived in March, on an ordinary Wednesday, and it was gone by June — and the most unsettling part is not that it was taken. It is that nobody took it.

A partner at a mid-sized consultancy is asked to do something she has never been asked to do in twenty-three years of practice. Her firm has bought an AI drafting system, and to configure it properly the implementation team needs one thing from her: a document describing what a good client deliverable looks like.

She is, by universal agreement, the best judge of that in the building. Juniors circulate drafts to her before they circulate them to anyone else. She reads three pages, makes four marks, and the thing is transformed. Nobody has ever successfully described what she does. They just say she has an eye.

She books a morning to write the document. She writes: “Clear. Rigorous. Client-ready.” She looks at it. It is true, and it is worthless — every deliverable she has ever rejected was also aiming to be clear, rigorous and client-ready. She tries again, harder, and produces two pages of the kind of prose that appears in every consultancy’s quality handbook and has never once changed an outcome. By lunchtime she has a document that would not stop a single one of the four thousand mistakes she has caught in her career.

The system is configured anyway, using her two pages. It produces work that is clear, rigorous and client-ready, and that she rejects at roughly the rate she has always rejected things. When the implementation team asks her why, she says the thing she has always said, the thing that used to be enough: it’s not quite there.

Nothing here is a failure of intelligence, effort or seniority. She is not a bad expert. She is a superb expert running into a bill that has been quietly accruing beneath her entire career, and which has just been presented, in full, with no warning and no instalment plan.

The shield we were handed in 1966

In 1966, the chemist-turned-philosopher Michael Polanyi opened The Tacit Dimension with a sentence that would end up doing sixty years of quiet load-bearing work in how we think about human skill.

“I shall reconsider human knowledge by starting from the fact that we can know more than we can tell.”

— Michael Polanyi, The Tacit Dimension, 1966

His examples were homely and devastating. You can recognise a friend’s face in a crowd of ten thousand and cannot describe that face precisely enough for a stranger to do the same. You can ride a bicycle and cannot state the rule you are obeying — which is, in fact, a rather specific relationship between the angle of imbalance and the radius of the corrective turn, a rule almost no cyclist has ever heard of and every cyclist perfectly obeys. Knowing and telling are separate faculties, and the first is vastly larger than the second.

For most of the twentieth century this was an interesting philosophical observation. Then computers arrived, and it became an economic prophecy.

In 2014, the MIT labour economist David Autor formalised it in a paper for the National Bureau of Economic Research and gave it the name it now carries: Polanyi’s Paradox. His argument was structural. Automation requires specification. A machine can only be given a task that has been fully and explicitly described. Therefore the tasks that resist automation are precisely those that draw on tacit knowledge — the ones demanding flexibility, judgement and common sense that we understand only implicitly and cannot write down. The frontier of automation was not set by how clever machines were. It was set by how articulate humans could be.

This became the great professional reassurance of the last decade, repeated in every keynote, every reskilling deck, every reassuring memo about the future of work. The routine goes; the judgement stays. And it was correct. The thing worth noticing now is why it was correct — because that reason has changed.

Tacit knowledge was safe not because it was valuable, but because it was unspeakable. Your expertise was protected by a wall built out of your own inability to describe it. That is a strange thing to have been relying on. It is stranger still that nobody thought to ask what would happen if a machine turned up that did not need the wall taken down brick by brick — only a rough sketch of what was on the other side.

How the shield became an invoice

Every previous wave of automation demanded complete specification. To automate a task with software you had to decompose it into unambiguous rules covering every branch and every edge case. That was an enormous, expensive, often impossible act of articulation, and Polanyi’s paradox stood in the way of nearly all of it. The bar was so high that most human judgement never came close to being tested.

Generative AI lowered that bar to the floor, and in doing so changed the nature of the problem entirely. It does not need your task decomposed. It arrives already carrying an enormous amount of general competence — it can write, structure, analyse, design, summarise and argue at a baseline level without being told how. What it needs from you is not a procedure. It is a standard: a statement of what, among the millions of competent outputs it could produce, counts as the right one.

That sounds easier than writing rules. It is in fact the harder half, because it is the half that was always tacit. The procedure was never the secret. Any competent junior can execute the procedure. The secret was the judgement that separates the acceptable version from the correct one — and that judgement has spent your entire career living in your hands and your gut, never in language.

So the paradox has not been solved. It has been inverted. It used to protect you by making your knowledge unautomatable. It now bills you by making your knowledge unusable — because the machine will happily do the work, at speed and at scale, at exactly the level of the standard you were able to state. And you were never trained to state one.

Worse: the gaps do not stay empty. This is the part most organisations have not yet noticed. When you fail to specify something, you do not get a blank where your standard should be. You get the model’s default — a plausible, well-formed, statistically central choice drawn from everything it has ever seen. Every unarticulated preference in your brief is silently filled in by somebody else’s average. The output is not missing a standard. It is carrying one you did not choose and cannot see.

And there is a second-order cruelty that the decision-making literature has documented for decades. Articulation is not merely difficult; it is unreliable. Gary Klein spent a career studying how genuine experts decide under pressure — fireground commanders, intensive-care nurses, military officers — and found they do not compare options at all. They recognise patterns, and the right course of action presents itself already formed. Crucially, when asked to explain, they produce accounts that are confident, coherent and frequently wrong about their own process. This is precisely why Klein and his colleagues had to invent Cognitive Task Analysis: an entire elicitation methodology, involving structured critical-incident interviews and probe questions, built on the premise that you cannot get expertise out of an expert by asking them for it.

Which means the first specification you write is almost never your standard. It is your theory of your standard. And the machine will build your theory, faithfully, at scale, while you stare at the result knowing something is wrong and unable to say what.

What the debt actually costs

The bill is already showing up in the aggregate numbers, misfiled as something else.

Forrester’s research on organisational “AI quotient” found that the share of employees who understand prompt engineering rose from 22% in 2024 to 26% in 2025 — four percentage points in a year of the most intense technology adoption in modern corporate history. That is usually reported as a training failure. It is more honestly read as evidence that the underlying capability is not a technique you can teach in an afternoon. You are not asking people to learn a syntax. You are asking them to make explicit something that has been implicit for their entire working lives.

Meanwhile, a Workday report in January 2026 found that almost 40% of apparent AI productivity gains in the workplace were being lost to rework and low-quality output. Nearly half the benefit, evaporating. The instinctive explanation is that the models are not good enough yet. Set that against the fact that model capability has risen sharply across the same period while the rework figure has not correspondingly fallen, and a different reading becomes hard to avoid: a large share of that 40% is not the machine failing to meet the standard. It is the standard never having been stated. That is the interest payment on the articulation debt, and it is being booked as a technology problem.

The distributional effect is the part that will reshape careers, and it runs in a direction almost nobody has priced in.

Seniority, in most professions, is a store of accumulated tacit judgement. Twenty years of pattern recognition; a library of cases; a finely calibrated sense of when something is off. In a world where work is executed by hand, that store converts directly into value — you do the work, and your judgement is embedded in every line of it, without ever needing to leave your head. In a world where work is executed by instruction, that same store converts into value only through the narrow aperture of what you can say. The library is still there. The door out of it has shrunk to the width of a sentence.

The consequence is genuinely uncomfortable: the twenty-year veteran with immaculate taste and no vocabulary for it can be functionally slower, right now, than a five-year mid-level who has less judgement but can state precisely what they want. Not because the veteran knows less. Because less of what they know is transmissible, and transmissibility has just become the constraint on everything.

And then there is the loss nobody is measuring at all. Standards that are never articulated are not merely underused — they are unpropagated. They do not enter the systems, the briefs, the templates or the model configurations. Which means the taste layer of an entire organisation quietly converges on whatever the model’s defaults happen to be, while the actual expertise sits in three or four people’s heads, retiring on schedule.

The four debtors

The debt does not present the same way to everyone. Four recognisable positions, and most people are standing in one of them right now.

• The Silent Master. Decades of genuine, hard-won judgement and almost no language for any of it. Their feedback vocabulary is a small set of gestures — not quite, closer, that’s it — which worked perfectly for thirty years because the recipient was a human who could triangulate from tone, context and repetition. A model cannot triangulate. The Silent Master carries the largest debt in the building and, painfully, the least awareness of carrying it, because until now their inability to explain was simply read as the mystique of expertise. The tell is a rejection rate that stays high while the quality of the instructions never improves.

• The Over-Specifier. Has correctly grasped that specification matters, and drawn exactly the wrong conclusion from it. Now writes twelve-hundred-word briefs describing tone, audience, structure, format, length and philosophy — everything, that is, except the specific discriminating judgement that actually separates the good version from the merely adequate one. They have confused volume with precision. Detail is not discrimination. A brief that describes everything equally has ranked nothing, and a model given a flat landscape will pick a plausible point on it. The Over-Specifier’s output is reliably competent, reliably generic, and reliably not what they wanted.

• The Borrowed Standard. Has no articulated standard of their own and, rather than confronting that, has quietly adopted the model’s output as the definition of good. Debt-free by way of bankruptcy. This is the most dangerous position of the four, because from the outside it looks like fluency: their work is fast, polished, well-formed and entirely uncontroversial. What has actually happened is that a human judgement layer has been removed from the system and nobody noticed, because the thing that replaced it produces confident, professional-looking work. Scale this across a department and the organisation’s taste becomes the statistical centre of the internet, applied consistently.

• The Translator. The rare person who both holds a real standard and can put it into words another person — or a machine — could act on. For most of their career this was not treated as a talent. It was treated as a personality trait, usually a slightly irritating one: the colleague who wanted the style guide written, who asked what “good” meant in the kickoff meeting, who could never let a vague brief through. They were the pedant. They are now, abruptly, the most leveraged individual in the organisation, because their judgement is the only judgement that can be installed anywhere other than in their own hands.

  ---

How to start paying it down

The first move is to stop treating this as a prompting problem. Prompting is a syntax, and syntaxes are learned in weeks. Articulating a standard is an act of self-excavation, and it is slow, effortful, and — this is the part that surprises people — genuinely improves the underlying expertise. Klein’s whole field exists because forcing tacit knowledge into language does not just extract it. It sharpens it.

• If you are the expert: stop trying to write your standard from a blank page. That is the one method the research says will not work; you will produce your theory of your standard and nothing more. Work from artefacts instead. Take twenty pieces of past work you rejected and twenty you approved, and for each rejected one write a single sentence naming the specific thing that made it wrong. The pattern will emerge from the pile, not from introspection. This is Cognitive Task Analysis in cheap, self-service form, and it is far more effective than any amount of staring thoughtfully out of a window.

• If you are early in your career: you carry far less debt, and that is a real, temporary and rapidly appreciating advantage. Build the habit now of writing down what good looks like before you generate, not after you are disappointed. But do not mistake being articulate for having taste. Specifying well and knowing well are different capabilities, and only one of them takes years to acquire. The person who can say precisely what they want and wants the wrong thing has an efficient route to being confidently mediocre at scale.

• If you manage the work: treat articulation as work. Not a documentation chore appended to Friday afternoons, not a wiki page created in a burst of enthusiasm and abandoned in a fortnight — a funded activity with time, attention and someone accountable for it. The highest-return exercise available to most teams right now is embarrassingly low-tech: sit your best practitioner beside your best writer and have the writer interview the practitioner about specific past decisions until a usable standard falls out. Then also change what you reward. If you promote on output volume, you will get Over-Specifiers and Borrowed Standards. If you reward the person who made the team’s judgement legible to everyone else, you will build an organisation whose expertise survives its experts.

• If you lead the organisation: understand what is actually on your balance sheet. Your firm’s competitive advantage is very likely a set of standards that exist nowhere except in the heads of a handful of long-tenured people, and every one of those heads has a departure date. Historically this was a slow-burning succession risk. It is now an acute one, because the competitor who articulates their standards can install them into systems that run continuously, at scale, without fatigue — and you cannot. Ask a blunt diagnostic question: if your three best people left tomorrow, what would remain of their judgement? If the honest answer is “nothing written down,” you do not have an AI strategy problem. You have an articulation debt, and it is your largest unrecorded liability.

  ---

The uncomfortable truth

There is a story we have been telling ourselves since roughly 2023, and it is a flattering one. In the story, AI takes the routine work and humans ascend to the judgement work — the taste, the discernment, the irreducibly human layer. It is comforting, it is widely believed, and it rests on an assumption nobody has examined: that judgement, once elevated, can actually be applied to a system that only receives language.

It cannot, for most of us, yet. Not because we lack judgement, but because our judgement has spent its entire existence in a form that never had to be spoken. We built careers on a paradox we treated as a fortress, and it turns out to have been a debt facility. Every year you were valued for knowing more than you could tell was a year of borrowing, drawn against a future in which someone would eventually need you to tell.

That future arrived without an announcement. The bill is not a training gap or a technology gap. It is the accumulated distance between what you know and what you have ever been able to say — and it is now the single most expensive gap in professional work.

The professionals who come through this well will not be the ones with the deepest expertise. They will be the ones who did the unglamorous, uncomfortable, genuinely difficult work of dragging their expertise into language, and discovered — as everyone who attempts it does — that some of what they thought they knew does not survive being said out loud.

For 100 years we were paid for knowing more than we could tell. The debt was always there. AI did not create it. It simply called it in, all at once, and asked us to speak.

A radiologist at a large teaching hospital gets a new AI triage tool. It reads each incoming chest scan, flags the ones that look urgent, and pre-marks suspected findings on the image before she ever opens it. Her job is unchanged on paper: she reads every scan and signs off. In practice her job has changed entirely. She is no longer the first pair of eyes. She is the second — the one confirming or overruling a machine that, in validation, agreed with expert consensus more than nine times in ten.

For the first fortnight she reads each scan as she always has, eyes tracking the lungs edge to edge, treating the AI’s marks as a colleague’s opinion rather than a verdict. Then the pattern of the work teaches her what the pattern of the work always teaches. The tool is right. It is right on Monday and right on Thursday and right for a hundred scans in a row. Her overrules get rarer. Her reads get faster. Somewhere around scan four hundred, without a single decision to do so, she stops reading the image and starts reading the marks — checking whether the AI’s boxes look plausible rather than searching the scan for what the boxes might have missed.

Then comes the scan the tool reads as clear, on a patient whose tumour sits in exactly the region the model is weakest on. The finding is faint but visible — to a radiologist who is looking. She is no longer looking. She is confirming. She signs off in nine seconds and moves to the next.

Nothing about this radiologist is negligent. She is highly trained, highly motivated, and doing precisely what the system was designed to make her do. That is the point. The failure was not a lapse in her character. It was the predictable output of asking a human to do the one thing humans are engineered to fail at — and then calling that arrangement a safety control.

A problem we solved, and named, in 1948

During the Second World War, the British military had a problem with radar and sonar operators. The equipment worked. The operators, staring at screens for hours waiting for the rare blip of an enemy submarine, kept missing the blips. Not through incompetence — through something more fundamental, and more troubling, because it could not be trained or disciplined away.

A young psychologist named Norman Mackworth was asked to study it. He built an apparatus now known as the Mackworth Clock: a plain clock face with a hand that ticked in regular steps and, at random and infrequent intervals, jumped two steps instead of one. The observer’s only task was to catch every double-jump across a two-hour watch. It is about the simplest sustained-attention task imaginable.

The result, published in 1948 and replicated thousands of times since, founded an entire field. Detection accuracy did not hold steady and then tail off with fatigue at the end. It collapsed early. Performance fell by ten to fifteen per cent within the first thirty minutes — and kept declining from there. This is the vigilance decrement, and its most important property is the one that makes it so dangerous: it afflicts motivated, rested, well-trained people who know they are being tested. Willpower does not touch it. It is not a flaw in particular workers. It is a property of the attentional system every human is issued at birth.

The cruel structure of it is this. Human attention is not built to be sustained on a source of information where almost nothing happens. The rarer the signal, the faster and deeper the decrement. A monitoring task where events are frequent keeps the mind engaged. A monitoring task where the thing you are watching for almost never occurs is precisely the task the brain cannot hold.

Which sets up the central irony, articulated in 1983 by the cognitive psychologist Lisanne Bainbridge in a four-page paper, “Ironies of Automation,” that has aged better than almost anything written about technology since. Her argument was deceptively simple: we automate the tasks humans are bad at, and in doing so we hand the human the residual task of monitoring the automation — a task humans are even worse at.

“We know from many ‘vigilance’ studies that it is impossible for even a highly motivated human being to maintain effective visual attention towards a source of information on which very little happens, for more than about half an hour.”

— Lisanne Bainbridge, “Ironies of Automation,” Automatica, 1983

Bainbridge’s paper has survived for four decades because it does not describe a fact about 1983 technology. It describes a structural relationship between humans and reliable machines — one that gets worse, not better, as the machine improves. And we have just handed that exact relationship to two billion knowledge workers, and given it a reassuring name.

Why AI is the most dangerous automation we have ever monitored

Every previous generation of automation that put a human on watch at least had the decency to look mechanical. A cruise control, an autopilot, an assembly-line sensor — you could tell you were supervising a machine. Generative AI removed that tell. It produces work that looks exactly like the work of a competent, careful colleague. It writes in complete sentences. It cites. It hedges where a thoughtful person would hedge. It sounds certain where a confident person would sound certain. And it is right often enough that scrutinising it begins to feel not just unnecessary but faintly insulting, like re-checking the arithmetic of someone who has never once got the sum wrong.

This fluency is not a cosmetic feature. It is the mechanism by which the trap springs harder. The research literature on automation bias — our tendency to over-trust automated systems — identifies two distinct failure modes, and fluent AI aggravates both simultaneously. The first is the omission error: you fail to notice a problem because the system did not flag it. The radiologist who stops searching the scan and searches the boxes instead is committing omission errors by the hundred; she has outsourced not just the analysis but the very act of looking. The second is the commission error: you actively follow the system’s recommendation even when other evidence in front of you contradicts it. The more authoritative the system sounds, the more its confident wrong answer overrides your quiet correct doubt.

What makes this worse than the old radar screen is that the AI is not a source of information on which very little happens. It is a source on which a great deal happens — a torrent of plausible, polished output — almost all of which is fine. That is the vigilance decrement’s ideal habitat, dressed up as its opposite. The operator feels busy, engaged, productive. There is no dead screen to fight. There is a stream of competent work sliding past, and the rare defect is a single deformed shape in a river of well-formed ones, moving at the speed of your approval clicks.

And here is the reliability paradox in its purest form: the better the model gets, the more certainly the human oversight fails. A model that is wrong ten per cent of the time keeps you at least intermittently alert, because you catch errors often enough to stay in the game. A model that is wrong a tenth of a per cent of the time lulls you completely — and then delivers its rare error to a reviewer who has not truly reviewed anything in weeks. Improving the AI does not reduce the risk of the human-plus-AI system. Past a certain point, it increases it.

What gets lost, and the debt nobody records

The evidence that we have already surrendered the oversight role is not subtle. In the 2025 global study of trust in AI conducted by KPMG and the University of Melbourne — roughly 48,000 people across 47 countries, one of the largest surveys of its kind — 66% of respondents said they rely on AI output without evaluating its accuracy. 56% admitted they have made mistakes in their work because of AI.

These are not the numbers of a species diligently keeping AI in the loop. They are the numbers of a species that has quietly stepped out of it while leaving a mannequin in the chair.

The cost of this does not appear anywhere it can be easily seen, and that is precisely what makes it accumulate. Every hour AI saves is booked immediately, visibly, and enthusiastically — in dashboards, in time-saved metrics, in the genuine relief of people whose drudgery just evaporated. But that saved hour is not free. A portion of it is a loan drawn against future vigilance, and the repayment falls due unpredictably, all at once, on whoever happens to be holding the rare error when it finally ships.

Call it the vigilance debt: the silent, compounding liability created every time speed is purchased by withdrawing attention, recorded on no ledger because a monitoring failure produces no artefact until the moment it produces a catastrophe.

Ordinary work leaves a trail of near-misses. You notice yourself almost making a mistake, and the noticing keeps you sharp. Vigilance failure leaves no such trail. The scans you signed off without reading look identical to the scans you signed off after reading. The letters you approved on rhythm look identical to the letters you approved on judgement. There is no feedback, no wobble, no warning tremor — right up until the one that was wrong goes out to four thousand customers or one patient, and the incident review discovers, to everyone’s genuine surprise, that the human control everyone was relying on had quietly stopped controlling anything months ago.

Regulators have noticed what most organisations have not. The European Union’s AI Act, in its provisions on human oversight, does something almost unheard of in legislation: it names a specific cognitive bias. Automation bias is the single psychological phenomenon the Act explicitly calls out as a threat to effective human oversight of high-risk systems. The lawmakers understood what the average deployment plan does not — that writing “a human will review the output” into a process does not make the review happen. Vigilance is not summoned by an org chart. It is defeated by one.

The four faces of the vigilance trap

The trap does not present the same way to everyone. It has recognisable archetypes, and naming them is the first step to noticing yourself becoming one.

• The Rhythm Clicker. This is the reviewer whose approvals have become motor memory. The task has a cadence — open, glance, approve, next — and once the cadence sets, the content stops entering conscious attention at all. The Rhythm Clicker is not lazy; they are efficient, and their efficiency is the disease. They have optimised a task down to the point where the one part that mattered, the actual looking, has been optimised away. The tell is that they could not tell you anything specific about the last ten things they approved.

• The Fluency Truster. This reviewer still reads — but reads for polish rather than for truth. Because the AI’s output is articulate, structured and confident, the Fluency Truster’s brain accepts fluency as a proxy for correctness, which it never was. They will catch a typo and miss a fabricated statistic, because the typo disrupts the surface and the fabrication does not. This is the most insidious archetype, because it feels like diligence. They are working hard. They are just checking the wrong layer.

• The Volume Drowner. Before AI, this person reviewed ten things a week and could give each real attention. AI raised the inflow to a hundred, and nobody adjusted the reviewing capacity, because reviewing was not the thing the rollout was about. The Volume Drowner is trying to be vigilant and is structurally prevented from it. Their failure is not attentional but arithmetic: genuine scrutiny of a hundred artefacts a week was never on the menu, so what gets called review is triage at best and theatre at worst.

• The Lonely Sentinel. This is the single human designated as the check on a system that runs thousands of times a day — the one radiologist, the one approver, the one “human in the loop” whose sign-off legitimises the entire pipeline. The Lonely Sentinel carries the full weight of a safety story they cannot possibly bear, and everyone upstream and downstream behaves as though the sentinel’s presence has made the system safe. Their existence is what allows the organisation to stop worrying. Their existence is the reason it should worry more.

  ---

What to actually do about it

The first and most important move is to stop treating vigilance as a matter of will. Bainbridge told us in 1983 that you cannot exhort your way out of the vigilance decrement, and seventy-five years of data agree. A policy that says “reviewers must carefully check all AI output” is not a control. It is a wish, and worse, it is a wish that transfers liability onto individuals for a failure the system made inevitable. Every intervention that actually works is a design intervention, not a motivational one.

• If you are an individual reviewer: your value has quietly inverted. The scarce, rising skill is no longer producing output faster — the machine wins that outright and permanently. It is being the person who genuinely reads the thing when everyone around you has stopped. Cultivate adversarial review as a deliberate practice: begin from the assumption that the plausible answer in front of you is wrong, and spend your attention hunting for where. Read for truth, not polish. And protect your own attentional conditions, because a Rhythm Clicker is just a good reviewer who was never given a reason to stay awake.

• If you manage the work: the highest-leverage thing you can do is design the reviewing task so that a normal human can actually perform it. Cap the length of unbroken monitoring stretches — the decrement bites at thirty minutes, so rotate reviewers before that, not after. Deliberately re-introduce signal: seed a known error rate into the stream so that the reviewer catches something often enough to stay engaged, the way pilots train on simulated failures and airport screeners are shown planted threat images to keep detection rates up. Measure catch rates, not throughput. And never let one person be the whole control for a high-volume system; the Lonely Sentinel is an org-design failure wearing a job title.

• If you lead the organisation: internalise the reliability paradox before it internalises you. The safety of a human-plus-AI system does not rise monotonically with model quality — it can fall as the model improves, because improvement erodes the human check. This means your risk is highest precisely when your metrics look best and your people feel most confident. Instrument for it. Ask not “how much time did AI save?” but “when did a human last catch something the AI got wrong, and how would we know if they had stopped?” If you cannot answer the second half of that question, you do not have oversight. You have a green button and a person to press it.

• If you design these systems: the goal is not to add a human to the loop. It is to build a loop a human can actually stay awake inside. That means surfacing the model’s uncertainty rather than hiding it behind uniform fluency, varying the presentation so the reviewer cannot settle into rhythm, forcing genuine engagement at the moments that matter rather than requesting blanket attention that cannot be given, and — hardest of all — resisting the commercial pull to make the output so smooth and confident that trusting it becomes the path of least resistance. Every increment of polish you add is an increment of vigilance you take away.

  ---

The uncomfortable truth

We told ourselves a story about AI and human judgement. In the story, the machine handles the volume and the human provides the wisdom — the oversight, the discernment, the final responsible glance that keeps the whole thing safe. It is a comforting story, and it is the story written into a thousand deployment plans and at least one major piece of European legislation. The trouble is that it depends entirely on a capability we have known for three-quarters of a century that humans do not reliably possess: sustained vigilance over a system that rarely fails.

The deeper irony is that we are running the experiment backwards. We are pouring effort into making the models more reliable, and each gain in reliability quietly weakens the human check that our safety story rests on. The better it gets, the less we watch. The less we watch, the more the rare failure matters. And the rare failure is the only kind a highly reliable system produces.

“Human-in-the-loop” was never a solution. It was the problem, restated as a reassurance. The organisations that come through this well will be the ones honest enough to admit that a human posted to watch a nearly-perfect machine is not a safeguard but a sedative — and to redesign the work around the reviewer we actually are, rather than the tireless sentinel we keep pretending to be.

The machine’s job is to be right almost always. Ours was supposed to be catching the moment it isn’t. We should probably find out whether anyone is still watching.

The bids team at a mid-sized engineering firm used to take three weeks to turn a client brief into a submitted proposal. After the AI rollout, the first draft came back in a day and a half. It was good. Not good enough to send unedited, but genuinely good — structured, costed, referencing the right precedents. The team lead sent it to legal on a Tuesday morning and, with entirely reasonable confidence, told the client to expect it by Friday.

Legal was one person. Legal had been one person before the rollout, and nobody had thought to change that, because the rollout was about drafting and legal review is not drafting.

Within a quarter the bids team was producing eleven proposals a week. Legal could review two. The firm’s average response time to clients — the only number the client actually experiences — moved from twenty-two days to twenty-six. It took another full quarter for anyone to notice, because every internal metric anyone was watching had improved dramatically. Time to first draft: down 90%. Drafts produced per head: up 14x. Employee-reported time saved: substantial. Every one of those numbers was true.

This is not a story about a badly run rollout. It is very close to the median story, and the aggregate data says so. In July 2025, MIT’s Project NANDA published The GenAI Divide: State of AI in Business 2025, based on more than 300 enterprise deployments, 52 structured case-study interviews and 153 leadership surveys. Their headline finding — that roughly 95% of an estimated 30 to 40 billion dollars in enterprise generative AI spend had produced no measurable profit-and-loss impact — briefly moved markets and was then mostly absorbed into the general noise of AI commentary, filed under the reassuring heading of “early days.”

It is not early days. It is the third year. And the reason so much of that spend evaporated is not that the technology underdelivered. In most of these deployments the technology did exactly what it said on the tin. The problem is that the thing it made faster was never the thing that was slow.

What the report actually said, and what everyone heard

The MIT authors were reasonably direct about where the failure sits.

“The core barrier to scaling is not infrastructure, regulation, or talent. It is learning.”

— Aditya Challapally, Chris Pease, Ramesh Raskar et al., The GenAI Divide: State of AI in Business 2025, MIT Project NANDA

What most readers heard was a claim about the models — that the tools do not retain feedback or adapt to context, which the report also says. But read the sentence again with an organisational eye and it says something broader and more uncomfortable: the barrier is not any of the things a company can purchase. It is the thing a company has to change about itself. And the specific thing most companies did not change is the shape of the path work takes through people.

Gallup’s 2026 workplace research gives that gap a number. For the first time in their measurement, half of employed American adults say they use AI in their role at least occasionally, with 13% using it daily. And yet only about one in ten employees inside AI-adopting organisations strongly agree that AI has transformed how work gets done where they work. Gallup’s own reading is that the benefits are concentrated at the level of individual tasks rather than workplace systems, because most organisations have not redesigned workflows, roles or processes around the technology at all.

Individual speed: up. Organisational throughput: essentially unchanged. That is not a paradox in need of resolving. It is the predictable result of optimising a link in a chain that was never the constraint.

We have run this experiment before

None of this should have surprised anyone, because there is an unusually well-documented precedent. In 1987 Robert Solow made the observation that became the productivity paradox: you could see the computer age everywhere except in the productivity statistics. Firms had spent enormously on computing through the late 1970s and 1980s, and measured productivity growth had gone sideways. The eventual resolution, worked out across a large literature over the following two decades, was not that computers were useless. It was that the gains required complementary organisational investment — different processes, different job designs, different decision rights — and those investments take years, cost real money, and depress measured productivity while they are being made. Firms that bought the computers and kept the workflow got the paradox. Firms that redesigned around the computers got the gains, roughly five to seven years later.

There is an even older piece of arithmetic underneath it. Amdahl’s law, formulated for parallel computing in 1967, says that the speedup available from accelerating any part of a system is capped by the fraction of the system you did not accelerate. Make one stage of a process infinitely fast and the overall improvement is still bounded by everything else. If the drafting stage was 20% of the elapsed time from brief to submitted proposal, then reducing drafting to zero — not by 90%, to zero — improves the whole process by at most 20%. And drafting was almost never 20%. In most knowledge-work chains it is closer to 10%, with the remaining 90% consisting of waiting: for a review, for an approval, for a meeting where three teams can be in the same room, for a decision-maker to have thirty uninterrupted minutes.

The Atlanta Fed’s March 2026 working paper, drawing on a survey of nearly 750 corporate executives, gives the current version of the Solow observation its formal name. The authors document what they call a productivity paradox in the AI era: perceived productivity gains are consistently larger than measured productivity gains. Executives are not lying. They are reporting something real — the acceleration of the visible, individual, task-level work — and then discovering that it does not appear in the aggregate, because the aggregate was never governed by that stage.

The mechanism: why it is a tax and not merely a disappointment

Here is where the story turns from “less benefit than hoped” to something with a genuine cost attached.

Treat any work chain as a queueing system, because that is what it is. Work arrives at a stage, waits, gets processed, moves on. In a system like this, total throughput is governed entirely by the slowest stage, and the only reliable way to make the system faster is to widen that stage. Speeding up a stage that was not the constraint does not improve throughput at all. Worse, it increases the arrival rate at the next stage downstream. If the downstream stage’s capacity is unchanged, the queue in front of it grows, and the average time work spends in the system goes up.

This is not a metaphor. It is the actual mechanism in the proposal story: eleven drafts a week arriving at a reviewer with a capacity of two produces a growing backlog, a longer average wait, and a client experience that gets measurably worse while every internal dashboard reports triumph. The organisation has not eliminated a bottleneck. It has fed one.

The coordination tax is the charge this system levies at every point where work passes between a person doing it and a person who must respond to it. It is paid per handoff. It is levied in elapsed time, in re-explanation, in meetings convened to resolve what a handoff failed to transmit. And it is almost completely indifferent to how fast the work itself was produced — which means that as production accelerates, the tax does not fall as a share of the total. It rises. It becomes, in relative terms, nearly the whole cost.

Most organisations have never measured it, because no individual pays it. It is charged to the space between people, and no one owns that space. It appears on no team’s dashboard, in no one’s objectives, and in no vendor’s ROI model. It is precisely the sort of cost that survives a decade of efficiency programmes untouched, and then quietly consumes the entire benefit of the next one.

The second charge: coordination gets more expensive, not just no cheaper

If the tax merely stayed constant while production accelerated, the picture would be disappointing but stable. The evidence suggests something worse: introducing AI into collaborative work tends to raise the cost of the coordination itself.

Reviewing the human-AI teaming literature, Schmutz and colleagues find that adding AI to a team frequently reduces coordination, communication and trust — that human-AI teams underperform not because the AI is weak but because team cognition and mutual understanding degrade. The mechanism is intuitive once stated. Coordination between people runs on a shared model of what each other knows and has actually thought about. When a colleague sends you a document, how carefully you check it has always depended on your read of how carefully they made it. Remove your ability to infer that — because the artefact could plausibly have been produced with five minutes of prompting or five hours of reasoning, and looks identical either way — and your only rational response is to check more.

So verification effort per handoff rises at exactly the moment volume per handoff rises. Reviewers who could previously skim a colleague’s work because they knew the colleague now have to read it properly, on three times as many documents. Approvers who could take a recommendation on trust now ask for the reasoning. Alignment meetings that used to converge on two credible options now have to work through six, because generating options got cheap and choosing between them did not.

This is the part organisations consistently fail to anticipate. They model AI adoption as a reduction in production cost with coordination cost held constant. In practice production cost falls, volume rises, and coordination cost per unit rises simultaneously. Three variables moving, only one of them modelled.

The four tolls

The coordination tax is not a single charge. It is collected at four distinct gates, and separating them matters, because each one has a different remedy and only one of them got cheaper after AI arrived.

• The Approval Toll. Work waits for someone with authority to say yes. This is the purest form of the tax, and the most brutally unaffected by AI: the time cost is set entirely by the approver’s calendar and attention, both of which are fixed. Where an approval chain has three signatures, AI can compress the drafting from three weeks to two days and the elapsed time will fall from, say, thirty-five days to thirty-one. The visible work was never the number. The signatures were.

• The Context Toll. Every handoff requires the receiver to be brought up to speed on what the sender knows. AI makes this toll worse in a specific and under-appreciated way: it dramatically increases the number of artefacts in circulation without increasing the shared understanding behind them. More documents, more variants, more options — each of which arrives at a colleague carrying less implicit context than a hand-made artefact would have, because the person who sent it may not hold all of that context themselves.

• The Alignment Toll. Multiple parties must agree before anything moves. This toll scales with the number of options on the table and the number of people who must converge, and AI inflates the first term hard. When producing a strategic option costs an afternoon rather than a fortnight, teams arrive at the alignment meeting with six credible directions instead of two. Generation got a hundred times cheaper; agreement got no cheaper at all, and now has three times as much to agree about.

• The Verification Toll. Someone must check the work before it counts. This is the toll that AI has damaged most directly, for the reason described above: volume rises while confidence per artefact falls, so total verification load rises on both axes at once. It is also the toll most likely to be quietly skipped under pressure, which converts a coordination problem into a risk problem without anyone deciding to accept the risk.

Read the four together and the pattern is stark. Of the four gates where organisational time actually goes, AI has meaningfully improved exactly none, made three actively more expensive, and been sold on its ability to accelerate a fifth stage — production — which was rarely the binding constraint in the first place.

The variable that actually predicts it working

There is a version of this story where AI produces real, compounding, measurable returns. It exists. The 5% in the MIT study are not statistical noise; they were extracting millions in value. The interesting question is what separates them, and the most useful answer in the current literature has nothing to do with tooling.

Wang, Feng and Sun built an inter-firm mobility network from 460 million job records covering more than 16,000 US companies, tracking how AI capability spreads between firms through the movement of AI workers. They found productivity spillovers two to three times larger than those historically associated with traditional IT — a striking result on its own. But the effect was conditional in a way that should reframe how leaders think about AI strategy entirely: hiring AI talent out of flatter, more lean-startup-intensive firms generated significant productivity gains, while hiring from firms lacking those traits yielded little benefit. Their mechanism tests point to flat and lean organisations producing more versatile AI generalists who carry richer, more transferable knowledge with them.

Same technology. Same models, available to everyone at the same price. Comparable people. Different organisational shape, radically different outcome. The authors’ broader framing is that AI spillovers, unlike IT spillovers, depend on experimental and integrative environments rather than on scale and process standardisation — which is another way of saying that the returns accrue to organisations with short paths between the work and the decision.

Which is the whole argument in a sentence. AI is not a productivity technology at organisational scale. It is an amplifier of whatever coordination structure you already had. In a flat organisation with few gates it compounds. In a layered one it congests. And the standard response to congestion — buying more AI, because the first deployment “clearly worked, look at the time-saved numbers” — is paying the same toll twice.

What to actually do about it

• For individuals, the strategic implication is uncomfortable but clarifying: your visible output has stopped being a differentiator, because everyone’s rose at the same time. What is now scarce is the ability to move work through people — knowing who actually decides, what they need in order to decide, how to get a clean yes on the first pass rather than the third, and when to skip a gate entirely because it exists for historical reasons nobody can articulate. This was always undervalued relative to production skill. It is now the constraint, which means it is where the returns are.

• For managers, the single highest-leverage act is to instrument the whole chain rather than the fast link. Almost every AI dashboard currently in use measures time saved per task, adoption rate, and self-reported productivity — three metrics that will report success indefinitely regardless of whether anything reaches a customer faster. Replace them, or at least outrank them, with elapsed time from request to delivered, and with queue depth at each internal handoff. The second number is the one that will be unpleasant, and it is the one that tells you where your ceiling is.

• For leaders, the arithmetic to internalise is Amdahl’s: the return on accelerating any stage is capped by the fraction of elapsed time that stage represents. Before the next tool purchase, map where elapsed time actually goes for one important piece of work — genuinely map it, in days, from request to delivery — and count how much of it is production versus waiting. In most organisations the ratio is bad enough to make the case on its own. If 85% of elapsed time is waiting at four gates, then removing a gate is worth more than any conceivable improvement in the model behind the drafting.

And for anyone designing an organisation from here: the Wang, Feng and Sun result suggests the highest-return AI investment available to most firms is not an AI investment at all. It is delayering. Every gate removed is a toll no longer charged, on every piece of work, forever — and unlike a tooling upgrade, nobody can buy the same advantage next quarter.

The ceiling nobody put on the slide

The seductive thing about the coordination tax is that it is invisible from inside every role that pays it. The person drafting sees their own work getting faster and concludes the system is faster. The reviewer sees an unmanageable queue and concludes the drafters are being careless. The executive sees a dashboard of time-saved metrics and concludes the rollout succeeded. The client sees a response arriving twenty-six days after they asked, four days later than last year, and concludes nothing at all, because they have nothing to compare it against.

Not one of these people has an incorrect view of what they can see. The tax lives in the space between them, which is exactly the space no dashboard covers, no individual owns, and no vendor sells a fix for. It is the reason 95% of an enormous investment produced nothing measurable, and it is also the reason the remaining 5% produced so much: they had fewer gates to begin with, not better models.

The uncomfortable truth is that the last three years have been a very expensive natural experiment testing a hypothesis nobody stated out loud — that the constraint on knowledge work was individual production capacity. The experiment has now returned a fairly clear result, at a cost of tens of billions of dollars, and the result is no. The constraint was the structure. It always was. AI simply removed the last remaining excuse for pretending otherwise, by making the production side so fast that there is nothing left to blame.

You did not buy a faster organisation. You bought a faster way to reach the queue.

Daniel Okafor had been writing Python for three years when his team asked him to pick up Trio, a concurrency library nobody else on staff knew well either. He worked through the self-guided tutorial the way he worked through everything now: with an AI assistant open in a side pane, ready to generate a working example the moment he got stuck. He finished the tutorial in the time the syllabus allotted. He built the sample project. Everything ran.

Three weeks later, a production job using Trio hung intermittently in a way the tutorial had never covered, and Daniel stared at the stack trace with the specific, sinking recognition of someone who has never actually debugged the thing he supposedly learned. He had finished the course. He had not learned the library. He had learned how to get an AI to finish the course for him, which turns out to be a different skill entirely, and one with a much shorter half-life once the assistant isn’t in the room.

Daniel is not a real name, but his experience is close to a real result. In January 2026, Anthropic researchers Judy Hanwen Shen and Alex Tamkin ran fifty-two software developers through exactly this scenario in a randomized controlled trial: learn the Trio library, half with an AI assistant able to generate correct code on demand, half without. The two groups finished in roughly the same amount of time — AI conferred no meaningful efficiency gain in this setting. What it did confer was a gap: the AI-assisted group scored 17% lower on a follow-up quiz testing conceptual understanding, code reading, and debugging, with the largest losses concentrated in debugging specifically — the skill you only build by sitting with an error until you understand it.

The researchers’ own summary of the result is blunt enough to serve as this edition’s epigraph.

“AI-enhanced productivity is not a shortcut to competence.”

— Judy Hanwen Shen and Alex Tamkin, “How AI Impacts Skill Formation”, Anthropic, 2026

That sentence deserves to be read twice, because it contradicts the entire pitch of the last three years of workplace AI adoption. The pitch was never “this will make you incompetent faster.” The pitch was speed, and speed delivered — sometimes. But underneath the speed, in domain after domain, researchers keep finding the same second effect, arriving quietly and mostly unmeasured: something that used to get built through the ordinary friction of work has stopped getting built, because the friction is gone.

That something is feedback. This is an account of the vacuum left behind when it disappears.

What feedback actually was

It is worth being precise about what got removed, because “feedback” has become one of those words flattened by overuse in performance-review culture until it barely means anything. Strip it back to the mechanism. A feedback loop, in the sense that matters here, has four steps: you attempt something, using your current best understanding; the world returns a signal about whether the attempt worked; you notice the gap between what you expected and what happened; and you update your understanding to close that gap next time. Attempt, signal, noticing, update. Run that loop enough times in a domain and the update process compresses — you stop consciously reasoning through problems you’ve solved before and start recognising them. That compression is what “expertise” is. It is not a personality trait or a certificate. It is thousands of completed loops, most of them too small and too fast to remember individually.

Every established profession built its training around this mechanism, usually without describing it in those terms. Apprenticeship systems put novices next to masters precisely so that attempts could be corrected quickly and cheaply, on low-stakes work, before the stakes rose. Medical residencies are built entirely around supervised attempts with rapid correction — the resident makes the call, the attending confirms or corrects it, and the resident updates in real time, thousands of times, before being trusted alone. Newsrooms had sub-editors whose entire function was to hand a piece back covered in corrections, which is a brutal but extremely efficient feedback mechanism, one many journalists will tell you they learned more from than journalism school. None of these systems were designed by anyone thinking about “feedback loops” as a formal concept. They were designed by people who noticed, empirically, that competence didn’t develop any other way.

What all of them share is a structural feature that AI assistance quietly breaks: the correction has to arrive after the attempt, and the person has to notice it as a correction. If the “attempt” step is skipped — if the AI generates the answer before you’ve formed your own — there is nothing for the correction to correct. If the correction happens invisibly, folded into a tool’s silent auto-fix, there is a correction but no noticing. Either way, the loop that builds competence never completes, even though, from the outside, work is visibly getting done, on schedule, to a high standard. That is the trap. The vacuum is invisible from the output side. It only shows up later, in the moment the tool isn’t there.

The mechanism: where AI actually intervenes

It’s tempting to describe this as AI simply “doing the work for you,” but that framing misses where the damage happens. AI doing routine, low-skill work for you is not the problem — nobody develops expertise from filing expense reports, and automating them is a pure gain. The Feedback Vacuum opens specifically when AI intervenes in tasks you are still in the process of learning, and it opens at one precise point in the loop: between the attempt and the noticing.

Consider the two most common patterns of use. In the first, the assistant generates the answer before you’ve formed your own hypothesis — you describe the problem, it returns working code, a finished paragraph, a completed diagnosis. There is no attempt for a correction to attach to, so there is no loop, just output. In the second, you do form an attempt, but the assistant “helps” by silently repairing the parts that were wrong — autocomplete finishing your syntax before your typo becomes an error message, a grammar tool smoothing your sentence before its awkwardness registers as a lesson, a code-review bot auto-patching an edge case you never saw. Here an attempt exists, but the signal that would have told you it was wrong is intercepted before it reaches you. Both patterns look identical from the outside: correct output, produced efficiently. Only one of them — neither, in fact — leaves anything behind.

METR’s 2025 randomized trial of sixteen experienced open-source developers is the sharpest illustration of how invisible this is even to the people it’s happening to. Given AI tools for real coding tasks in mature codebases, the developers took 19% longer to finish than the control group working unassisted. Beforehand, they had forecast that AI would cut their time by 24%. Afterward — having just been measurably slower — they estimated that AI had made them about 20% faster. The gap between the forecast, the outcome, and the post-hoc perception is not a rounding error. It is what a severed feedback loop looks like applied reflexively to the tool itself: the developers had no reliable internal signal for whether the AI was helping, because the very faculty that would normally supply that signal — noticing when something worked and when it didn’t — was itself one of the casualties.

A 2026 theoretical model from Venkat Ram Reddy Ganuthula and a co-author, published in Human Behavior and Emerging Technologies, formalises why this perception gap is not a coincidence. Their model treats skill as a running balance between learning-through-practice and forgetting-through-disuse, with AI assistance shifting both terms of the equation at once: it raises visible, current-period performance — you produce good output today — while simultaneously reducing the practice that maintains skill over time, because the assistant is doing the practising for you. Run the simulation forward and you get a curve that rises before it falls: performance improves in the short run, exactly as advertised, then erodes in the long run, exactly as nobody advertised, and the two phases are far enough apart in time that the erosion is easy to misattribute to something else entirely — a market downturn, a personal slump, “just not what I used to be.”

Whose loop, exactly

The Feedback Vacuum is often discussed as though it is purely a software engineering phenomenon, because software engineering happens to be the domain with the cleanest randomized trials. It is not confined there, and the clearest evidence of that comes from medicine, where the stakes make the pattern much harder to ignore.

In a study published in Radiology in 2023, twenty-seven radiologists — spanning inexperienced, moderately experienced, and highly experienced readers — assessed fifty mammograms with the assistance of a purported AI system offering a BI-RADS category suggestion. When the AI’s suggestion was correct, inexperienced readers matched it correctly almost 80% of the time — a strong result, on its face. When the researchers had the AI suggest an incorrect category, the same inexperienced readers’ accuracy collapsed to under 20%. They were not reading the mammogram and consulting the AI. For all practical purposes, they were reading the AI, and checking the mammogram only enough to notice that it hadn’t obviously contradicted it.

This is the Feedback Vacuum in a setting where the “attempt” step is a diagnosis and the “signal” is, eventually, a biopsy result — a genuinely high-stakes loop that took decades to build into radiology training. The AI did not remove that loop by design; it removed it by convenience, one confident BI-RADS suggestion at a time, in exactly the cases — inexperienced readers, still forming their own independent read — where the loop mattered most. The radiologists who resisted the bias were, unsurprisingly, the most experienced ones: readers who had already banked enough completed loops, before the tool arrived, to have an independent judgment worth defending. The vacuum does not open evenly. It opens fastest in exactly the people who have the least stored expertise to fall back on — which is to say, the people the profession most needs to be developing.

The four shapes of feedback loss

Not everyone loses the loop the same way. Four recognisable patterns of AI use show up across the research and across ordinary workplace observation, and only one of them keeps the loop intact.

• The Vending Machine User. Describes the problem, receives the answer, ships it. No hypothesis is ever formed before the AI’s output arrives, so there is no attempt for a correction to attach to — the loop never opens in the first place. This is the largest group, the default mode most tools are designed to encourage, and the pattern behind the 17-point comprehension gap in the Anthropic study: not misuse, just the path of least resistance, taken by people with every reasonable intention of doing good work.

• The Ghost Editor. Forms a genuine attempt, but works inside a tool that silently repairs the parts that were wrong before they’re consciously registered — autocomplete finishing the syntax, a smoothing pass cleaning the prose, an auto-fix patching the edge case. An error occurred. Nobody noticed, including the person who made it. This is the subtlest version of the vacuum, because from the inside it feels like fluency, not assistance.

• The Deferred Debtor. The loop is not erased so much as pushed downstream, past the point where it’s still useful. An AI-drafted client email goes out with a subtly wrong tone; the client doesn’t reply for six weeks, and the account quietly cools for reasons nobody connects back to that one email. A hiring rubric drafted with AI screens out a strong candidate; the gap in the team’s output shows up as a vague sense that “we’re just not hiring as well as we used to,” eighteen months and several unrelated decisions later. The feedback still technically arrives. It arrives too late, and too disconnected from its cause, to teach anyone anything.

• The Sparring Partner. The protective pattern, and the rarest. Forms an independent attempt first — a real hypothesis, a real first draft, a real diagnosis — before consulting AI at all, then uses the AI’s response specifically as a check against that independent view, actively seeking out the disagreement rather than the confirmation. This maps closely onto the specific interaction patterns the Anthropic researchers found did preserve learning: participants who asked for explanations rather than finished answers, or asked conceptual questions rather than requesting generated code, scored no worse than the unassisted control group. The loop survived because they kept forcing themselves back into the “attempt” step the tool was otherwise happy to skip on their behalf.

Most people are not consistently one archetype. A single professional is often a Sparring Partner in their specialism and a Vending Machine User everywhere else — which is, if anything, the right allocation of a scarce resource: the friction is worth defending exactly where you’re still building expertise, and worth discarding everywhere you already have it.

What to actually do about it

• For individuals, the operating rule is simple to state and hard to practise: identify which parts of your job you are still learning, as opposed to already competent at, and change your AI usage pattern accordingly. In the domain you’re building, form your own attempt before you open the assistant, and when you do consult it, ask for critique of your reasoning rather than a replacement for it — the research is specific here, not vague self-improvement advice: explanation-seeking and concept-checking preserved learning outcomes; answer-generation didn’t. In domains where you’re already expert, hand over the routine freely. The friction is only valuable where the loop is still under construction.

• For managers, the practical move is to stop treating “time to completion” as a sufficient quality metric, because it is exactly the metric the Feedback Vacuum leaves undisturbed while the underlying skill erodes underneath it. Build in comprehension checks that don’t allow the assistant in the room — a debugging question asked cold, a diagnosis defended without the tool’s suggestion visible, a draft explained line by line. This will feel like an unnecessary tax on people who are visibly delivering good output on schedule. It is not a tax. It is the only way to distinguish real judgment from a well-camouflaged vacuum before the day arrives that the tool is wrong and no one in the room can tell.

• For leaders, the Ganuthula model’s central warning belongs in every AI adoption plan: the performance gain and the skill loss are on different clocks, and the gain will always report first. A rollout that looks unambiguously successful at the six-month mark can be quietly hollowing out the judgment it will need at the three-year mark, and the metrics that would catch this — comprehension without assistance, debugging under time pressure, an independent second read — are not metrics most AI adoption dashboards currently track, because nobody built the dashboard with the vacuum in mind.

• For hiring, treat “fast with AI tools” and “skilled” as two different claims requiring two different tests. A candidate who has spent two years as a fluent Vending Machine User will interview well on anything the assistant can be present for and poorly on anything it can’t — and the gap will not show up unless you specifically design for it to show up, by asking them to debug, diagnose, or explain with the tool switched off.

  ---

The tuition nobody budgeted for

There is a version of this argument that lands as nostalgia for productive suffering, and that is not what this is. Nobody is arguing for friction as a virtue in itself, or for a return to debugging without a debugger, reading film negatives instead of digital scans, or drafting on a typewriter. The argument is narrower and, hopefully, more useful: friction was never decorative. It was the delivery mechanism for a signal that competence has always depended on, and removing the friction without replacing the signal does not make the signal unnecessary. It just makes it disappear, silently, exactly where you can least afford not to notice.

The uncomfortable truth of the Feedback Vacuum is that it is invisible by construction. Every measure an organisation typically watches — output quality, time to completion, error rates on delivered work — can look flat or improving while the underlying capacity to produce that output without assistance quietly erodes underneath it, because none of those measures were built to detect the absence of a loop that used to run silently in the background of ordinary work. You do not find out you are in the vacuum by looking at your metrics. You find out the day the assistant is wrong, or unavailable, or simply not in the room — and discover, in that moment, whether the last several years actually taught you anything, or just taught you how to ask.

Priya Nair had spent 14 years handling claims at a mid-sized insurer in Leeds, the last six as the person colleagues brought the strange ones to. Her queue in 2024 held sixty-odd claims a day. Fifty of them were routine — a rear-end collision, a burst pipe, a stolen bike. Ten required actual thought. Perhaps one was genuinely strange: dates that didn’t quite line up, an invoice from a garage that existed mostly on paper, a claimant whose story was a little too polished.

She caught those. She caught them because — although she would never have put it this way — the fifty routine claims were her instrument panel. Ten thousand ordinary claims a year had taught her exactly what ordinary looked like: how a genuine invoice was formatted, when a real claimant rings twice, what an honest delay sounds like. The strange ones announced themselves by their wrongness against a baseline she was refreshing daily without noticing she was doing it.

In early 2025, the insurer deployed claims automation, and it was good. Within a year it was settling the routine end-to-end: intake, verification, payment, done. Priya’s queue fell from sixty to fourteen. Management called it “elevating our people to higher-value work”, and to be fair to them, they meant it.

Every one of the fourteen was hard. Complex liability. Distressed claimants. Suspected fraud — flagged, now, by a model, rather than by the prickle at the back of Priya’s neck.

Eighteen months in, Priya had noticed two things. The first was that she was more tired than she had ever been in the job, in a way that weekends didn’t fix. Every case was a summit; there were no valleys between them any more. The second was harder to name: her sense of what normal looked like was going soft. When a flagged claim was strange, she increasingly could not say what it was strange against. The machine handled the ordinary — and the ordinary, it turned out, had been the source of her radar.

Priya’s job had become harder and shallower at the same time. Nobody designed that. It is simply what is left when you subtract the routine.

A warning from 1983

The clearest description of what is happening to Priya was published more than four decades ago. In 1983, the cognitive psychologist Lisanne Bainbridge wrote a short paper in Automatica called “Ironies of Automation”, about industrial control rooms and autopilots. It has been quietly famous among human-factors researchers ever since, and it reads today like a memo about the 2026 workplace that arrived forty-three years early.

Bainbridge’s argument was structural, not sentimental. When designers automate a process, they automate the parts that can be specified — the routine, the repeatable, the easy. Whatever cannot be specified is left to the human. That produces two ironies. The first: the human is left holding precisely the tasks the designers could not figure out how to automate — an arbitrary collection of the hardest, rarest, most ambiguous work. The second is crueller: the automation simultaneously removes the routine practice through which the human maintained the very skills those hard, rare moments require.

“By taking away the easy parts of the task, automation can make the difficult parts of the human operator’s task more difficult.”

— Lisanne Bainbridge, “Ironies of Automation”, Automatica, 1983

Aviation lived this first, and paid for the lesson in accident reports. Autopilot flies the cruise better than any human, so pilots stopped hand-flying — and their manual instincts, maintained for decades by thousands of uneventful hours, began to decay. By 1997, American Airlines’ training department was warning its pilots about becoming “children of the magenta line” — crews so accustomed to following the automation’s guidance that they struggled in the moments it handed the aircraft back. The warnings became case files. When Air France 447 was lost over the Atlantic in 2009, investigators found a crew confronted, at altitude, with a situation the automation had always previously handled — and manual instincts that thousands of uneventful autopilot hours had quietly let atrophy. The industry’s eventual answer was telling: regulators and airlines began mandating manual-flying practice. Deliberately reintroduced routine, at real cost, as maintenance for the instincts that the exceptional moments require.

Aviation could do that because it is one industry, tightly regulated, with its failure modes written up by accident investigators. Knowledge work in 2026 is running the same experiment across the entire economy at once — with no regulator, no accident report, and no mandated hand-flying.

The machine keeps the routine. You keep the residue

What the current generation of AI absorbs is, almost by definition, the routine: the standard ticket, the reconciliation, the first draft, the ordinary claim, the meeting notes, the status report. That is what it means for work to be automatable — it is regular enough to be learned from examples. So the machine eats from the middle of the bell curve outward, and what it leaves behind is the residue: the complex, the emotional, the ambiguous, the genuinely novel.

Klarna ran this experiment in public. In 2024 the company announced its AI assistant was doing the work of 700 customer service agents. In 2025 came the quieter second act: the company started hiring humans again. CEO Sebastian Siemiatkowski’s post-mortem, delivered to Bloomberg, was admirably blunt — “We focused too much on cost. The result was lower quality.” The arrangement Klarna landed on is the template now being stamped across the economy: the AI keeps the routine volume; a smaller pool of humans handles the complex, the emotional and the judgment-heavy. Klarna’s reversal was reported as a retreat from automation. It was nothing of the sort. It was the Exception Economy finding its stable form.

The economics are seductive, and they are real. The landmark study of generative AI in the workplace — Brynjolfsson, Li and Raymond’s “Generative AI at Work”, published in The Quarterly Journal of Economics in 2025 — followed 5,179 customer support agents and found the tool raised productivity 14% on average, and a remarkable 34% for novice workers, because it hands the newest people the accumulated playbook of the most experienced. Routine work, it turns out, is exactly the work AI does best on a human’s behalf. Which is why it is disappearing first.

There is a second stream feeding the residue, and it is easy to miss: the machine’s own output. Automated work does not arrive finished; it arrives plausible. Somebody has to read it, judge it, and catch the small fraction that is confidently wrong — and that somebody is doing exception work too, because every review is a judgment call with no routine attached to it. The Upwork Research Institute found 39% of AI-using employees now spend more time reviewing AI-generated content than they expected the tools to save, and nearly half cannot see how the promised productivity gains are supposed to materialise. The routine did not simply leave the workday. It came back as verification — the one form of routine that offers neither rest nor reps.

But watch where everything flows. Two conveyor belts now run through every AI-adopting organisation. The routine flows to the machine. The exceptions flow to the most capable human available. Reporting through 2026 keeps surfacing this second routing rule: when agents handle the clean, documented process, the exceptions they generate get escalated to the strongest people on the floor — usually senior staff who once managed the whole process and now manage only its failures. Their role quietly shifts from execution to escalation. Harvard Business Review put the resulting pattern on its front page in February 2026 with a headline that could serve as this edition’s subtitle: AI doesn’t reduce work — it intensifies it.

And the remaining human work runs at machine tempo. Microsoft’s 2025 Work Trend Index, drawing on 31,000 workers across 31 markets, measured the texture of the modern workday: an interruption every two minutes during core hours — 275 a day — 153 Teams messages and 117 emails daily, meetings after 8pm up 16% year on year. Nearly half of employees, 48%, describe their work as chaotic and fragmented. That is what a workday feels like when the valleys have been removed from it.

The Drudgery Dividend: what the boring work was paying for

Here is the accounting error at the centre of almost every AI business case: routine work was booked as pure cost. Minutes to be recovered, salaries to be redeployed, toil to be eliminated. What never appeared on any ledger is that the routine was simultaneously paying four subsidies to the organisation that hosted it — call it the Drudgery Dividend. Cancel the routine and you cancel the dividend, whether you noticed you were receiving it or not.

Recovery. Sustainable jobs have rhythm: hard calls and easy stretches, peaks and valleys. The valleys were not wasted time — they were where the peaks got metabolised. Emergency medicine, the one profession that has always done all-exception work, learned this the hard way and engineered around it: shift limits, rotations, decompression, staffing ratios. Knowledge work is now importing the emergency department’s intensity with none of its safeguards. The strain is already measurable: the Upwork Research Institute found that while 96% of C-suite leaders expect AI to lift productivity, 77% of employees using AI say it has increased their workload — and 71% report burnout. By 2026, studies were finding frequent AI users reporting markedly higher burnout than non-users — the opposite of the promise, and exactly what you would predict if the tools were stripping the recovery out of the workday.

Rehearsal. Skills are not stored; they are maintained. The easy reps were the maintenance schedule — the ordinary cases through which a professional’s pattern library stayed current without anyone calling it training. Bainbridge’s second irony, operating at economy scale: the same automation that handles the routine removes the practice that kept humans capable of the exceptions. The QJE study’s most striking number has a shadow side here. If a novice with AI performs near an experienced worker’s level on routine work, organisations will conclude the routine no longer needs to be humanly practised at all. Those were the reps the next generation of instinct was going to be built from.

Radar. Anomaly detection is a by-product of exposure to the normal. Priya could spot the fraudulent claim because she had processed ten thousand honest ones; auditors, security analysts, editors and underwriters all train their sense of wrongness on volume — ordinary volume. Security operations teams discovered this early: hand the tier-one alert queue to a model, and within a year the senior analysts reviewing its escalations have lost the ambient feel for the network’s ordinary weather — the baseline hum against which an intrusion used to stand out. Route the ordinary through the machine, and the human reviewing its escalations is comparing them against a fading memory of what normal looked like. The twist is genuinely cruel: the better the automation gets, the less often its humans see the ordinary, and the weaker their radar becomes — precisely as the cases reaching them get stranger.

Ramps. The easy work was the on-ramp. I wrote in May about the apprenticeship pipeline collapsing as AI absorbs entry-level tasks; the Exception Economy generalises the problem. Ramps were never only for juniors — they were how career switchers crossed domains, how returners rebuilt confidence, how the newly promoted learned an adjacent craft sideways. A workplace made entirely of hard cases has no shallow end for anyone. Every entrance is a cliff edge.

The four job shapes of the Exception Economy

Subtract the routine and the remaining human roles start collapsing into four recognisable shapes. Most organisations contain all four already, unnamed.

• The Escalation Magnet. The most capable person on the team, to whom everything the machine cannot handle now flows. Their reward for excellence is more exceptions; their calendar is other people’s edge cases; their job title still describes a role that no longer exists. They are the first to burn out — and when they leave, their queue reroutes to the next most capable person, who inherits both the workload and the trajectory.

• The Standing Reserve. Retained “for oversight”. They review machine output, approve, and wait. Their risk is not overload but hollowing: they are Bainbridge’s control-room operator, monitoring a system that almost never needs them, their skills decaying in place — until the day the system fails in a way that requires everything they used to be able to do.

• The Emergency Generalist. Common in smaller firms: the human who spans whatever the agents cannot do that day. Half-finished automations, weird vendor cases, the customer who insists on a person. They context-switch across domains at machine tempo and master none of them. Microsoft’s every-two-minutes interruption figure is not a statistic to them; it is a biography.

• The Valley Keeper. The rarest shape, and the only deliberately designed one: someone whose role intentionally retains a flow of ordinary work — the audit sample, the manual Friday pass, the rotation through routine cases the machine could have handled cheaper. On a cost spreadsheet they look like waste. They are the last people in the building whose sense of normal is current, whose reps are fresh, and next to whom a junior can still learn. Organisations that keep Valley Keepers are not being sentimental. They are maintaining the baseline everyone else’s judgment silently depends on.

  ---

What to do about it

• For individuals, the shift is to treat recovery and routine as skill maintenance rather than slack. Schedule valleys as deliberately as you schedule meetings. Keep a personal flow of ordinary reps in your core craft — write the ordinary brief yourself sometimes, work a handful of unremarkable cases end-to-end each week — not from nostalgia, but because your radar and your instincts are calibrated on exposure you are no longer getting by default. And track your own exception load: count what share of your week is escalations. When it approaches all of it, the problem is the role’s design, not your resilience — and the negotiation you need is about the role.

• For managers, the emergency department is the manual. Rotate people on and off the escalation queue the way hospitals rotate trauma shifts, and treat sustained exception load as an occupational exposure to be measured and capped, not a badge of seniority. Then maintain the baseline deliberately: route a sample of ordinary cases through human hands even when the machine could handle them — auditors have worked this way for a century — and protect real routine work for development purposes, priced honestly as training rather than disguised as inefficiency.

• For leaders, the ask is to put the Drudgery Dividend into the automation business case itself. The four subsidies do not disappear as needs when the routine disappears as work; they become unfunded liabilities. Recovery now costs headroom. Rehearsal now costs mandated reps. Radar now costs sampled volume. Ramps now cost deliberately inefficient junior work. An automation plan that books the savings without booking these replacement costs is not a productivity plan — it is a deferral schedule, and the interest compounds in the working lives of your best people until it gets paid all at once, in resignations. Klarna paid it in public. Most organisations will pay it quietly.

• For hiring, stop writing roles that are 100% exception handling and calling them senior. Screen for ambiguity tolerance, yes — but a role designed without valleys will consume whoever you hire into it, and the more capable they are, the faster the Exception Economy will find them. The interview question worth adding is not about coping with pressure — it is about what the candidate does between peaks, because in the role you are designing, there is no between unless you build one.

  ---

The part nobody budgeted for

The promise was that AI would free people for “more meaningful work”, and it is keeping that promise with a precision nobody requested. Meaning, distilled, at machine tempo, with all the filler removed. It turns out the filler was doing something. A workday of pure significance is not a gift; it is a load rating exceeded — and the people carrying it are, by the routing logic of the whole system, the ones you can least afford to lose.

The uncomfortable truth of the Exception Economy is that the boring parts of the job were never the obstacle to the best work. They were the price the job was quietly paying to keep people capable of the best work — the recovery that made intensity survivable, the rehearsal that kept instincts current, the radar that noticed what the machine could not, the ramps the next generation climbed in on. AI has stopped paying that price. The need did not go away.

Someone still has to pay it. Right now, by default, it is being paid in the working lives of your most capable people — invisibly, until the Tuesday the machine meets something it has never seen and hands it, at last, to a human whose radar has faded, whose reps have lapsed, and whose last valley was eighteen months ago. That is the moment an organisation discovers what the routine was really for. The wise ones will have discovered it on purpose, earlier, and paid for it while it was still cheap.

Elena Bosch had underwritten commercial credit for eleven years at a mid-sized bank in Rotterdam. Colleagues who worked with her described her as “unnervingly right” — not infallible, but right often enough, and wrong in ways that were legible even to her, that people trusted her instinct on files that didn’t fit the standard scoring model. Her default-rate predictions had landed within two points of actual outcomes for six consecutive years. She could tell you, unprompted, the three times in the past decade she had been badly wrong, and exactly what she had missed each time.

In early 2025, the bank rolled out an AI underwriting assistant to the commercial lending team. It read financials, cross-referenced sector risk data, and produced a recommendation with a confidence score attached. Elena adopted a habit almost immediately, without being told to: she read the file cold, wrote down her own view — approve, decline, or flag — before opening the tool, and only then checked it against the model’s output. When they agreed, she moved fast. When they diverged, she went back into the file to find out why.

Her colleague Jorik, two desks over, adopted a different habit, just as unconsciously. He opened the file and the tool at the same time. He read the model’s recommendation first, then skimmed the underlying file to confirm it looked reasonable, then approved. It was faster. It felt, if anything, more rigorous — he was, after all, reading the same financials Elena was, just in a different order.

Eight months later, a portfolio quality review turned up something that unsettled the risk committee more than any single bad loan could have. Elena’s calibration — the statistical tightness between how confident she reported being in a decision and how often that decision proved correct — hadn’t moved from her pre-AI baseline. Jorik’s had come apart. On the subset of cases where the model’s recommendation later proved wrong, Jorik’s own accuracy had fallen sharply — but his self-reported confidence in those same decisions remained exactly as high as it had always been. He had no idea anything had changed. Nobody had told him his judgement was slipping, because the correction signal that would normally tell a professional that — the felt discomfort of being wrong, attributed clearly to your own call — had never arrived. The model was right often enough, and confident enough in its wrongness, that Jorik’s errors were absorbed silently into decisions that felt, from the inside, exactly as sound as they always had.

This is the Confidence Illusion. It is not a story about AI making people less capable. Every available measure of Jorik’s underlying analytical skill — his ability to read a balance sheet, to spot the sector risk his manager wanted him to spot — remained intact. What broke was something upstream of skill: the loop that let his confidence track his own accuracy. And once that loop breaks, a professional can be simultaneously skilled and structurally unable to know it.

Why confidence used to mean something

To understand what has broken, it helps to understand what calibration actually is, and where it comes from. It is not a personality trait. It is a trained response, built through a specific, repeatable mechanism: a person forms a judgement under uncertainty, reality responds, and the gap between the prediction and the outcome — assuming the person actually notices it — recalibrates the next prediction.

This is, essentially, the mechanism behind every serious account of how expertise develops. Gary Klein’s decades of research into naturalistic decision-making — how firefighters, ICU nurses, and fighter pilots make fast, accurate calls under time pressure — found that expert intuition is built from exactly this loop, repeated thousands of times, until pattern recognition becomes near-instantaneous. Philip Tetlock’s forecasting research, most visible in the “superforecasters” work, found that the individuals who consistently outperformed professional analysts on geopolitical predictions shared one habit above all others: they tracked their own predictions against outcomes, relentlessly, and adjusted. Calibration, in Tetlock’s data, was learnable — but only through that unforgiving feedback loop. People who never checked their predictions against outcomes stayed miscalibrated indefinitely, regardless of how experienced they became.

The common thread across all of this research is that confidence was never supposed to be a free-floating feeling. It was supposed to be downstream of a track record — built, case by case, through the friction of being wrong and finding out. Professionals who worked without any decision-support tool got this feedback by default, because there was no alternative: they had to form a view, because nothing else was going to form one for them, and reality corrected them whether they liked it or not.

AI has not removed reality’s ability to correct people. It has removed the requirement that a person form an independent view before consulting a second opinion — and it turns out that requirement was doing almost all of the calibrating work.

The mechanism: how offloading breaks the loop

The distinction the research keeps drawing is between two different things people can offload to AI: memory and judgement. Offloading memory — letting a tool hold facts you’d otherwise need to recall — has a long history and a relatively mild effect on calibration; people have used notebooks, databases, and search engines for this without losing their grip on how confident they should be in their own conclusions.

Offloading judgement is different. A 2026 study presented at the CHI Conference on Human Factors in Computing Systems — “Accurate but Not Confident or Confident but Not Accurate? Cognitive Offloading Impairs Confidence Calibration in Human-AI Teams” — tested this distinction directly. Participants who worked entirely unaided showed the tightest alignment between their stated confidence and their actual accuracy of any condition in the study. Participants who offloaded judgement specifically — not memory, but the act of forming an initial view — showed the sharpest overconfidence effect measured. Combined offloading of both memory and judgement produced a different, but equally distorting, metacognitive bias. The finding was blunt: the moment a person stops forming their own view before seeing an external recommendation, their stated confidence stops being a reliable signal of anything.

This connects to a separate and older body of research on automation bias and complacency. A recent review in the journal AI & Society offers a useful, precise definition of the mechanism: complacency is what happens when trust closes the gap between “I believe this system works” and “I no longer evaluate whether it worked this time.” That second clause is the entire mechanism. It is not that people trust AI too much in the abstract. It is that, case by case, they stop running the small internal check — does this feel right to me, independent of what the tool says — that used to catch errors before they became decisions. The review notes that this collapse is especially likely to occur when a tool’s recommendation aligns with a person’s initial instinct, which creates a reinforcing loop: agreement breeds trust, trust breeds less checking, less checking means agreement is confirmed more often because disagreement is never investigated.

Complacency happens when trust closes the gap between “I believe this system works” and “I no longer evaluate whether it has worked this time.” This is where users catch errors, flag edge cases, and discover the AI’s limitations — and where automation-assisted decision-makers stop doing so.

— Automation bias in human–AI collaboration, AI & Society, 2025–2026

The mechanism, stated plainly: an unaided professional forms a view, gets it wrong sometimes, and feels that wrongness directly — this is uncomfortable, and discomfort is what recalibrates confidence. An AI-assisted professional who skips the independent view never generates the raw material discomfort needs. There’s nothing for the wrongness to attach to. It gets absorbed into a decision that, from the inside, felt entirely reasonable — because it was never really the professional’s own decision to feel wrong about.

What gets lost: two faces of the same failure

The workforce data suggests this mechanism is producing two distinct, opposite-looking outcomes — and it’s worth being precise about why they’re actually the same failure.

ManpowerGroup’s 2026 Global Talent Barometer, based on interviews with nearly 14,000 workers across 19 countries, found that regular AI use jumped 13% in a single year, to 45% of the workforce — while confidence in using the technology fell 18% over the same period, the first overall decline in worker confidence the barometer had recorded in three years. The drop was steepest among the most experienced workers: down 35% among Baby Boomers, 25% among Gen X. And yet 89% of workers still report confidence in the skills required for their current role. This is not a story about people losing faith in their own competence generally. It is a much narrower and stranger story: people are losing trust specifically in their own independent judgement, in the moments where that judgement now runs alongside a machine’s.

A separate 2026 study, reported by the American Psychological Association, adds a piece that clarifies the picture further: overreliance on AI at work does not appear to measurably reduce raw cognitive ability. What it erodes is confidence in independent reasoning, and people’s sense of ownership over their own ideas — their felt sense that a conclusion is actually theirs, rather than borrowed and merely endorsed. Researchers flagged the long-term risk plainly: not that AI use makes people less intelligent, but that some professionals become less engaged in the kind of effortful, generative thinking that produces genuinely novel judgement — because the tool has quietly become the place where the thinking happens first.

Put the two findings together and the shape of the problem sharpens. One group — the Jorik pattern — offloads the initial judgement entirely, stops noticing when they’re wrong, and drifts toward silent overconfidence that nobody, including them, can see from the inside. A second group — visible in the Manpower data, concentrated among the most experienced workers — retains the instinct to form an independent view, notices that view increasingly gets second-guessed or overridden by a fluent machine recommendation, and starts to doubt instincts that were never actually unreliable. Both groups have lost the same thing: a working relationship between what they believe and what is true. One believes too much. The other no longer knows what to believe. Neither can currently tell you, reliably, which of their own judgements to trust — which is precisely the capability the feedback loop used to provide.

Three archetypes

Three professional postures are visible in this transition, and naming them helps predict where a given person or team is heading before a performance review or an audit catches it.

• The Silent Drifter has adopted the Jorik pattern without noticing. They read the AI’s output before forming their own view, or alongside it rather than before it, and their confidence has not moved even as their independent accuracy — measurable only by looking at the subset of cases where the AI later proves wrong — quietly declines. They are, by every self-report metric, doing fine. The gap between their confidence and their accuracy is invisible to them by construction: nothing in their daily experience produces the discomfort that would reveal it. This is the most structurally dangerous archetype, because it self-conceals; the only way to detect it is to deliberately audit decisions after the fact, which almost nobody does until something goes wrong at scale.

• The Doubt Spiral is the mirror image, and disproportionately made up of experienced professionals — the ManpowerGroup data’s Boomer and Gen X respondents. Their independent judgement is intact, arguably as strong as it has ever been, but it now runs constantly alongside a fluent, confident machine output that occasionally disagrees with it. Each disagreement produces a small, corrosive moment of self-doubt, even when the human was right — because the AI’s confidence is expressed with the same fluent certainty whether it is correct or not, and there is no longer an obvious signal for which voice to trust. Over time, this group’s stated confidence in their own judgement erodes even though their underlying accuracy has not. They are, in a real sense, the collateral damage of a workplace that has not built any mechanism for telling people when their own instinct was the right one.

• The Calibrated Checker is the smallest and most valuable group, and their habit is almost embarrassingly simple: they form an independent view before they open the tool, every time, as a fixed discipline rather than an occasional practice. When the AI agrees, they move fast, exactly as anyone would. When it disagrees, they treat the disagreement as diagnostic information rather than as an automatic override in either direction — sometimes the AI is right and they update; sometimes they are right and the AI’s confident output was simply wrong in a way that fluent language obscured. Crucially, they track the outcomes of these disagreements over time, which means their confidence keeps being disciplined by the exact mechanism that built expertise in the first place. They are not smarter than the Silent Drifters or more skilled than the Doubt Spirals. They have simply refused to let the tool remove the one step — form your own view first — that made confidence mean something.

  ---

What this means in practice

For individuals, the practical implication is almost aggressively simple, which is part of why it is so easy to skip: write down your own view, even briefly, before you open the tool. Not because your first instinct is always right — often it won’t be — but because the act of committing to a view, and then finding out whether it held up, is the only mechanism that keeps your confidence tethered to your actual accuracy. Skipping this step doesn’t just cost you a marginally worse decision today. It costs you the raw material your judgement needs to stay calibrated over the next decade.

For organisations, the implication is that AI adoption without a parallel discipline for preserving independent judgement is quietly manufacturing both of the Confidence Illusion’s failure modes at once — a population that is dangerously overconfident in eroding judgement, sitting alongside a population that has stopped trusting judgement that was never actually impaired. Neither failure shows up in standard productivity or output-quality metrics in the short term, because AI-assisted output is, on average, good. It shows up later, unpredictably, in the moments the tool is wrong and nobody in the workflow retained the instinct — or the confidence — to catch it.

The practical fix organisations are experimenting with is structural rather than motivational: build the independent-judgement step into the workflow itself, so it doesn’t depend on individual discipline. Some firms are formalising this as a two-step approval process — record a human recommendation before the AI recommendation is visible, then compare. It is a small amount of friction, deliberately reintroduced, in a system that has spent two years trying to remove friction wherever it can be found. The firms doing this are, in effect, treating the independent-judgement step the way clinical trials treat a blind assessment: not as inefficiency, but as the only mechanism that keeps the measurement honest.

For hiring and development, the shift is from screening for confidence — historically an easy, visible, interview-friendly proxy for competence — to screening for calibration, which is much harder to fake and much more diagnostic. A candidate who can describe, specifically, a time they were confidently wrong and how they found out is demonstrating the exact muscle the Confidence Illusion is quietly atrophying across the workforce. A candidate who cannot produce that story, or who describes only times they were confidently right, is a much weaker signal than it used to be — high self-reported confidence, on its own, no longer reliably predicts good judgement in an AI-assisted environment. It may increasingly predict the opposite.

The closing uncomfortable truth

There is a final, uncomfortable feature of the Confidence Illusion worth naming directly: it is nearly undetectable from the inside. A Silent Drifter does not feel like someone whose judgement is eroding. They feel exactly like someone who is right most of the time, because they are — the AI is usually right, and their approval of its usually-right output feels, moment to moment, indistinguishable from good judgement. A Doubt Spiral does not feel like someone whose instincts are being unfairly eroded. They feel like someone who is simply, reasonably, less sure than they used to be, in a world that has gotten more complicated. Neither has access to the information that would tell them which category they’re in, because the information that used to provide it — the direct, attributable, felt experience of being wrong — has been quietly rerouted around them.

The people whose confidence still means something are not, on average, more talented than everyone else. They have simply refused to let the tool remove the one habit that made confidence a real signal in the first place: forming a view of their own, before they see anyone else’s — including a machine’s — and finding out, deliberately and repeatedly, whether they were right.

Everyone else is walking around with a level of confidence that has quietly stopped being about anything at all. The unsettling part is that there is currently no way to tell, from the outside or from the inside, who is who — until the moment it matters, and the gap between what someone believed and what was actually true finally becomes visible to everyone except the person who should have seen it first.

The consultant who could not brief a machine

Marco had spent 14 years at a mid-size strategy consultancy in Zurich. In those 14 years, he had built a reputation for something genuinely rare in his profession: he could walk into a client situation he had never seen before, ask 3 questions, and have a thesis. He was, by every metric that mattered in consulting, an excellent thinker. Partners sought him for the ambiguous engagements. Clients requested him by name.

In February 2026, his firm deployed AI agents across the practice. Every consultant received access to the same system and the same three-week onboarding. Three months later, the pilot review threw up a pattern that the managing partner described as “inexplicable.” The highest-performing consultants in the pilot were not the most experienced. They were not the most technically capable. By most internal performance ratings, they were unremarkable — except for a single habit that their colleagues had always found slightly excessive.

They wrote very long, very precise briefs. Before they touched the agent, they spent the first hours of any engagement writing a detailed problem statement: naming not just what they needed, but why they needed it, what they already knew, what constraints were non-negotiable, and what judgment calls the agent should leave to them.

Marco wrote briefs the way senior partners write briefs: short, confident, jargon-efficient. The kind a junior colleague who had worked with him for years could interpret without clarification. The kind that presumed substantial shared context.

The agent had no shared context. It took what it was given and produced, from it, a comprehensive, confident, thoroughly wrong analysis.

This is the Legibility Deficit. It is not a story about AI fluency. It is a story about the sudden, brutal surfacing of a skill gap that the social infrastructure of knowledge work had spent decades papering over — and the professional advantage now accruing to the people who, for one reason or another, never developed the habit of being vague.

Why knowledge work was never legibility-tested before

Legibility — the ability to make your thinking visible, structured, and interpretable by others — has always been a valued skill. The best writers, strategists, engineers, and lawyers are, among other things, unusually good at externalising their reasoning in a form that others can act on without supplementary interpretation.

But the operative phrase is “supplementary interpretation.” Human colleagues are astonishingly good at it. They infer the unstated context. They read the relationship, the tone, the institutional history. They ask the clarifying question. They complete the half-formed thought. They compensate, in thousands of micro-interactions every day, for the gaps between what was said and what was meant.

This compensatory infrastructure is so pervasive and so effective that it is essentially invisible. It is the reason that knowledge work organisations can function despite the fact that most of the communication within them is, by any rigorous standard, massively underspecified. “Can you take a look at this?” “We need to think about the client situation.” “The numbers feel off.” These instructions contain almost no information — yet work gets done, because the person receiving them can infer, from context, relationship, and institutional knowledge, approximately what is required.

This infrastructure has a name in organisational theory: tacit knowledge. It is the accumulated, uncodified understanding that lives in the social fabric of an organisation — in its norms, its shared vocabulary, its accumulated history of past projects and past failures. Michael Polanyi described it in 1966 as “the kind of knowledge that we know more of than we can tell.” Most of what enables knowledge workers to be productive is tacit. And none of it is transferable to an AI agent through a brief.

This is not a criticism of AI agents. It is a precise description of the challenge they surface. A human colleague carries tacit knowledge that lets them compensate for an underspecified instruction. An AI agent takes the instruction as given. It has no relationship with you, no memory of your previous projects, no sense of the institutional norms that would tell a human reader what “take a look at this” actually means in your firm. It produces output from exactly what you provide — and what most professionals provide, once the human compensators are removed, turns out to be remarkably thin.

The AI-agent transition is, at its core, a legibility audit. It is the first time in the history of knowledge work that the quality of thinking — specifically, the ability to externalise that thinking precisely enough for it to be acted on without inference — has been systematically tested, at scale, across an entire workforce, in real time. The results are uncomfortable.

The mechanism: how legibility becomes leverage

The economist’s version of the Legibility Deficit is simple. When AI agents can do the execution, the returns flow to whoever can most precisely specify what should be executed — because that specification is the input the agent depends on and cannot generate on its own.

Specification — the act of making intent explicit, naming constraints, articulating context, and identifying the judgment calls that require human decision — is legibility. And like all scarce complements to an abundant input, it commands a premium that rises as the input becomes more abundant.

A 2025 paper tracking skill emergence and obsolescence across more than 12 million job postings in the LLM era found that the fastest-growing skill categories were not technical. They were what the authors called “articulation-adjacent”: structured communication, specification writing, problem framing, context-setting, and the ability to translate a complex, multi-constraint situation into an actionable instruction. These skills were growing at three to five times the rate of even AI-specific technical skills, and they were commanding the sharpest wage premiums of any category tracked.

McKinsey’s 2025 State of AI report put numbers on the organisational version of this. Firms that had developed systematic legibility practices — structured brief formats, explicit context-capture workflows, defined escalation criteria for judgment calls — showed “significantly higher AI performance and adoption rates” than those that had provided tool access and assumed use would follow. The variable that predicted outcome was not the AI system. It was the quality of the human specification the AI system received.

A 2025 cross-sector employer survey found that 91% of firms reported effective AI use required genuine language competence — not vocabulary, but the ability to state intent, name constraints, and identify what “correct” looks like before the work begins. 92% said this capacity was becoming more important in their organisations as AI tool adoption increased.

The paradox — and there is always a paradox — is that the tool designed to make work more productive has made the quality of the human thinking that precedes the work more consequential, not less.

What the social infrastructure was hiding

The most uncomfortable implication of the Legibility Deficit is that it retroactively reveals how much of what passed for “strong communication” in professional settings was actually the recipient’s effort being invisibly compensated for.

Consider the anatomy of a typical knowledge-work brief. A partner at a law firm sends a junior associate a one-line instruction: “Look into the precedent here — I want to understand our exposure.” What has actually been communicated? The area of law (implied by context). The specific risk the partner is concerned about (not named, but inferable from the matter). The level of depth required (not specified, but guessable from the time available). The format expected (assumed based on firm norms). The judgment calls the associate should make independently versus escalate (left entirely to instinct, shaped by relationship).

The associate completes the brief correctly — or at least correctly enough — by deploying all of this tacit knowledge. The partner, watching the work arrive at the right level of depth and in the right format, concludes: “I gave her a clear brief and she delivered.” An AI agent, receiving the same brief, would ask: Which jurisdiction? Which precedent? What specific risk? What format? How much depth? In the absence of those answers, it proceeds with default assumptions that may or may not align with what the partner meant. The output is technically responsive to the instruction and potentially useless to the person who gave it.

The brief was not clear. It was always underspecified. The clarity was being supplied, invisibly, by the recipient’s interpretive effort. AI agents have removed that effort from the equation, and what remains is the instruction itself — which, stripped of its compensatory infrastructure, is often a great deal thinner than the person who wrote it believed.

“Humans often struggle to explicitly articulate their goals and objectives” — and this has been identified as one of the core challenges in designing AI agents that align with human intent, because the social systems that previously compensated for that struggle are absent in an agentic workflow.

— Overseeing Agents Without Constant Oversight, arXiv 2025

The legibility gap is not uniformly distributed across the workforce. It is strongly correlated with seniority — because seniority, in most organisations, is precisely the state in which the most tacit context exists and the most compensatory effort is applied by the people around you. Junior professionals have always had to be more explicit — their colleagues know less about what they mean. Senior professionals have been able to be less explicit for years, because their accumulated institutional capital means everyone around them already knows a great deal about what they mean before they speak.

AI agents invert this hierarchy. The junior professional, who was always forced to externalise their thinking, may adapt faster than expected. The senior professional, whose vagueness has been compensated for by a decade of relationship equity, is discovering that the agent has no relationship equity with them at all.

Three archetypes

It helps to name the professional postures emerging in this transition, because they predict trajectories before the performance reviews do.

• The Tacit Fluency Professional is genuinely excellent at thinking, but their thinking has always been externalised through conversation, relationship, and social iteration rather than through writing. They work best in rooms: they read the situation, adjust in real time, clarify through dialogue. Their intelligence is real, but it is interactional rather than compositional — it lives in the exchange, not the brief. With AI agents, they are discovering that the exchange does not happen first. The brief does. And their briefs, stripped of the interactional channel through which their actual thinking emerges, are thin. Their output quality is declining not because they think less well, but because they are being asked to think in a medium they have never needed to develop fluency in.

• The Legibility Native was always explicit. They started every project by writing the problem statement. They asked the clarifying question before they started work, not during. They specified the success criteria before the brief went out. Colleagues often found this excessive — surely everyone already knew the situation? — but the habit was formed and sustained because it was intrinsically useful to them, independent of the interpretive infrastructure around them. With AI agents, their habit turned out to be exactly the skill the moment required. They did not change what they did. The value of what they had always done changed around them.

• The Legibility Developer has recognised the shift and is deliberately rebuilding their practice. They are learning to write the problem statement before they touch the agent. They are developing brief templates that force explicit constraint-naming. They are reviewing their outputs not just for quality but for specification — asking, after the fact, what the agent would have needed to know to produce this correctly, and building that back into their upfront process. This is learnable. It is harder, the more senior you are, because the habits of implicit communication are deeply grooved — but the ceiling is genuinely high for those who take it seriously.

  ---

What this means in practice

For organisations, the Legibility Deficit changes what the most valuable professional development investment is. It is not AI tool training. It is legibility training — the practice of writing the problem statement before starting the work, naming the constraints explicitly, specifying the desired outcome unambiguously, and identifying the judgment calls that the agent should not make without human sign-off.

These are teachable skills. The practice is not complicated: write the problem statement before you write the prompt. Name the constraints you would usually leave implicit. Specify what “correct” looks like before you ask for the output. Identify the judgment call that the agent should not make without you. These are habits, not talents. They can be built deliberately, at any career stage, by anyone willing to invest the first ten minutes of every task in clarity rather than starting with the prompt.

The structural risk for firms is the same pattern that emerged in every previous wave of automation: the technology gets deployed without the complementary human-skill development, early performance disappointments are attributed to the technology rather than to the human specification it received, and the firms that invested in both the tool and the legibility training pull ahead while everyone else debates the tool quality.

• For individuals, the implication is sharpening fast. Fluency with AI tools is becoming table stakes — a baseline competency, not an edge. The edge flows to the complement: the thing the tool cannot supply and the market cannot sell you. In this case, that complement is legibility. The capacity to state what you need in a form precise enough that a system without your tacit context can act on it correctly is becoming the fastest-growing professional premium in the knowledge economy.

• For early-career professionals, this is clarifying. The discipline that was always demanded of you — be explicit, ask the clarifying question, write the problem statement — turns out to be the discipline the market is now rewarding. The habit of externalising your thinking before you start was not just a professional expectation. It was practising for a world that would eventually require it of everyone.

• For senior professionals, the implication is sharper and less comfortable. The vagueness that felt like authority — the ability to give an underspecified directive and trust that the organisation would interpret it correctly — was always a tax on the people below you. It is now a tax on your own output. Learning to say things precisely, after fourteen years of not having to, turns out to be harder than it looks.

The closing uncomfortable truth

There is a final, generational dimension to this that deserves naming. Legibility is learned through the discipline of writing — through the habit of committing to a position before you know how the room will react, of stating what you need without the benefit of immediate feedback, of externalising your reasoning into a form that has to stand on its own. A professional cohort that forms its habits in an environment where the agent produces the draft and the human edits it may never develop the specification muscle at all — because the muscle is built in the act of writing the brief, not in the act of editing the response.

The organisations now deploying AI agents without a parallel investment in legibility development are not just creating a short-term performance gap. They are potentially creating a long-term capability gap — a cohort of professionals who are fluent with the tools and vague about what they need from them, and who, unlike their senior colleagues, have no accumulated tacit context that even theoretically justifies the vagueness.

This is not inevitable. The window for deciding to get there intentionally is shorter than it looks. The Legibility Deficit is already present in the data. It is already showing up in performance reviews, in AI pilot results, in the quiet reputational shifts inside organisations that have deployed AI agents and are watching, with some puzzlement, who is thriving and who is not.

This is, in the end, a story not about AI but about thinking. AI agents have not changed what excellent thinking looks like. They have made the quality of thinking legible — in a way that the social infrastructure of the workplace had, for decades, allowed people to avoid demonstrating.

The ones who always thought clearly just got very expensive. And the rest of us have a season or two to decide whether to get there deliberately — or to wait and discover, the hard way, what was actually in our briefs all along.

There is a financial services firm in Singapore — call it a composite, because the pattern is now common enough to be a type rather than a case — that spent 2025 acquiring the most advanced AI capability money could buy. Frontier models on enterprise contracts. A dedicated compute allocation that would have been a national-scale resource a decade ago. A new internal platform team. A board deck with a slide titled, without irony, “AI Moat.”

By any procurement standard, they had won. They owned more raw AI capability than any competitor in their market. And for a while, the dashboards agreed: faster reporting, quicker turnaround, lower cost per task.

Eighteen months later, a mid-sized competitor — running older, cheaper, demonstrably less powerful models — was consistently making better calls in exactly the situations that mattered most: the ambiguous ones, the novel ones, the ones with no precedent in the training data. The Singapore firm had more AI. The competitor was getting more from it. And the gap between them was not closing. It was widening, quarter on quarter, at an accelerating rate.

This is the Compounding Divide. It is the most important strategic distinction of the current AI moment, and most organisations are on the wrong side of it without knowing it. The divide is not between firms that have AI and firms that don’t; that race is essentially over, and almost everyone has it. The divide is between firms whose AI gets smarter every time it’s used and firms whose AI simply runs faster while the people around it slowly stop thinking.

The accounting distinction nobody expected from a CEO

In June 2026, Satya Nadella published a long essay — “A frontier without an ecosystem is not stable” — that was viewed more than 28 million times. What made it land was not a product announcement or a benchmark. It was a piece of accounting.

Nadella proposed that two kinds of capital now define a firm in the AI era. The first he called token capital: the AI capability a company builds and owns — its proprietary systems, its data, its evaluations, its tuned and orchestrated workflows sitting on top of foundation models. (The “token” is the unit of text a model reads and generates. It has nothing to do with cryptocurrency.) The second is human capital: the knowledge, judgement, relationships, ingenuity, and pattern recognition of a company’s people.

Then came the line that most readers treated as reassurance and that I read as a warning:

“Human capital does not become less valuable as token capital grows. It only becomes more valuable… Without human direction, you have compute running in circles.”

— Satya Nadella, June 2026

The reassuring reading is: don’t worry, humans still matter. The harder reading — the one Nadella’s own framing actually supports — is that the two kinds of capital behave completely differently as assets. One you can buy. One you cannot. And only one of them compounds. Understanding why is the whole game.

Why token capital converges

Token capital has a seductive property: it is purchasable. You can sign a contract, provision the compute, license the model, stand up the platform, and own meaningful AI capability within a quarter. It is legible to a board, it appears on a balance sheet, and it can be acquired with a decision rather than earned through years of practice.

That same property is its weakness. Anything you can buy, your competitor can also buy — from the same providers, on roughly the same terms, at roughly the same time.

Frontier capability is steadily becoming a utility: metered, broadly available, and increasingly undifferentiated at the point of use. The history of general-purpose technologies is unambiguous on this point. Electrification, in its early decades, conferred enormous advantage on the firms that had it. Then everyone had it, and having it stopped being an advantage: it became the precondition for being in business at all. The same arc ran through enterprise software, through cloud computing, through broadband. The general-purpose layer commoditises. The advantage migrates to the complement, the thing the technology cannot supply and the market cannot sell you.

The economist’s version of this is straightforward: when a valuable input becomes abundant and cheap, the returns flow to whatever scarce input it depends on.

Make compute and frontier models abundant, and the scarce complement is the human judgement that directs them and the proprietary loop that improves them. PwC’s 2026 Global AI Jobs Barometer puts numbers on the migration: the wage premium for AI skills has climbed to 62%, and “professionalised” roles that combine domain judgement with AI are growing roughly twice as fast as “democratised” ones, with 42% faster wage growth since 2021. The market is already repricing the complement. Most corporate AI strategies are still buying the commodity.

The depreciation default

To see why the loop is so unusual, it helps to remember how almost every asset a firm owns actually behaves: it depreciates. A machine wears. A patent expires. A software platform ages into legacy. A trained workforce forgets, leaves, and has to be retrained. The natural tendency of value, in the physical and organisational world, is to leak away. Most of management is, in effect, the work of fighting depreciation — maintenance, retraining, reinvestment, replacement. The default direction of an asset is down.

Token capital obeys this default with unusual speed. A frontier model that is state of the art today is mid-tier within a year and a commodity within two, as the next generation arrives and the price of the previous one collapses. The compute you provisioned is a depreciating capital expense. The platform you built ages the moment it ships. Bought capability does not merely fail to compound — it actively decays, and it decays on the same schedule for you as for every competitor who bought the same thing.

A genuine learning loop is the rare asset that runs the other way. Because it is fed by accumulated judgement and proprietary feedback rather than by hardware or licences, each cycle leaves it slightly more capable than the last. It appreciates through use rather than depreciating despite it. This is what Nadella means when he calls it “unlike most assets.” The strategic error of the Token Buyer is not that they bought a bad asset. It is that they bought a depreciating one and assumed it would behave like an appreciating one — that owning the engine was the same as owning the compounding.

The mechanism: why the loop compounds

Nadella’s real argument is not about either kind of capital in isolation. It is about the engine that links them — what he calls a learning loop. The idea is deceptively simple. Every time your people use your AI systems, both sides should get smarter. People learn from what the AI surfaces. The AI capability improves from the data, corrections, evaluations, and judgement your people feed back into it. Run that loop repeatedly and it accumulates value the way compound interest does — rather than depreciating the way most assets do. Nadella describes the result as “a hill-climbing machine” that becomes “the new IP of the firm.”

This is the part you cannot buy. There is no contract for it. You can only build it, slowly, through use — through thousands of cycles in which a human applies judgement to an AI output, the output improves, the human’s understanding sharpens, and the improved system raises the ceiling on what the next cycle can attempt.

And here is the structural fact that makes the divide so unforgiving: compounding is exponential. A firm that improves its loop by a small percentage each cycle, sustained over hundreds of cycles, does not end up slightly ahead of a firm that simply bought capability and left it static. It ends up exponentially ahead. The distance between them does not grow at a constant rate. It grows at an accelerating one. This is why the Singapore firm’s competitor kept pulling further away rather than being caught: the leader had purchased a level; the competitor had built a slope.

PwC’s data shows what this looks like in aggregate. The top 20% of the most AI-exposed companies achieved average labour-productivity growth of 163% relative to 2018: nearly 5 times higher than the most AI-exposed companies overall. Same access to frontier AI. Radically divergent outcomes. The variable that separates them is not how much AI they bought. It is whether the AI was wired into a loop that compounds human judgement, or simply deployed to run existing tasks faster.

What gets starved

If the loop is the source of all the compounding, then the human judgement that feeds it is the single most important input in the system. Which makes the central irony of the current moment almost unbearable: that input is precisely what most organisations are quietly degrading.

The Microsoft and Carnegie Mellon University 2025 study of 319 knowledge workers found that higher confidence in generative AI inversely correlated with critical thinking. The more a worker trusted the AI’s output, the less they scrutinised it. Participants reported applying no critical thinking at all to a substantial share of their AI-assisted tasks. Read through Nadella’s framework, this is not a productivity footnote. It is the fuel line to the engine being cut. The corrections, the caught errors, the applied judgement — the very inputs that make the loop compound — stop being supplied at exactly the moment the organisation believes it is becoming more efficient.

The training and mentorship picture compounds the problem. PwC found that even as AI skills command a 62% wage premium, more than half the global workforce reported no recent training and 57% lacked access to mentorship. Organisations are accumulating token capital on the asset side while letting human capital — the only thing that makes token capital compound — go uncultivated. They are buying more engine and draining more fuel, and reading the short-term speed as success.

There is a generational edge to this that deserves naming. The corrective judgement that feeds the loop is itself built through years of doing the work — forming positions, being wrong, being challenged, and recalibrating. A junior cohort that arrives into an environment where the AI produces the position and the human merely approves it never builds that judgement in the first place. The independent research is directionally consistent here: studies of cognitive offloading have repeatedly found the effect strongest in younger users, precisely the group entering the workforce now and forming the professional instincts they will rely on for decades. An organisation can therefore be starving its loop on two horizons at once — degrading the judgement of the people it has, and failing to grow the judgement of the people it is hiring. The fuel line is being cut at both the present and the future end.

The Work AI Index 2026 offers the inverse, hopeful signal. Its “high AI achievers” do not simply prompt and accept. They spend more time reviewing and correcting AI output than low achievers (40% of their AI time versus 33%), and they are markedly more likely to deliberately decline to use AI on certain tasks where their own judgement is the better instrument. The differentiator at the individual level is not usage volume. It is the deliberate application of judgement — the behaviour that feeds the loop. The people getting the most out of AI are the ones doing the most thinking around it, not the least.

Three archetypes

It helps to name the postures organisations are adopting, because the labels make the trajectory visible before the financials do.

• The Token Buyers have defined AI as a procurement problem. They measure success by capability acquired and cost removed. Their dashboards are green. Their AI runs their existing processes faster, and they have mistaken that speed for a moat. They are, in Nadella’s phrase, running compute in circles — accumulating a depreciating asset on equal terms with every competitor who signed the same contracts, while the human judgement that might have made it compound goes unexercised. The bill has not arrived, so they believe there is no bill.

• The Loop Tourists understand, in principle, that the loop matters. They have read the essay. They talk about feedback and human-in-the-loop and continuous improvement. But they have not changed a single incentive, workflow, or metric to make it real. Their people are still rewarded for output speed, not for the judgement that improves the system. The loop is in the strategy deck and absent from the working day. This is, currently, the largest category — and the most dangerous, because the language of compounding provides cover for an organisation that is not actually doing it.

• The Loop Builders — the rarest — have rebuilt their operating model around the loop. They instrument the points where human judgement meets AI output and treat those points as the firm’s most valuable real estate. They reward the analyst who caught the subtle error, not just the one who shipped fast. They protect the cognitive work that feeds the system even when it looks, on a quarterly dashboard, like friction. They are often slower than the Token Buyers in any given quarter. They are building the slope while everyone else buys the level.

What this means in practice

For organisations, the strategic question has to change. “How much AI capability have we acquired?” is now close to meaningless — the honest answer for most firms is “roughly the same as our competitors, and increasingly so.” The question that actually predicts the future is: is our organisation getting measurably smarter every time it uses its AI, or just faster? If you cannot point to the specific mechanism by which judgement is captured, fed back, and accumulated, you do not have a learning loop. You have a faster way of doing what you already did, available on equal terms to everyone you compete with.

The practical moves follow from this. Instrument the human-AI interface and treat the corrections, overrides, and judgements that happen there as a first-class asset, not as exhaust. Reward the application of judgement, not just the velocity of output — because the moment your incentives favour speed over scrutiny, you are training your people to stop feeding the loop. Protect the cognitive work that builds expertise even when it reads as inefficiency, because expertise is the input the loop runs on, and an organisation that has automated away its own judgement has quietly switched off its compounding.

For individuals, the implication is sharper and, I think, clarifying. The era in which “I can use AI” was a differentiator is ending. Everyone can use AI. Fluency with the tools is becoming what literacy with a spreadsheet became — a baseline, not an edge.

The durable position is to be the human judgement that makes the loop compound: the person who notices what the model missed, who distrusts the plausible answer, who can correct the system rather than merely operate it, and who feeds that correction back so the next cycle starts higher. Being a fluent prompter makes you interchangeable with everyone else who prompts fluently. Being the judgement in the loop makes you load-bearing — the scarce complement to an abundant, commoditising input.

This is the actionable core of Nadella’s “only becomes more valuable.” Human capital becomes more valuable not automatically, and not for everyone, but specifically for the people and firms positioned at the point where judgement compounds AI capability. For everyone else, the abundance of cheap, capable AI is not an opportunity. It is a rising tide that lifts competitors who built loops and slowly submerges those who only bought tokens.

The closing uncomfortable truth

The deepest discomfort in the Compounding Divide is its timing. The divide becomes most dangerous at the exact moment AI feels most democratised. When everyone has access to frontier capability on similar terms, the natural conclusion is that the playing field has been levelled. The opposite is true. Equal access to a commoditising input is precisely the condition under which the only remaining differentiator — the compounding loop you cannot buy — matters most.

So the firms that feel safest are often the most exposed: they have the AI, the dashboards are green, the speed is real, and none of it is compounding because the human judgement that would make it compound has been optimised into silence. And the firms that feel behind — the ones whose people still argue with the model, still correct it, still insist on understanding before they ship — are quietly building the one asset that accumulates while everyone else’s depreciates.

Token capital is the part you can buy. That was always going to make it the part that converges. The advantage was never going to live in the asset everyone can acquire on equal terms. It lives in the loop — and the loop only turns if a human keeps feeding it judgement.

You can buy the tokens. You have to earn the loop. And the gap between the two is already widening faster than anyone buying their way in can see.

There is a management consulting firm in Amsterdam — call it a composite, because the pattern it represents is now widespread — that spent 2025 systematically removing friction from its knowledge work. Not the bureaucratic friction of approvals and sign-off chains, but the intellectual friction: the first-draft writing that analysts used to do themselves, the three-hour debates about strategic framing, the laborious manual synthesis of research before a client presentation.

The AI-assisted workflow they designed is, by any conventional productivity measure, excellent. Client-ready documents in a third of the time. Proposal turnaround in 48 hours instead of two weeks. Partners freed from review bottlenecks. Senior consultants finally able to focus on client relationships rather than document production. The metrics trend positive across the board.

Eighteen months in, a senior partner raised a problem that had no metric attached to it. The junior analysts — two cohorts’ worth, hired in 2024 and 2025 — could not explain why they believed the things they were recommending. Not in the sense of being unable to defend the AI’s logic. In the sense of having no formed view at all. The AI had produced a position. They had adopted it. The friction that would have produced the position in them — the struggle to structure an argument, the debate that forced articulation, the constraint that demanded creative problem-solving — had been optimised away along with everything else.

The firm had become very fast at producing work it no longer deeply understood.

This is the Friction Premium problem. It is not a technology problem. It is a structural one — and it is arriving across almost every organisation that has deployed AI substantively in knowledge work.

The science they left out of the AI strategy deck

The case against removing all friction from knowledge work has been building in the academic literature for thirty years. It simply hasn’t made it into most enterprise AI adoption frameworks.

In 1994, UCLA cognitive psychologist Robert Bjork coined the term “desirable difficulties” to describe a counterintuitive finding that had been accumulating across learning research for decades. Certain forms of resistance during learning — spaced retrieval rather than massed review, interleaved practice rather than block repetition, varied conditions rather than stable ones — consistently improve long-term retention and the transfer of skills to new contexts, even though they make learning feel harder in the moment and produce worse performance during the learning period itself.

The mechanism is not mysterious once you understand it.

Difficulty creates what psychologists call “generation effects” — when you have to retrieve, reconstruct, or produce something rather than simply recognise it, you engage different and deeper cognitive processes. You build connections between concepts. You construct understanding rather than merely encountering it. Those connections produce durable knowledge that transfers when the situation changes.

The smooth path — re-reading, summarising, reviewing already-formatted material — produces the sensation of understanding without the underlying encoding that makes knowledge usable in genuinely novel situations. You feel like you know. You may not know in the way that matters when the context shifts.

Bjork called the inverse of this the Fluency Trap: when processing feels easy, we interpret that ease as comprehension. The problem is that ease of processing and depth of understanding are not the same thing. Fluency is a feeling, not a fact. And it is the feeling that AI is extraordinarily good at producing — on behalf of the people who use it.

The Fluency Trap has an organisational-level equivalent. When an organisation runs its knowledge work through AI systems that produce fluent, plausible, well-structured outputs, the cognitive signal to evaluate critically is dampened across the board. The output looks like understanding. It carries the surface structure of careful thought. It may or may not have the underlying depth.

The empirical case at scale

The Microsoft and Carnegie Mellon University 2025 study of 319 knowledge workers is the sharpest quantitative picture of what this looks like in practice. Participants reported applying no critical thinking to 40% of their AI-assisted tasks. Not reduced critical thinking — none. More significantly, the study found that user confidence in GenAI inversely correlated with critical thinking application: the professionals who trusted AI most scrutinised its outputs least.

This is the Fluency Trap made visible at scale. When the output is fluent and confident in tone, the cognitive trigger for critical evaluation is not activated. When the human reviewer has high confidence in the AI system, their own evaluation is further reduced. Two compounding effects that together produce, across thousands of decisions per day, a systematic withdrawal of the cognitive work that was previously producing understanding.

The directional consistency across independent studies is striking. Gerlich’s 2025 study, published in MDPI’s AI Tools in Society, found a negative correlation between frequent AI usage and critical thinking ability, with the effect stronger in younger participants — exactly the cohort entering knowledge work organisations now and developing the professional judgement that will define their careers. A 2026 University of Technology Sydney analysis of cognitive offloading found that extensive reliance on AI tools can lead to a decline in cognitive engagement and skill development, though the effect is responsive to deliberate interaction design.

The mitigation finding is important. The problem is not the technology. It is the absence of intentional design. Most organisations have not designed their AI deployments to preserve cognitive friction. They have deployed AI to eliminate it, measured the resulting productivity gains, and called the job done.

What the friction was actually for

To understand what is being lost, it helps to decompose productive friction into its three distinct forms. They are different in mechanism, they produce different kinds of value, and they require different interventions to preserve.

Cognitive friction is the effort of articulation — the work of formulating your own argument before the AI does it for you. The labour of structuring a position, identifying its weaknesses, finding the right evidence, and presenting it in a form that holds together under scrutiny. This process does not merely produce a document. It produces understanding in the person doing it. The output is a side effect. The cognitive work is the point.

When AI writes the first draft, the human no longer does the work of articulation. They do the work of evaluation — which is genuinely valuable, but a different and often shallower cognitive task, particularly when the output is fluent and the evaluator has high confidence in the AI’s capability. The Fluency Trap reduces the quality of the evaluation step. The generation step — where real understanding is built — is skipped entirely.

Over time, the person who never articulates their position in writing has not built the same depth of understanding as the person who did. The outputs may be indistinguishable in the short term. The expertise is not. This difference only becomes visible when something genuinely novel arrives and the depth of understanding has to transfer to a situation the AI’s training data did not anticipate.

Social friction is the structured disagreement that forces better thinking. The meeting where two people with genuinely different frameworks argue until the logic becomes clearer for both. The pre-mortem that requires someone to articulate specifically and on paper why the plan will fail. The adversarial collaboration where a committed sceptic’s challenge produces a more robust position than any amount of internal consensus would have generated. These interactions are often uncomfortable. They are productive precisely because of the discomfort — the discomfort is the signal that genuine thinking is happening, that positions are being tested rather than ratified.

AI-assisted workflows tend to reduce social friction systematically. When the AI synthesises the debate rather than the debate happening in full, when the AI identifies risks rather than a sceptic articulating and defending them, when the AI generates the pre-mortem rather than a team constructing it through genuine anticipation — the social process that produced the improvement in thinking is compressed or eliminated. The conclusions may be similar. The understanding built in the room is not.

Structural friction is the constraint that forces creative solutions. The budget limit that requires rethinking the entire approach rather than adding resources. The technical restriction that pushes engineers toward solutions they would not have considered with unlimited flexibility. The deadline that eliminates the option of the perfect solution and requires the good-enough one that, in the search for it, reveals something the perfect solution would have permanently hidden.

“Training is frequently non-optimal because it fails to incorporate the variability, delays, uncertainties, and other challenges the learner can be expected to face in a real-world job setting.”

— Robert Bjork, UCLA Bjork Learning and Forgetting Lab

What gets lost: expertise versus the output of expertise

The most important distinction the Fluency Trap obscures is between expertise and the output of expertise.

AI can produce the output of expertise remarkably well. Given sufficient training data and a capable model, an AI system can produce a strategy document, a due diligence report, a legal brief, a financial model, or a market analysis that is indistinguishable from the work of a domain expert. This is real. It is valuable. It is one of the genuine and transformative capabilities AI brings to knowledge work.

What AI cannot produce is the expertise itself — the durable, transferable, judgement-forming understanding that exists in the person who would otherwise have done the work. In a world where AI routinely produces the output, it becomes far less obvious that this matters. The output is there. The client is satisfied. The process completes. Why should it matter whether any specific human deeply understood what was produced?

It matters for three reasons that only become visible at critical moments. First, evaluation: someone has to be able to tell whether the AI’s output is right, subtly wrong, or dangerously wrong in ways that fluent presentation conceals. The consultant who has never constructed their own analysis from first principles doesn’t know what a misleading analysis looks like from the inside. Second, adaptation: when conditions change in ways that weren’t in the training data, the organisation needs people who understand the domain, not people who can prompt an AI that understands it. The difference is invisible under normal conditions and critical under unusual ones. Third, succession: when the people who carry genuine expertise leave — which they always do — the organisation needs the next generation to have built expertise through their own cognitive work. If the junior generation never built expertise because AI produced the output throughout their development years, the knowledge exists nowhere a human can reach it.

Bain’s 2025 Innovation Report captures one dimension of this in findings about ideation quality. AI-assisted ideation produces a high volume of solutions with strong forecasted value. Human-generated ideas remain significantly stronger on novelty — particularly for breakthrough, category-defining innovation. The serendipitous interactions, the unexpected collaborations, the constraint-forced lateral thinking: these are precisely what AI cannot replicate, because they are produced by the conditions AI is optimised to eliminate.

Three archetypes: how organisations relate to productive friction

• The Friction Eliminators have defined their AI strategy primarily in terms of process efficiency. Friction is categorised as unnecessary by default — anything that slows output is a candidate for elimination. Their short-term productivity metrics are improving. Their talent development outcomes, their capacity to handle genuinely novel problems, and their organisational resilience when conditions shift are not being measured. The bill has not arrived yet.

• The Friction Preservers have maintained pre-AI working practices alongside AI deployment — requiring analysts to write first drafts before using AI assistance for refinement, maintaining structured debate in strategy sessions rather than letting AI-generated summaries replace the debate itself. They are being described internally as “not using AI to its full potential.” They are building organisations that can think independently of their AI tools.

• The Friction Designers — the rarest category — have distinguished systematically between process friction and productive friction and have used AI to eliminate the former while deliberately designing the latter into their workflows. They require AI-assisted drafts to be critically annotated before any recommendation is made. They use AI to generate the pre-mortem and then require the team to argue against it. They set explicit constraints on AI-assisted option generation to preserve the creative problem-solving that unconstrained generation would eliminate. They are building AI-augmented workflows that make the cognitive work harder, not easier, at the points where the cognitive work is most valuable.

The friction audit

For any organisation deploying AI in knowledge work, the practical starting point is what I call the friction audit. Not “what friction have we eliminated?” — the answer is measurable and probably positive by any conventional metric. But the harder question: what category was the friction we eliminated?

Process friction — time spent on handoffs, approvals, administration, formatting, and coordination that moves information between people and systems without transforming it — is generally waste. Eliminating it releases time for higher-value work. This is real and worth pursuing.

Cognitive, social, and structural friction are different in kind. Eliminating them does not release capacity for higher-value work. It eliminates the mechanism by which higher-value work was being produced. The time saved is real. The capability that was building in that time is quietly gone.

The audit question for each workflow where AI has removed friction is specific: what was the human cognitive, social, or structural work occurring inside that friction? If that work was producing understanding, judgement, or creative insight — even as a side effect of something that looked inefficient from the outside — its elimination has a hidden cost that the productivity metrics will not capture until the capability it was building is needed and is not there.

The closing uncomfortable truth

The Fluency Trap, at the organisational level, runs like this: the organisation that deploys AI most aggressively in knowledge work may become the most fluent at producing work it doesn’t deeply understand. The outputs will be excellent. Clients will be satisfied. The metrics will trend positive. The organisation will feel capable and efficient.

The exposure arrives when conditions change — when the novel problem appears that requires genuine understanding rather than fluent pattern-matching, when the expert who actually knew things has left, when the AI’s training data no longer maps to the situation at hand, when someone needs to explain not just what was concluded but why it was right.

The Acar, Tarakci, and van Knippenberg meta-review of 145 empirical studies found that a healthy dose of constraints benefits individuals, teams, and organisations alike. The constraints organisations are optimising away are not purely inefficiency. Some of them are the conditions under which people develop the judgement to do this work well.

The Friction Premium is not a case against AI deployment. It is a case for precision — for understanding the difference between friction that was genuinely waste and friction that was structural value disguised as inefficiency.

The organisations that understand this distinction will build AI deployments that eliminate process friction while preserving productive friction. They will be slower than the Friction Eliminators in the short run. They will be more capable when the conditions that reveal the difference between fluency and understanding finally arrive.

Speed is not intelligence. Smoothness is not competence.

The premium, in the long run, goes to organisations that know the difference — and had the discipline to preserve the friction that was worth keeping.

On August 2, 2026, the EU AI Act reaches full enforcement. For the first time in any major jurisdiction, organisations deploying high-risk AI systems are legally required to assign human oversight responsibilities to a named individual — someone with the specific competence to understand the system’s outputs and the actual authority to act when something goes wrong. Under corporate law, directors who consciously disregard significant AI governance obligations now face personal liability, not just organisational exposure.

The regulation is forcing an answer to the question most organisations have not asked. Not “do we have a responsible AI policy?” — 89% of organisations now do, according to Stanford’s 2026 AI Index.

But the harder question: who, specifically, is personally accountable when an AI-driven decision causes harm to a specific person?

These are not the same question. One is answered by a document. The other requires a name, a defined scope, and a consequence attached to that name. Organisations have produced documents at scale. They have not produced names.

The accountability void is the gap between those two things — and it is, right now, one of the most consequential and least examined structural problems in AI deployment.

How the void forms

The accountability void is not the product of deliberate evasion. It is the product of how AI deployment actually happened — through a series of individually reasonable decisions that collectively produced an unreasonable result.

When organisations began embedding AI into consequential workflows — screening job applications, scoring creditworthiness, triaging medical records, flagging compliance risk — they typically did so by treating AI output as an input to a human decision. The framing was: the AI recommends, the human decides. The human reviewing the recommendation was accountable for the decision. The AI was a tool, not a decision-maker. This framing made the accountability architecture feel settled.

But as AI capability expanded and the volume of AI-assisted decisions grew, the human review step became faster, more cursory, and less substantive.

What had been genuine evaluation became ratification. The AI recommendation increasingly was the decision. The human approval was a procedural step — performed, in many cases, without the time, the contextual information, or the technical understanding required to meaningfully evaluate what the AI had concluded.

The accountability architecture never moved. It still formally points at the human decision-maker — who is now on the legal hook for a decision they did not meaningfully make. The practical decision-making sits with a system that has no legal subjectivity and cannot be held accountable.

The void is the space between these two positions: formal accountability assigned to a human, practical authority exercised by an algorithm.

What makes the void structurally stable — and therefore dangerous — is that it is invisible as long as everything works. The AI recommendation is accepted. The human ratifies. The outcome is fine. No audit trail reveals that the human didn’t actually evaluate the recommendation; they simply processed it. The void is only exposed at the moment it matters most: when the AI is wrong, the harm is real, and someone needs to answer for it.

The evidence: incidents rising, response quality falling

Stanford’s 2026 AI Index provides the most comprehensive quantitative picture of how this void is developing. The AI Incident Database recorded 362 documented AI incidents in 2025 — up 55% from 233 in 2024. The rate of increase significantly outpaces the growth in AI deployment, suggesting that the tail-risk exposure is growing as a proportion of the deployed base, not just in absolute terms.

More telling than the incident count is what happened to organisational response capability in the same period. The share of organisations rating their own AI incident response as “excellent” dropped from 28% in 2024 to 18% in 2025. The organisations deploying more AI are, by their own assessment, becoming less capable of handling its failures.

Simultaneously, the Foundation Model Transparency Index dropped from an average score of 58 in 2024 to 40 in 2025. The systems that organisations are deploying are becoming less transparent about their own workings at exactly the moment when those organisations need to be more accountable for the outcomes those systems produce.

When the void becomes a lawsuit

The legal exposure of the accountability void moved from theoretical to concrete in 2025. In May, a US federal court certified a nationwide collective action in Mobley v. Workday. Workday’s AI-powered hiring platform was accused of systematically screening out applicants over 40, using a model that evaluated skills alignment and provided ranked recommendations to employers. Applicants received automated rejections without any disclosed AI involvement.

A parallel case targeting Eightfold AI raised the same structural question: applicants were scored and rejected without knowing AI was involved, without access to the criteria, and without a meaningful mechanism for human review. The legal question at the centre of both cases is the accountability void made concrete: if a candidate is scored and rejected by an AI system, and the human reviewer never substantively evaluated the underlying case, who is accountable for the outcome?

“The rapid deployment of AI systems at scale could create substantial responsibility gaps, increasing the likelihood of damages, losses, and wrong expectations.”

— London Business School, “The Governance Gap at the Heart of the AI Boom” (2025)

The policy illusion

89% of organisations now have a formal responsible AI policy. The documents are real. The governance intentions are, in many cases, genuine. The problem is structural: responsible AI policies are written at the organisational level. Accountability for a specific decision affecting a specific person must sit at the individual level.

Organisations have the former. They almost universally lack the latter.

Grant Thornton’s 2026 AI Impact Survey found that most organisational AI governance frameworks remain internal guidelines with no external verification, no shared standards, and no mechanism for assigning individual accountability to specific AI systems. They specify what the organisation believes about AI use. They do not specify who is personally responsible when a specific AI output causes a specific harm.

Three archetypes of nominal accountability

The Policy Owner has genuine accountability for the policy’s existence. They have no meaningful accountability for the specific outputs of specific AI systems. When a specific AI decision causes harm, they can point to the policy’s existence; they cannot answer for the decision.

The Approving Manager signed off on deployment. Their ongoing accountability for specific outputs is formal rather than substantive. Their accountability is institutional — they said yes to the system — but not individual: they did not make the decision that produced the harm.

The Human in the Loop is the formal oversight mechanism. In practice, they are processing volume that makes genuine evaluation impossible. A recruiter reviewing 400 AI-screened candidates in a day cannot meaningfully evaluate 400 AI recommendations. Their formal accountability is intact; their practical capacity to exercise it is not.

What genuine accountability looks like

Genuine accountability for an AI system requires four specific elements.

• First, a named individual — not a team, not a policy document — with defined scope.

• Second, the competence to understand those outputs: the named individual must have access to the system’s logic, training, known failure modes, and performance data.

• Third, the authority to act: the named individual must have the formal power to override, pause, or escalate without requiring multiple levels of approval.

• Fourth, a defined consequence: what happens, specifically, when the named individual fails in their oversight function?

Most organisations have zero of these four elements specified for most of their AI systems.

The EU AI Act is forcing all four for high-risk systems. The rest remains in the void.

The closing uncomfortable truth

The accountability void will not resolve itself through continued policy development. It will resolve — in most organisations — through a serious failure that forces the question of who is personally accountable into a context where the answer can no longer be deferred.

The organisations best positioned when that moment arrives are those doing the unglamorous work now: naming the humans accountable for specific AI systems, equipping them with the competence and authority their role requires, and attaching a consequence to that accountability that is real rather than nominal.

Most organisations have a responsible AI policy. Most organisations cannot name the specific human accountable for a specific AI output that harmed a specific person.

One is a document. The other is governance. The document exists. The governance hasn’t arrived yet.

The corporate hierarchy was not designed to survive this

It was designed to manage information in a world where information was expensive to move. The seven-to-nine-layer enterprise org chart — the one that defines almost every company of consequence built between 1920 and 2015 — is, at its structural core, a solution to a 1920 problem: how do you get information from the front line to the person who needs to act on it, and instructions back to the front line, without it dissolving into noise along the way?

The answer was layers. Each layer served as a compression function. Front-line workers generated data. First-line supervisors synthesised it into summaries. Middle managers translated those summaries into cross-functional coordination. Senior managers turned coordination into strategy inputs. The C-suite made decisions and sent instructions back down the same chain. The org chart was, in every operational sense, an information routing architecture built from people.

McKinsey’s 2026 “State of Organizations” report found something that might be the most understated finding of the year: two-thirds of leaders say their organisations are overly complex and inefficient, and nearly 40% say that redefining process flows — not restructuring, not headcount reduction — is the single biggest productivity unlock available to them right now. The complexity they are describing is not accidental. It is the residue of an information routing architecture that is no longer needed for the purpose it was designed for.

The AI Stack changes the economics of information relay at every layer. A generative engine synthesises. A predictive engine anticipates. A perceptive engine monitors in real time. An agentic engine routes, escalates, and executes. An optimisation engine continuously improves. The five-layer AI Stack — the subject of the first piece in this series — performs, automatically and at scale, the compression and relay function that occupied most of the coordination middle of the traditional org chart.

The question is not whether the hierarchy will change. Gartner estimates 20% of organisations will eliminate more than half their middle management positions through AI-driven restructuring by 2026. Korn Ferry found that 41% of employees already report their company has trimmed management layers in 2024–25. The question is whether the change is intentional — or just a smaller, more confused version of the same broken architecture.

This is the fifth and final piece in my series on the Organisation of the Future. The first four named the building blocks: the AI Stack (the engines), the Operator (the humans who run them), the Verifier (the humans who ensure the outputs earn trust), the Workflow Designer (the humans who architect the road). This piece draws the org chart they all live on.

How the hierarchy got so tall

To understand why the org chart is changing, it helps to understand why it became what it is.

Frederick Taylor formalised the division of labour in the 1910s. Alfred Sloan built the modern divisional corporation at General Motors in the 1920s — the model that the Fortune 500 spent the next fifty years replicating. The divisional structure needed layers because it was managing scale without computers. A VP of Operations at GM in 1950 had no automated monitoring of production lines across twelve plants. They had reports from plant managers, who had reports from shift supervisors, who had reports from team leads. Each layer existed because someone needed to synthesise, check, and relay what was happening in the layer below.

The corporate computer age, beginning in the 1970s and accelerating through the 1990s, compressed some of this — enterprise resource planning systems made data more visible across layers, and the internet enabled real-time communication that shortened some relay chains. But the fundamental management architecture of the industrial corporation survived into the 2020s largely intact. Most Fortune 500 companies still had seven to nine reporting layers between a frontline worker and the CEO in 2024.

The reason those layers persisted, even as technology improved, is that information synthesis — the judgment call about what matters and what doesn’t, what to escalate and what to absorb — was still fundamentally a human task. An ERP system could show you that SKU 7843 was running 14% under plan. It could not tell you whether that mattered, why it was happening, who needed to know, or what to do about it. That judgment lived in the middle layer. The middle layer persisted because judgment was scarce.

The AI Stack changes the scarcity calculus on judgment at three levels simultaneously. It automates the monitoring (perceptive engines). It synthesises the signal from the noise (predictive engines). It routes the escalation to the right human seat, with the right context packet, at the right time (agentic engines). The judgment that middle management spent 70% of its time preparing for — the ten minutes of actual decision-making that justified the hour of synthesis and the two hours of reporting — can now be delivered pre-synthesised to the person who needs to make it.

What gets lost

Before drawing the new org chart, there is something important to acknowledge about what is structurally lost in the transition.

Middle management is not simply a coordination mechanism. It is also an institutional memory system. The manager who has been in post for six years and knows that the finance-sales handoff has a seventeen-step workaround because of a legacy system migration in 2019 is carrying institutional knowledge that does not live in any process document, any ERP system, or any AI training dataset. That knowledge exists as context in a person. When that person leaves — whether voluntarily as part of a restructuring or involuntarily as part of a headcount reduction — the knowledge leaves with them.

This is not a small problem. Gartner’s May 2026 analysis found that AI-driven restructuring is creating budget room but not delivering the productivity returns companies expected. The mechanism is this: companies cut the coordination layer, expecting the AI Stack to absorb the coordination function, but find that the AI Stack was not trained on the institutional context those people carried. The workflow breaks in subtler ways. The escalation goes to the manager who no longer exists. The synthesis misses the context the predictive engine was never given. The agentic engine routes correctly according to the workflow specification but incorrectly according to the institutional reality.

The Workflow Designer’s primary job — mapping the actual workflow, not the documented workflow — is partly a knowledge transfer mechanism. The Mapper archetype of the Workflow Designer role exists precisely because the gap between the process document and the real process is where institutional knowledge lives. Before you can eliminate the coordination middle, you have to extract what it knows. Most companies doing AI-driven restructuring in 2025–26 are not doing this extraction. They are eliminating the relay runners and hoping the documentation is good enough. It rarely is.

The five-layer architecture

The new org chart has five genuine layers. They are not five reporting levels — they are five functional layers with different relationships to the work, to the AI engines, and to each other. A Verifier may formally report to a Director of Operations, but their functional role is in the Trust Layer regardless of where they sit on a reporting tree. The org chart of the future is a functional map, not just a hierarchy chart.

• Layer 1: The Engine Room. The AI Stack runs in the background of every workflow the organisation executes. Generative, predictive, perceptive, agentic, and optimisation engines — the five types named in the first piece in this series — are infrastructure, not personnel. They are the most important layer in the new org because they are the one that changes the economics of everything above them. And they are the one most organisations are building without the other four layers in place. The Engine Room is not empty of humans — it requires engineers, architects, and calibration specialists — but those humans are infrastructure roles, not coordination roles.

• Layer 2: The Execution Layer. The Operators — Conductors, Translators, Mechanics, and Surgeons, as named in the second piece — are the humans who run the AI-augmented workflows. They are not managing people. They are managing processes that combine human judgment with AI output at every consequential step. The Execution Layer is also the largest human layer in the new org by headcount. Most of the people who survived the reduction of the coordination middle are now Operators, whether or not their job title says so.

• Layer 3: The Trust Layer. The Verifiers — Domain Expert, Critic, Auditor, Red Team, as named in the third piece — are the humans who ensure that outputs from the AI Stack earn organisational trust before they are acted on, communicated, or escalated. The Trust Layer is functionally new. In the old org, trust was built through hierarchy: something was trustworthy because a director had signed off on it. In the new org, the director-as-trust-mechanism is being replaced by a systematic verification function. In regulated domains — finance, healthcare, hiring, content moderation — this layer is now legally load-bearing under the EU AI Act, which reached full enforcement in August 2026.

• Layer 4: The Architecture Layer. The Workflow Designers — Mapper, Boundary Setter, Recovery Designer, Composer — are the structural engineers of the new org. They sit between the Direction Layer and the Execution Layer, translating strategic goals into designed workflows. They are the new coordination mechanism: instead of a human relay chain synthesising information in real time, a designed workflow specifies in advance where information flows, where it gets synthesised, where humans make decisions, and what happens when the workflow breaks. The Architecture Layer is the smallest human layer in the new org by headcount but the highest-leverage: a single Composer designing a well-specified workflow can enable fifty Operators to execute reliably at scale.

• Layer 5: The Direction Layer. The senior leaders who set goals, make strategic bets, and own accountability. What changes at this layer is not title or authority but operating reality. The Direction Layer in the old org received synthesised information through a relay chain and made decisions based on what the relay produced. In the new org, the relay chain has been replaced by designed workflows and AI synthesis — which means the Direction Layer has direct sight-lines to the actual work, without the compression and distortion that six layers of relay introduced. This is both more powerful and more demanding. A senior leader with direct sight-lines to AI-synthesised workflow data is better informed, faster. They are also being asked to make more consequential decisions with less buffer — because the buffer was the middle.

The deleted layer and the coordination question

The layer that the five-band model has no place for is middle management as a coordination mechanism. Not as a person type — but as a functional role: the human whose primary job is to synthesise information from below, coordinate across functions, and relay upward. That coordination function has been absorbed by three parts of the new architecture: the Workflow Designer’s Architecture Layer (who designs the coordination logic in advance), the AI Stack’s agentic and predictive engines (who execute the coordination logic in real time), and the Operator’s Execution Layer (who own the workflow outcomes and make the in-context judgment calls that the design cannot fully specify in advance).

“Shifting attention from structure to flow — with the biggest productivity payoff lying not in reorganising reporting lines but in radically simplifying and unifying processes across the enterprise.”

— McKinsey, The State of Organizations 2026

What this means for people currently in coordination-heavy middle management roles: the function is not disappearing, but its location in the org is changing. The judgment, escalation, and domain expertise components of middle management roles are being elevated into the Operator and Verifier tiers. The synthesis and relay components are being absorbed by the AI Stack and the Workflow Designer. The transition is not “middle managers are being replaced by AI.” It is “the coordination-heavy version of middle management is being replaced by a combination of designed workflows and AI engines, while the judgment-heavy version is being elevated into named functional roles with higher accountability.” That is a harder story to tell in a restructuring announcement, which is why most companies are not telling it.

Building it deliberately

The difference between a company that captures the productivity gain of the new org and one that is still running pilots in 2028 comes down to one question: did they build all five layers, or just the Engine Room?

McKinsey’s finding that only 21% of companies have redesigned their workflows end-to-end is, read through the lens of this series, a finding about which companies have invested in all five layers. The 21% have Operators, Verifiers, and Workflow Designers in place around their AI Stack. The 79% have engines and hope.

Building deliberately means four specific decisions. First: name the role families explicitly. Operator, Verifier, and Workflow Designer need to exist as intentional job descriptions, not as retrofits onto existing titles. Second: do the knowledge extraction before the restructuring. Map the real workflows — with Mapper-archetype thinking — before you eliminate the people who carry the institutional context those workflows depend on. The order matters enormously. Most companies are getting it backwards. Third: design the Trust Layer for your regulatory context. In high-risk domains, the Trust Layer is now legally required, not just good practice. The companies that treat the Verifier as optional are building legal liability into the org chart. Fourth: widen spans of control intentionally, not accidentally. Span-widening only works if the Workflow Designers have done their job. If the Architecture Layer is under-resourced, wider spans of control at the top become a recipe for chaos.

The closing uncomfortable truth

Gartner’s most quietly important finding of 2026 is this: AI will create more jobs than it eliminates — but not until 2028. Through 2026 and 2027, the net effect on global employment is roughly neutral. We are in the transition window — the period between old org and new org in which both are partially visible.

The companies that build the new shape deliberately during this window will be the ones with the Architecture Layer, the Trust Layer, and the Execution Layer staffed and operational by the time the transition completes. The ones that cut the relay middle without installing the new structure will find themselves in 2028 with AI engines, no coordination architecture, and a growing list of workflows that break in ways nobody can diagnose — because the institutional knowledge left with the people who were made redundant in 2026.

The org chart of 2030 has fewer layers than today’s.

The people who remain carry more judgment, more accountability, and more leverage than anyone in the 2020 version of the same company. The flat org is not a smaller version of the tall org. It is a different shape entirely.

Drawing that shape on purpose is the only version of this that works.

This is the fifth and final piece in my series on the Organisation of the Future. The full series: The AI Stack (May 27) → The Operator (June 3) → The Verifier (June 10) → The Workflow Designer (June 17) → The New Org Chart (June 24). Thanks for reading!

In early 2026, MIT’s NANDA Initiative published the most-quoted line of the year in enterprise AI. Across more than 1,800 generative AI deployments they had benchmarked in 2024 and 2025, 95% had delivered zero measurable P&L impact. Not “small but real”. Zero. The number ran through every CFO Slack channel by the end of the week. S&P Global ran its own audit shortly after and found that 42% of companies had abandoned most of their AI initiatives in 2025, up from 17% the year before. IBM’s Institute for Business Value put the share of enterprise AI initiatives delivering expected ROI at 25%. Gartner — based on a poll of 3,400+ organisations actively investing in agentic AI — now forecasts that 40%+ of agentic AI projects will be cancelled or fail to reach production by the end of 2027.

These are not the numbers of a market failing to spend. The same year produced a record $675 billion in hyperscaler AI infrastructure spend, with cumulative investment headed toward $3–4 trillion by the end of the decade. The capex is overwhelmingly there. The pilots are overwhelmingly there. The P&L impact is overwhelmingly not.

McKinsey, looking at the same gap, found something cleaner. The roughly 6% of organisations they call “AI high performers” — those attributing more than 5% of EBIT to AI — capture about three times the value of everyone else. They are not better at picking models. They are not better at writing prompts. They are not better at vendor selection. The single biggest factor separating them from the rest is that they redesigned their workflows end-to-end. Only 21% of all companies have. The other 79% are running new tools through old plumbing — and producing the headline numbers above.

Three weeks ago, in the first piece of this series, I introduced the five AI engines — generative, predictive, perceptive, agentic, optimisation — that will run the 2030 enterprise. Two weeks ago, the AI Operator: the new orchestration role that supervises the stack and splits into four archetypes (Conductor, Translator, Mechanic, Surgeon). Last week, the AI Verifier: the role that checks the work, and splits into four more archetypes (Domain Expert, Critic, Auditor, Red Team).

Each of those pieces answered the question “who does the work in the new stack?” This week’s question comes before all of them, and is the question that the McKinsey 21% number quietly puts on the table: who designs the work itself?

The answer is a structurally new role I’ll call the AI Workflow Designer. Like the Operator and the Verifier, it does not arrive as a single job. It splits into four named archetypes: Mapper, Boundary Setter, Recovery Designer, Composer. Each one handles a different part of the architecture. A serious 2027 workflow function has at least two of these archetypes. The high performers — the McKinsey 6% — have all four. Most organisations today have none of them, by name, on the org chart.

“Most companies are applying AI to individual tasks rather than redesigning entire workflows, but the real productivity unlock comes from reimagining workflows so people, agents, and robots each do what they do best.” — McKinsey Global Institute, 2026

Why the workflow became the bottleneck

Most of the AI failures of 2024 and 2025 were not model failures. The model was capable. The vendor demo worked. The proof of concept was credible. The board deck looked respectable. The deployment then collapsed somewhere between the pilot and the production load — and the post-mortem, almost without exception, did not say “the model was wrong”. It said some combination of “the data wasn’t where we thought it was”, “the escalation path didn’t exist”, “the handoff broke”, “the recovery wasn’t designed”, “the boundary between agent and human action wasn’t drawn”, “we hadn’t mapped the actual workflow before we automated it”. Architecture failures. Not model failures.

Microsoft’s AI Red Team published a taxonomy of agent failure modes in 2025 and updated it in early 2026. Almost every production agent failure they have catalogued, across their internal estate and the customers they advise, traces back to one of five repeating patterns. Four of the five are architecture issues — bad handoff design, missing fallback, unclear authority boundary, brittle composition between engines. Only one is squarely a model defect. The model is not the bottleneck. The workflow is.

Look at customer support, the largest single AI-touched workflow in the enterprise. 2026 CX research consistently finds that only 15% of consumers experience a seamless AI-to-human handoff. The other 85% report disjointed transitions where they have to repeat their issue, where the human agent lacks context, where the transfer takes minutes, where the chatbot has just tried to argue with them after they explicitly asked for a human. One in three human agents reports lacking the customer context they needed to resolve the issue after the AI handoff. None of these is a model accuracy failure. All of them are workflow design failures.

Look at software engineering. Recent industry analyses of agentic coding workflows in 2026 found that teams without structured delegation primitives — defined boundaries between what the agent decides and what the human decides — saw a 23% increase in bug density and a 12% increase in time spent on code review. The agent was capable. The workflow around it was not.

Look at credit decisions, claims triage, hiring, content moderation. The pattern repeats. The model performs at or above human baseline on the discrete task. The system around it produces outcomes that range from mildly disappointing to publicly catastrophic. The story is almost never “the model was wrong” alone. It is “the workflow let the wrong output ship”.

This is why the Workflow Designer is the highest-leverage IC role of the next five years. The model is the engine. The Operator drives it. The Verifier checks it. The Workflow Designer is the person who decides where the engine goes, where the driver sits, where the brake is, where the recovery lane is, and where the boundary between machine and human is drawn. Without that role, the rest of the stack runs on hope.

What the AI Workflow Designer actually does

Three things, none of which fit cleanly into the existing org chart.

• Architectural mapping. The AI Workflow Designer maps the real workflow before any model is added to it — the documented steps, the undocumented steps, the data that quietly passes between people on Slack, the implicit escalations, the unstated authority boundaries, the failure modes the team already knows about but has never written down. The first deliverable is never a tool decision. It is a picture of the workflow as it actually runs.

• Authority specification. The AI Workflow Designer draws, for each step in that workflow, the line between what the agent is permitted to decide, what the human seat is required to decide, what must be escalated, what must be flagged for the Verifier, what must be logged for the auditor, and what must never be done at all. This is the part the EU AI Act, which becomes fully enforceable on 2 August 2026, made legally load-bearing. A workflow without explicit authority boundaries is now a workflow with explicit legal exposure.

• Resilience design. The AI Workflow Designer specifies what happens when the workflow breaks. Not whether it breaks — when. The timeout. The fallback. The rollback. The escalation packet. The customer-facing communication. The internal incident path. The audit-trail capture. The thresholds that automatically trigger human review even when nothing has visibly failed. Most production AI failures of 2024–2025 had no resilience design at all. The first time the workflow broke, it broke loudly, publicly, expensively, and unrecoverably.

The four Workflow Designer archetypes

The role splits cleanly into four. None of them is a hierarchy. They are flavours, each indispensable to a different stage of the design.

• The Mapper. The systems thinker who sits with a domain expert for two hours and walks out with the actual workflow on paper — including the parts that aren’t in any process document. Their value is realism. They notice that the documented loan-approval flow has eleven steps and that the real one has nineteen, and that step fourteen is “Marie phones Pierre on Tuesday to clarify the income field”. They notice that the marketing approval workflow lists three reviewers and that two of them actually rubber-stamp without reading, and that the third is the one whose judgement everyone implicitly trusts. They notice that the order-fulfilment workflow officially has no manual exceptions, and that in practice three percent of orders are handled out of band on email because the system can’t represent them.

Mappers come from business analysis, operations, service design, lean manufacturing, internal consulting. Their skill is observation rather than imagination — they draw what is, not what should be. Best fit: any workflow about to receive AI for the first time, where the gap between the documented process and the real one is the precise gap in which the model will silently break. A Mapper who fails to surface the undocumented steps is the reason a pilot looks fine in demo and shatters in production. The good Mapper produces a workflow map that the domain experts read and quietly nod at: “yes, that is actually how it works”.

The Mapper’s risk is producing a map of what is, and stopping there. The mature Mapper is paired with at least one of the other three archetypes — usually the Boundary Setter — to turn the map into a design.

• The Boundary Setter. The decision architect who specifies, for each step in the mapped workflow, where AI is permitted to act and where the human seat is required. Their value is rigour. They write the policy that says: the agent may approve loans up to €25,000 with predicted-default-rate under X; between €25,000 and €100,000 the agent may recommend but the human must sign; above €100,000 the workflow exits the agentic system entirely. They write the policy that says: the content-moderation agent may delete spam and remove obvious hate speech; borderline political content escalates to a human reviewer within fifteen minutes; content involving named public figures is not actioned by the agent at all.

Boundary Setters come from product management, policy, risk, ethics, regulated-industry compliance, and senior platform engineering. Their habit is to think in terms of permissions, thresholds, and exceptions rather than features. Best fit: any workflow with consequence — financial decisions, hiring, healthcare, content moderation, customer-facing communication, any workflow inside the EU AI Act Annex III categories. The Boundary Setter’s output is now legally load-bearing under the EU AI Act, the UK AI policy framework, the emerging US state-level AI rules, and the major insurers’ policy renewals.

The Boundary Setter’s risk is over-specification — a policy so dense and conservative that the workflow falls back to humans for everything and the AI investment never lands. The mature Boundary Setter ships a policy that is permissive enough to capture the value and conservative enough to survive the worst week of the year.

• The Recovery Designer. The failure-mode specialist who designs what happens when the workflow breaks. Their value is graceful degradation. They specify the timeout — the agent must reach a verdict in fewer than seven seconds; otherwise the workflow falls back to a defined human queue. They specify the rollback — if any step in the chain produces an error, the system reverses the last three steps and notifies the operator. They specify the human-handover packet — what the human receives when an escalation arrives, in what format, with what context, what evidence, what suggested action. They specify the apology — what gets said to the customer, by whom, with what authority. They specify the audit trail — what is logged, where it is stored, who can access it, how long it is retained.

Recovery Designers come from site reliability engineering, incident response, customer experience leadership, safety engineering in regulated industries, and military operations planning. Their habit is to assume the workflow will fail and to design for that failure to be small, recoverable, and well-communicated. Best fit: every production agentic workflow — because by 2027, the question is not whether your workflow will fail in production but how visibly, how recoverably, and how cheaply it will fail.

The Recovery Designer’s risk is paranoia — a design so defensive that the agent cannot act without three layers of fallback, latency rises, and the workflow stops feeling like AI at all. The mature Recovery Designer designs for the failure that actually happens, not every failure that could be imagined.

• The Composer. The architect who takes the Mapper’s workflow, the Boundary Setter’s authority policy, the Recovery Designer’s resilience plan, the available AI engines, the available Operator and Verifier archetypes, and assembles them into a coherent end-to-end workflow that actually ships value. Their value is integration. They decide where the generative engine ends and the predictive engine begins. They decide which Operator archetype owns which segment of the workflow. They decide which Verifier gate sits at which threshold. They decide which engines never touch each other.

Composers come from senior product leadership, distinguished engineering, chief-of-staff backgrounds, technical-strategy consulting, and increasingly from a new wave of explicitly AI-architecture programmes. Their habit is to hold the whole flow in their head at once. Best fit: every workflow that touches more than one AI engine — which by 2027 will be the majority of production AI workflows. The Composer is the role most likely to grow into the Chief AI Architect title that does not yet exist in stable form on most org charts.

The Composer’s risk is elegance over operability — a beautifully integrated architecture that the Operators cannot actually run, the Verifiers cannot actually verify, and the team cannot actually maintain. The mature Composer designs for the team that exists, not the team they wish they had.

None of these four is a hierarchy.

The high-performer pattern is to have all four, with the Mapper and Boundary Setter working in tight pair on the front end of every new workflow, the Recovery Designer engaged from day one rather than after the first incident, and the Composer holding the integrated picture and signing off the architecture before the engines arrive.

The mentoring problem this surfaces

Here is the second-order failure mode that ties the Workflow Designer back to The Apprenticeship Implosion, The Originality Tax, and The AI Verifier: we are not training AI Workflow Designers either, and the existing seam between product, operations and ethics — where this role lives — is not somewhere any single university programme, bootcamp, MBA, or corporate L&D track currently delivers people from.

Product schools train feature design. Operations training trains process improvement. Ethics training, where it exists, trains review. Software engineering programmes train shipping. None of them trains the integrated muscle the Workflow Designer needs: the ability to sit with a domain expert and reverse-engineer their real workflow, then to set decision boundaries that survive the worst Tuesday of the year, then to design the recovery the system needs when (not if) it breaks, then to compose multiple AI engines into one shipping flow. That is product + ops + ethics + integration architecture in one head — and the job description does not yet exist on the major boards in stable form.

What this means in practice is that for the next two to three years, the Workflow Designer is overwhelmingly a promotion candidate, not a hire. The strongest candidates are senior product leads who have already shipped complex multi-team flows; the senior operations managers who have already mapped end-to-end processes for transformation programmes; the chief-of-staff types who have already composed across silos; the SRE leads who already think about failure modes professionally; and the regulated-industry compliance leads who already think about authority boundaries with legal precision. The market signal of the next twelve months will be the salary band these promotions land at, not the title.

What this means

• If you are early in your career: stop chasing the “AI engineer” title that increasingly means “good with prompts”. Build Workflow Designer evidence. Pick a workflow inside your organisation — a small one is fine — and map it end-to-end yourself, including the unwritten steps. Write the authority boundaries you would propose, with thresholds, escalation paths, and worst-case constraints. Design the recovery: what happens when the workflow breaks, who is told what, how the customer learns. Compose two AI engines into a single shipping flow, even if it is small. Publish what you find. In eighteen months, that portfolio will be worth more than any frontier-model fluency on its own. The market signal in 2027 will not be “I can ship with AI”. It will be “I designed the workflow that captured the value”.

• If you are hiring: add at least one AI Workflow Designer seat to every team running multiple AI engines. The McKinsey 21% data is unambiguous — the companies that capture the AI productivity gain are the ones that have done end-to-end workflow redesign, and they have done it because someone, by name, owns that work. If you cannot find the candidate on the market — and you mostly cannot — promote from inside. Your best senior product leads, principal operators, chiefs of staff, SREs, and compliance leads are your strongest candidates, and they already know your domain. Hire for the archetype, not the title. The market for the title will catch up in eighteen months.

• If you are leading: the AI Operator and the AI Verifier without the AI Workflow Designer are tactics without an architecture. The McKinsey 21% number is the number that will define who captures the AI productivity gain by 2028 and who is still running pilots. Three things to do this quarter. First, name the Workflow Designer seat explicitly on every team running more than one AI engine — not “the product manager handles it” but “Sofia is the Composer on the underwriting workflow; Idris is the Boundary Setter”. Second, fund the role at parity with senior product and principal engineering. The Workflow Designer is a senior IC role, not a junior coordinator. Third, mandate the architecture deliverable: no AI workflow ships to production without a workflow map, an authority policy, a recovery plan, and a composition diagram signed by the Workflow Designer of record. No exceptions. That signature is the audit trail when something goes wrong, and it is the asset that compounds into capability over time.

The uncomfortable truth

Most organisations are buying engines, hiring drivers, installing gates — and skipping the road.

The AI Workflow Designer is the role that builds the road. It sits in the seam between product, operations and ethics, three functions that historically have not talked to each other in any sustained way. It does not look like a growth story. It does not have a clean parent function. It is not what venture markets fund and it is not what bootcamps ship. It is the role that the McKinsey 21% have, by name, on their org chart, and that the other 79% have not yet realised they need.

The next eighteen months will rebalance this. A few visible enterprise AI failures will be reframed by their post-mortems as “we never designed the workflow”; a wave of EU AI Act enforcement actions will turn the Boundary Setter output from a nice-to-have into a regulatory line item; a handful of high-performer case studies — McKinsey’s preferred genre — will explicitly name the AI Workflow Designer function in the org chart that captured the EBIT. The title will then stabilise. The salary band will then climb. The market will then catch up.

The companies that hire ahead of that adjustment will quietly accumulate a two- to three-year advantage that, when the market catches up, will look like luck and will in fact be preparation.

Next week, the closing piece of this series: the New Org Chart. What the company that has staffed the Operators, the Verifiers, the Workflow Designers, and integrated the five engines actually looks like on a single sheet of paper. The shape of the 2030 enterprise.

In late 2025, Forrester ran the numbers on a question most enterprise CFOs had quietly been afraid to ask out loud. How much, on a per-employee basis, was the company actually spending on the labour of checking AI output? Not the licence fees. Not the implementation costs. The hours: people reading what the model produced, cross-referencing it, correcting it, second-guessing it, sometimes throwing it out and starting again. The answer came back at roughly $14,200 per employee per year.

In a 10,000-person company, that is $142 million a year on a line item the budget does not have.

It is not a one-off. Industry estimates put LLM hallucination rates on certain factual and citation tasks as high as 82% — and even the most heavily benchmarked frontier models still fail open-ended frontier reasoning tasks at non-trivial rates. The market for hallucination-detection tools alone grew 318% between 2023 and 2025. Seventy-six percent of enterprises now run a formal human-in-the-loop process to catch errors before output ships. The World Economic Forum’s 2026 talent report shows that only 14% of organisations believe they have the AI-security talent they need to keep pace, and ManpowerGroup ranks AI skills as the single hardest-to-fill capability worldwide.

Two weeks ago, in the first piece of this series, I introduced the five AI engines — generative, predictive, perceptive, agentic, optimisation — that will run the 2030 enterprise. Last week, the AI Operator: the new orchestration role that supervises the stack and splits into four archetypes (Conductor, Translator, Mechanic, Surgeon). This week, the question every strategy deck quietly defers and the EU AI Act has just made non-deferrable: who, by name, checks the work?

The answer is a structurally new role I’ll call the AI Verifier. Like the Operator, it does not arrive as a single job. It splits into four named archetypes: the Domain Expert, the Critic, the Auditor, and the Red Team. Each catches a different kind of error. Each is recruited from a different pre-AI profile. A serious 2027 verification function staffs at least two of the four, and increasingly three.

“We are training a generation to produce with AI, and almost no one to check it. The asymmetry will define the next decade of corporate risk.” — Adaptation of the WEF Future of Jobs framing, 2026

Why verification became the bottleneck

For most of the post-war knowledge economy, production and verification ran at roughly the same speed. The contract took a senior associate three hours to draft; it took a partner forty minutes to review. The financial model took an analyst two days to build; it took a director ninety minutes to challenge. The article took a journalist three hours to write; it took an editor one hour to edit. Verification was cheaper than production — perhaps a third of the cost — but it lived in the same order of magnitude. The roles were balanced.

Generative AI has broken that balance. A frontier model now drafts a contract clause in eight seconds, builds a credible-looking financial model in two minutes, and writes a publishable article in twenty. The drafting cost has collapsed by two orders of magnitude. The verification cost has not. Verifying the contract clause still takes the same forty minutes. Verifying the model still takes the same ninety. Verifying the article still takes the same hour. And — in many domains — verification has actually become harder, because the output is now plausible at the surface in a way that human-drafted output rarely was. Errors hide better when they are wearing the syntax of competence.

Step-level verification benchmarks released in 2025 (Hard2Verify, evaluating 29 generative critics and process reward models on frontier maths reasoning) confirmed the gap empirically. Most open-source verifier models still lag closed-source counterparts on identifying the first error in a chain of reasoning. The harder the domain, the worse the gap. In other words, the machines we built to check the machines are themselves not as good as the machines doing the work. Even Anthropic, OpenAI, and DeepMind — when they need a ground-truth verifier — still default to expensive human raters.

Add the legal pressure. On 2 August 2026, the EU AI Act becomes fully enforceable. Penalties for non-compliance run up to €35 million or 7% of global annual turnover — whichever is higher. Technical documentation must be drawn up before market placement, kept up to date throughout the system’s lifetime, and retained for ten years. Every high-risk system listed in Annex III — employment, credit, education, law enforcement, critical infrastructure — needs a defensible audit trail. The EU AI Office and member-state authorities have the power to demand documentation, conduct evaluations, and order corrective measures. A vague “we have a human in the loop” sentence will not survive an inspection.

Verification is no longer a discretionary post-process. It is the structural bottleneck of the agentic enterprise, and one of the most regulated parts of the stack.

What the AI Verifier actually does

Three things, none of which a generalist reviewer does well.

• Calibrated disagreement. The AI Verifier reads an AI output and decides — fast, with consequence attached — whether to trust it, edit it, escalate it, or reject it. The skill is not “find the error”. The skill is “set the right confidence interval on this output, against the cost of being wrong”. An AI Verifier who flags everything is a bottleneck. An AI Verifier who flags nothing is a rubber stamp. The work is in the middle, and the middle requires judgement of a specific kind.

• Failure-mode anticipation. The AI Verifier knows, before any output is read, the failure modes the engine in question is prone to. The generative engine fabricates citations. The predictive engine over-extrapolates from the training distribution. The perceptive engine fails silently in low-data subgroups. The agentic engine drifts when the environment changes. The optimisation engine maximises the wrong proxy. An AI Verifier walks into a review with a hypothesis about where this output is most likely to be wrong, not a blank mind.

• Defensible documentation. The AI Verifier produces a record. Not a slack message, not a vibe — a structured record that says what was checked, what was found, what the threshold was, what was approved and on what basis, who signed. This is the part the EU AI Act, the UK AI policy framework, the SEC’s emerging guidance, and the major insurers are all converging on. It is also the part that turns AI Verifier work from cost centre to asset: the documentation is what gets you through the audit, the lawsuit, the regulator visit, and the difficult board meeting.

  ---

The four AI Verifier archetypes

The AI Verifier role splits cleanly into four. Each catches a different kind of error. None of them is a hierarchy — they are flavours.

• The Domain Expert. The senior practitioner with twenty years of pattern recognition in their field, reading the AI output and immediately seeing what is wrong about it. Their value is depth. The Domain Expert is the surgeon who reads a model-generated differential diagnosis and notices the missing rare condition; the senior tax partner who reads a model-generated structuring memo and notices the obsolete treatment of carried interest; the chief structural engineer who reads a model-generated load calculation and notices the soil assumption that does not match the site. They do not need a checklist. They have one in their head, refined over decades.

The Domain Expert comes from senior practitioner roles — medicine, law, engineering, finance, scientific research, regulated trades. Best fit: any high-stakes domain where surface plausibility and substantive correctness routinely diverge, and where the cost of being wrong is paid in lives, balance sheets, or regulatory consequence. Domain Experts are the most expensive AI Verifier archetype to recruit because they are also the most expensive non-AI Verifier practitioners. The structural shift in the role is that — increasingly — their value lives in the verification work, not the production work the agent now handles.

The Domain Expert’s risk is over-reliance on tacit pattern matching. The instinct that catches the rare missing diagnosis is also the instinct that mistakes “this looks like what I have seen before” for “this is correct”. The mature Domain Expert pairs their tacit judgement with one of the other archetypes’ tools.

• The Critic. The structural thinker who tests the argument rather than the facts. The Critic reads an output and asks: where does this reasoning break? What was assumed but never stated? Where would a competent adversary find the hole? Where is the model substituting plausibility for inference? They are the editor who sees that the article’s central claim is not actually supported by the evidence presented. They are the consultant who sees that the slide deck’s conclusion does not follow from the analysis. They are the policy reviewer who sees that the strategy paper is internally consistent but resting on a premise the world has already moved past.

Critics come from editorial, academic, consulting, philosophy, debate, senior product, and senior strategy backgrounds. Their habit is to compress an argument to its load-bearing claims and test each one against the rest. Best fit: strategy outputs, analysis memos, decision papers, opinion pieces, planning documents — any artefact where the bug is in the reasoning rather than the data.

The Critic’s risk is the “anti-everything” failure mode: the Critic who cannot say “ship it” turns into the team’s friction tax. The mature Critic disagrees fast, supports fast, and earns the right to be heard by being correct often.

• The Auditor. The systematic verifier who tests against a documented standard. The Auditor’s value is reproducibility. They check that the output conforms to the policy, the regulation, the contract, the SOP, the data-handling rules, the bias controls, the licensing terms. They produce a record. They write the runbook. They build the eval harness that the rest of the team can run. They are why the company passes the regulatory inspection.

Auditors come from internal audit, compliance, quality engineering, model-risk management, financial controls, and security operations. Best fit: regulated workflows where the question is not “is this good?” but “can we prove we checked?”. The EU AI Act has just made this archetype non-optional in every Annex III system. The 2026 Deloitte State of AI in the Enterprise report puts only one in five companies as having a mature governance model for autonomous agents — meaning four out of five companies are currently running production AI workflows without the Auditor seat that, in twelve weeks, regulators will start asking by name.

The Auditor’s risk is process for its own sake — the runbook that grew to forty pages because every prior incident added a line, and which no one actually follows. The mature Auditor prunes ruthlessly, automates the routine checks, and reserves human review for the cases that matter.

• The Red Team. The adversarial tester who tries to break the system before someone less friendly does. The Red Team’s value is creative attack. Prompt injection. Jailbreak. Edge case. Bias probe. The question the model has not been asked yet because nobody on the build team thought to ask it. The Red Team’s job is to be the most resourceful adversary the system will encounter — under controlled conditions, before the actual adversary arrives.

Red teams come from security research, penetration testing, journalism, investigative analysis, and a small but growing number of explicitly trained AI-safety programmes. The talent market for this archetype is the tightest of the four. Microsoft has stated publicly that skilled LLM security practitioners are in high demand and low supply. Indeed has 70+ remote AI-red-team listings open at any given moment. Entry-level comp has crossed $90k at the high end. Director-level AI-focused security roles are commanding $250k–$500k+ at major firms. By projection, 60% of organisations will be using AI red-teaming in 2026.

The Red Team’s risk is theatre. There is a class of “red team” engagement that produces a glossy report nobody reads, finds nothing the team did not already know, and tells the procurement department what it paid to hear. The mature Red Team is uncomfortable to host, reports findings the build team did not want to know, and is treated as a strategic peer rather than a vendor.

None of the four is a hierarchy. A serious verification function has at least two of them, often three. The Domain Expert and the Auditor together cover most regulated production workflows. The Critic and the Red Team together cover most strategy-and-policy work. The Auditor and the Red Team together cover most security-critical agentic systems. A great team has all four — and pays for it.

The mentoring problem nobody is talking about

Here is the second-order failure mode, and the one that ties this piece back to The Apprenticeship Implosion and The Originality Tax: we are not training AI Verifiers.

Universities still train generation. The undergraduate writes the essay, builds the model, ships the prototype. The MBA still trains structuring; the case method is, at its core, a generation method — produce a recommendation, defend it. The bootcamp trains shipping. The engineering rotation programme trains feature delivery. Almost no professional formation pathway, at scale, trains the skill of reading an AI output and locating what is wrong with it, under time pressure, with consequence attached. The reading-against-the-grain instinct that a great senior editor, or a great senior partner, or a great senior reviewer has — that is what an AI Verifier is, and it is a skill the current pipeline does not produce.

Worse: junior people, the ones who should be apprenticing into the skill, are now being asked to produce more, faster, with less mentor time, because their managers’ attention is being absorbed by AI orchestration. The very generation that should be learning verification under apprenticeship conditions is instead being optimised away from it.

This is the structural problem the next eighteen months will surface, and the post-mortems will name. Most of the high-profile AI failures of 2026–2027 will not be failures of the model. They will be failures of verification — preventable, catchable, named in the audit findings. The model produced a confident wrong answer. The system shipped it. The AI Verifier seat was vacant, or staffed by someone whose own training had not given them the muscle to push back.

What this means

• If you are early in your career. Stop optimising your résumé only for what you can produce with AI. Add what you can verify in spite of AI. Build an AI Verifier portfolio. If your instinct is the Domain Expert’s — depth in one field — write the annotated review packs. Read the model output in your domain, find the errors, write up the patterns. If your instinct is the Critic’s — argument-testing — keep a structured log of critiques: AI outputs you tested, premises you found unstable, conclusions you reversed. If your instinct is the Auditor’s — systematic — build evaluation frameworks and publish them; ship the runbooks; document the controls you would put on a production agentic workflow. If your instinct is the Red Team’s — adversarial — submit jailbreaks and bias probes to the bug-bounty programmes the major model providers now run; build a public portfolio of findings.

The market signal in eighteen months will not be “I can ship with AI”. It will be “I can be trusted to check what AI ships”. Build that résumé now, while the market has not yet adjusted to it.

• If you are hiring. Add at least one AI Verifier seat to every team running a production agentic workflow. Most organisations have zero. The Deloitte 2026 numbers say only one in five firms has a mature governance model. The other four in five are running on velocity and luck. By August, the EU AI Act starts assigning a cost to that luck, and the major US regulators are tracking close behind.

Hire for the archetype, not the title. Most candidates will not call themselves “AI Verifiers” because the title does not yet exist in a stable form on job boards. They will call themselves senior tax partners, principal reviewers, model-risk officers, internal auditors, security researchers, senior editors, ML safety specialists. Read the work, not the label. The market gap, today, is a labelling problem more than a supply problem; the supply will tighten quickly once the labels stabilise.

• If you are leading. The Operator without the AI Verifier is a velocity bet without a brake. You are paying for speed and absorbing risk you have not measured. Three things to do this quarter. First, name the AI Verifier role explicitly on every AI-touching team — not “we have a human in the loop” but “Marie is the Domain Expert AI Verifier on the underwriting workflow; Jean is the Auditor; this is the documented threshold for escalation to the Surgeon”. Second, fund the role at parity with the Operator role. If your Operator is paid $X, your AI Verifier is paid $X. If you can only afford one, you have an AI Verifier-first problem, not a budget problem. Third, mandate the documentation. Every shipped AI output crosses an AI Verifier signature with a recorded confidence assessment. No exceptions, no shortcuts. This is the audit trail.

The organisations that do this in 2026 will be the organisations that pass the audits, win the regulated contracts, and survive the first wave of AI-incident lawsuits in 2027. The organisations that don’t will discover — too late — that the cheapest cost of all was the AI Verifier role they did not hire.

The uncomfortable truth

Generation makes you fast. Verification makes you trustworthy. Right now, the market is paying for fast. The training pipelines, the bootcamps, the MBA programmes, the corporate L&D budgets, the venture term sheets — all of them, today, optimise for generation. Build with AI. Ship with AI. Demo with AI. Go faster. Go faster. Go faster.

The market that survives 2027 will be paying for trustworthy. The premium will move — quickly, in some sectors; slowly in others — from the people who can produce the most with the least friction to the people who can certify what shipped, when, under what controls, with what confidence. The premium will move because the cost of being wrong will move. A few visible AI-driven failures, a handful of EU AI Act fines, one or two negligence lawsuits where the verification trail was the deciding evidence — and the market resets.

The AI Verifier is the role most current strategy decks are missing entirely. Not because it isn’t obvious — it is obvious — but because it does not look like a growth story. It looks like a cost. It is a cost, in the way insurance is a cost, and a brake is a cost, and a seatbelt is a cost. It is also the cost that lets the rest of the system run at speed.

Most organisations will discover this the hard way. A small minority will discover it now, hire the four archetypes early, build the documentation infrastructure, and quietly accumulate a compounding advantage that — by 2028 — looks like luck and is in fact preparation.

Next week: the Workflow Designer — the role that decides, before the AI Operator orchestrates and the AI Verifier checks, where the AI gates and the human seats actually meet in the workflow. The fourth piece of this series, and the one that ties the architecture together.

In October 2024, Gartner published a forecast that, if you read it slowly, has a particular cruelty to it. By the end of 2026, one in five organisations would use AI to eliminate more than half of their middle management roles. Read past the percentages and what the sentence actually says is this: within 24 months, in every fifth company, the people whose business cards currently say “Senior Manager”, “Director”, or “Head of” will, in the majority, not have jobs that match those titles any more.

It was a forecast. The forecast is now nearly resolved. Microsoft’s security engineering unit has doubled span of control from 5.5 to 10 direct reports per manager. Gusto’s analysis of US firms shows the average supervisor moved from three direct reports in 2019 to six in 2025. McKinsey has cut roughly 5,000 internal roles since 2023 — about a tenth of its workforce — and senior partner Rob Levin has described the operating model that follows in a sentence that should be pinned to every CHRO’s monitor: “a more flat network of human teams supervising AI agents.” Middle-management share of layoffs rose from 20% of all cuts in 2019 to 32% in 2023, a roughly 60% increase in their share of the pain.

Last week, in the first piece of this series, I introduced the five AI engines — generative, predictive, perceptive, agentic, optimisation — that will run the 2030 enterprise. This week’s question is the structural follow-up, and it is the one most organisations are not ready for: who runs them?

The answer is not “your existing middle manager, with an AI Copilot.” It is a structurally different role I’ll call the AI Operator — and importantly, it is not one role but four. The AI Operator arrives as four named archetypes: the AI Conductor, the AI Translator, the AI Mechanic, and the AI Surgeon. Each is suited to a different kind of workflow. Each is recruited from a different pre-AI profile. Two or three of the four typically have to coexist inside any production stack of any complexity. This piece is about what each archetype does, what they replace, who makes the jump into each one, and — most uncomfortably — why most of the people currently holding “manager” in their titles will not be the ones to occupy the new layer.

“Now we’re moving to this more flat network of human teams supervising AI agents.” — Rob Levin, Senior Partner, McKinsey, 2025

That sentence is doing a lot of work. It quietly replaces the entire mental model of an organisation as a stack of human reporting relationships with a different mental model: a network of small human pods, each of which orchestrates a fleet of non-human workers. It also implies, with no ceremony, that the supervisory layer in between — the layer that has dominated white-collar work since Sloan and Ford — is no longer the load-bearing structure.

What the middle manager actually does today

If you strip the title back to its operational components, the middle-management role is three jobs stitched together. They are recognisable in every department, every industry, every flavour of corporate work.

• Information triage. The middle manager reads the reports, the dashboards, the customer feedback, the engineering tickets, the financial summaries. They turn that intake into structured output: a status update for the next layer up, a brief for the next layer down, a translation across the lateral function that doesn’t speak the same vocabulary. They are, in network terms, a routing node — taking dense raw information from one part of the org and reformatting it for another.

• Work distribution. They decide who picks up the new ticket. They negotiate priorities when two stakeholders both want a thing now. They run the morning standup that aligns the four sub-teams. They escalate the blocker. They reshuffle the schedule when someone calls in sick. They are, in operating terms, the scheduling and dispatch layer.

• People coordination. They do one-on-ones, performance reviews, hiring loops, growth conversations, career planning, conflict mediation, and the slow, accumulating work of building a team that works well together. They are, in human terms, the steward of the unit’s social fabric.

For most of the 20th century, fusing these three jobs into one human role was not just sensible: it was structurally necessary. The same person had to read everything anyway. Information triage was the most expensive part of the role, and once you had paid the cost of one human reading everything, you might as well also have them distribute the work and coordinate the people. Three jobs, one salary, one office, one expensive sip of organisational attention.

That logic held until 2023.

How AI absorbs two of the three

The mechanism is unsentimental. Generative AI absorbs information triage — at first imperfectly, then competently, then, in well-tuned setups, better than the human did. Reports get summarised, dashboards get narrativised, customer feedback gets clustered into themes, engineering updates get translated into language the rest of the business understands. The work that used to consume two-thirds of a manager’s day is now a function call.

The agentic stack absorbs most of work distribution. Tickets route themselves, priorities resolve themselves based on stated rules, standups become asynchronous summaries generated overnight, schedules optimise themselves around availability and dependencies. The dispatcher function — once human — becomes a coordination layer that runs in the background of the workflow. The orchestration patterns are explicit: hierarchical supervisor-worker, swarm, pipeline. Databricks, Microsoft, IBM, and a dozen agent-framework vendors are shipping the reference architectures for it now.

What is left over is people coordination. Which is real, valuable, irreducible work — but it is also, on its own, a fraction of a role. You do not need a full-time manager whose only job is one-on-ones and hiring loops. You need that work done well by someone whose actual job is something more substantial.

This is the structural fact most organisations have not faced. When two-thirds of a role’s day-to-day evaporates into a workflow, you do not solve it by adding an AI Copilot to the manager and asking them to be 40% more strategic. You redraw the role.

The role you redraw to is the AI Operator.

What the AI Operator actually does

Three things, and none of them are what the middle manager does today.

• Stack design. Which AI engines run which parts of which workflow. Where the verification gates sit. Where a human stays in the loop and where the agent has authority to act unsupervised. What the escalation paths look like when something exceeds the agent’s confidence threshold. This is the work of an architect, not a coordinator. It requires knowing what each of the five engines can and cannot do, knowing where their failure modes sit, and knowing how to compose them into a workflow that produces a result the business can defend.

If you have not designed a multi-step workflow that involves more than one AI capability and more than one verification gate, you have not yet done the work an AI Operator does.

• Failure-mode engineering. What happens when the agent hallucinates a customer name. What happens when the predictive model drifts because the training distribution no longer matches the operational one. What happens when the perceptive model fails silently — confidently reporting “no defect” while the line keeps producing defects. What happens when the optimisation engine maximises the wrong proxy and the business goes sideways.

The AI Operator’s job is to know — in advance, by design — what failure modes exist and what the escalation, recovery, and fallback paths are. The job is closer to an SRE designing for fault tolerance than a manager designing for performance reviews. Gartner forecasts that 40% of agentic AI projects launched in 2025 and 2026 will be cancelled by 2027 — overwhelmingly not because the technology fails, but because nobody designed for the failures. That gap is the AI Operator-shaped hole in most current organisations.

• Throughput accountability. A small human team — three to five people, increasingly — plus a fleet of agents that may number in the dozens. McKinsey’s reference number, from Rob Levin, is that 50 to 100 agents can be supervised by two or three humans in a well-designed setup. The AI Operator owns the team’s output, the agents’ output, the cost of running the stack, and the quality bar across the whole.

This is closer to a plant manager than a people manager. Outcomes are measured at agent-level metrics, not human-hour productivity. The dashboard the AI Operator watches is not “tickets completed per person per week”; it is something more like “successful workflow completions per dollar of compute, with verified output, broken down by failure mode.”

The shape of an AI Operator’s day

To make this concrete: where the middle manager’s calendar in 2023 was 60% meetings, 25% email, 15% strategic thinking, the AI Operator’s calendar in 2027 looks closer to 30% workflow design, 25% failure-mode review, 20% quality audits and agent fine-tuning, 15% one-on-ones with the small human team, and 10% upward and lateral conversations with adjacent AI Operators.

The meetings that remain are mostly substantive: a weekly review of agent failure cases, a fortnightly stack-architecture review with peers, monthly performance conversations with the small human team.

The status meeting — the one that defined a generation of corporate work — is, for the AI Operator, gone. The agents file the status.

The AI Operator reads it the way an air-traffic controller reads a board: looking for the one signal that needs attention.

The four AI Operator archetypes

The AI Operator does not arrive as a single job. The work splits cleanly into four archetypes, each one suited to a different kind of workflow and a different temperament. Most production stacks need two or three of them spread across two or three humans. A handful of small teams will have someone who plausibly does all four — but they are rare, and increasingly expensive.

The four are the Conductor, the Translator, the Mechanic, and the Surgeon. They are not levels of seniority. They are flavours of the role.

• The Conductor. The Conductor sees the whole stack. They know which of the five engines from last week’s piece — generative, predictive, perceptive, agentic, optimisation — runs which part of which workflow, in what order, with which handoffs. Their value is sequencing. They are not the deepest practitioner in any single engine; they are the one who knows, when the workflow needs to move from a generative draft to a predictive risk score to an agentic action to an optimisation pass, where each baton change happens and what each downstream engine needs from the one upstream.

Conductors come from product management, technical programme management, and engineering team-lead backgrounds. They have shipped systems where the timing and dependency structure was the design problem. The instinct that makes a Conductor is the instinct to draw the workflow on the whiteboard before opening the editor — to see the score before playing it. The best Conductors I have observed in 2025 and 2026 were senior PMs who had previously shipped pipelined ML products; they already had the language for “this output is the input to that”.

The risk for a Conductor is over-orchestration. A workflow that has too many engines, too many gates, too many handoffs is also a workflow that breaks at every seam. The mature Conductor designs for the fewest crossings that produce the required result.

• The Translator. The Translator lives at the seams between functions. Their value is carrying intent across boundaries without it being deformed. A finance team articulates a need in cash-flow language; the workflow has to be specified in data-quality and confidence-threshold language; the customer-facing team needs to know what to say when the agent returns a low-confidence answer. Each translation is an opportunity for meaning to be lost, garbled, or quietly stripped of the constraint that mattered most. The Translator’s job is that nothing gets lost.

Translators come from hybrid backgrounds. Data analyst plus product. ML engineer plus business operations. Growth lead plus customer success. They are the people who have been on at least two sides of a thing and can speak both dialects without thinking about it. In the 2026 market the named title for this profile is usually “AI Product Manager” or “Workflow Architect”; the actual skill is fluency at the joins.

The risk for a Translator is performing translation without doing it. There is a class of person who sounds like a Translator — uses the right vocabulary in both rooms — but is structurally a Status Theatre Manager who happens to have an LLM in their toolkit. The test is whether the workflows they describe actually run and behave as advertised when you check.

• The Mechanic. The Mechanic lives in the failure logs. Their value is diagnostic. When agent confidence is degrading on a workflow and nobody can say why, the Mechanic is the one you call. They read the eval traces. They re-run the prompt against the held-out set. They check whether the embedding model has drifted. They notice that the perceptive model has started misclassifying a particular SKU since the lighting changed in the warehouse. The Mechanic’s instinct is that something is wrong and the cause is somewhere specific, and they will not be satisfied with the team’s first guess.

Mechanics come from site reliability engineering, MLOps, customer success in technical products, and quality engineering. They have a long history of being woken up at 3am to find the cause of an outage and an even longer history of being unimpressed by anyone who declares the system “mostly fine”. The single most undervalued profile in the 2026 talent market is the Mechanic. The market still pays them like operations people. By 2028 they will be paid like the architects they are — because organisations that lose their Mechanics lose their agents within months.

The risk for a Mechanic is becoming the bottleneck. A team that routes every diagnostic question to one human will, within a quarter, have a queue. The mature Mechanic builds the eval frameworks, the dashboards, and the runbooks that let the rest of the team diagnose without them — and saves themselves for the truly novel failure.

• The Surgeon. The Surgeon does not run the day-to-day workflow. They do not sit in the standup. They are not on the dashboard rota. They are on call for the exceptional cases — the ones where the agent has flagged confidence below the threshold, or the decision is too consequential to automate, or the situation is too edge-case for the model to be trusted on. The Surgeon’s value is precise, high-stakes judgement on the cases that matter most.

Surgeons come from senior individual contributors in judgement-rich domains: senior underwriters, principal engineers, senior consultants, attending physicians, senior legal counsel. They are the people whose careers have been built on being trusted with the call that nobody else could quite make. In an agentic system, their work is not displaced by AI; it is concentrated by it. The routine cases the Surgeon used to also handle are now handled by the agents. What is left is the residue — the 2% of cases that are dense with consequence — and the Surgeon’s day shifts entirely toward those.

The risk for a Surgeon is over-intervention. If the Surgeon is pulled into every borderline call, they re-introduce a human bottleneck into a system designed to run without one. The mature Surgeon designs (with the Conductor) the confidence thresholds and the escalation conditions, then steps back and only handles what crosses the line.

None of these four is a hierarchy. None is more senior than the others. A workflow without a Mechanic is brittle. A workflow without a Conductor is incoherent. A workflow without a Translator gets the intent wrong. A workflow without a Surgeon gets the high-stakes case wrong. The four are complementary — and at the team scale we are now operating at (three to five humans plus a fleet of agents), two of the four often have to live inside the same person.

Who actually makes the jump

The honest answer is: a minority of current middle managers, and not the ones the current succession plan would have predicted.

The middle managers who make the jump into any of the four archetypes are the ones who, today, already do some of the AI Operator work.

They run the architecture-review call. They own the failure post-mortem. They are the manager who, when the team is debugging, sits in with the engineers rather than waiting for the upward summary. They have one foot in the actual work product, not just the dashboard about it.

The middle managers who do not make the jump are the ones whose value, on inspection, lived almost entirely in three patterns:

• The first is meeting density — calendar full, decisions made in rooms, value measured in attendance.

• The second is upward narrative — translating the team’s work into the language the layer above wants to hear.

• The third is approval gatekeeping — being the necessary signature, the rubber stamp before the work moves forward.

None of these three patterns is wrong; all three were valuable in a hierarchy. None of them survives the move to a flat network of human teams supervising agents.

There is no shame in being in the second category. The role those managers were promoted into is being deprecated; the organisational technology has changed underneath them. There is, however, a refusal of clarity in not telling them so.

The mentoring problem nobody is talking about

One thing worth being explicit about, because it is the legitimate cost of this transition. The middle-management layer was, for a generation of knowledge workers, also the apprenticeship layer. The junior consultant learned how to think by watching the engagement manager think. The junior analyst learned what good looked like by watching the senior associate edit their work. The middle layer was where you absorbed taste, judgement, organisational instinct, and craft.

This is the same problem I wrote about in The Apprenticeship Implosion a few weeks ago, and it is real. The AI Operator role is a senior individual contributor role with a small team and a large fleet of agents; it does not have the bandwidth or, frankly, the structure for the slow patient transmission of craft that the middle layer used to provide. Junior people in 2027 will need to find their apprenticeship somewhere else — in deliberate communities, in mentorship programmes, in cross-team rotations, in working closely with one AI Operator instead of through a chain of supervisors.

Organisations that solve this deliberately will compound a talent advantage. Organisations that don’t will quietly hollow out their leadership pipeline, and notice in 2030 that they no longer have the people to fill the next generation of AI Operator roles.

What this means

• If you are early in your career. The AI Operator role — in any of its four flavours — is the highest-leverage management path in the 2026 economy. The market has not fully named it; job postings are still labelled “AI Program Manager”, “Workflow Lead”, “AI Operations Manager”, “Agent Supervisor”, “AI Ops Lead”. Same shape of job under different titles. Pick the archetype that matches your instinct, not the title. If you naturally see the sequence of a workflow before you see any single piece of it, build a Conductor portfolio: ship multi-engine workflows. If you naturally translate between two languages — data and product, ML and ops, engineering and customer success — build a Translator portfolio: ship documented handoffs across functions. If you naturally cannot stop tugging at why something is failing, build a Mechanic portfolio: ship eval frameworks and failure-mode playbooks. If you are the senior individual contributor who carries judgement on the calls nobody else can make, build a Surgeon portfolio: ship the case studies of the exceptions you caught.

Stop building a generic “leadership” résumé. Build an archetype-specific résumé.

• If you are hiring. Stop hiring middle managers. Start hiring for one of the four. Most organisations need a Conductor and a Mechanic urgently — those are the two roles that determine whether a production stack runs at all. The Translator becomes essential the moment a workflow touches more than two functions. The Surgeon is the final hire and the one that determines whether the system can be trusted in regulated or high-stakes work.

The procurement-side instinct will be to upgrade existing managers with AI training. The data does not support it. Across the firms I have seen do this in 2025, the success rate of retraining traditional middle managers into any of the four archetypes is somewhere in the 15–25% range — and the success rate is materially higher for those who were already, in their existing role, doing some of the archetype’s work (running the architecture call, owning the post-mortem, sitting in with the engineers). The success rate of hiring Mechanics and Conductors directly into the role is materially higher, and the comp gap — for the moment — still favours the buyer.

This will not last. By 2027 the four archetypes will be named, the market will be tight, and the firms that hired early will be holding the talent the rest of the market is trying to pay 50% more to acquire.

• If you are leading. The kindest thing you can do for your middle layer is be honest about the trajectory, and specific about the path. “Become an AI Operator” is not actionable. “I think you have the instincts of a Conductor — here is the training, here are the workflows you can shadow, here is the timeline” is actionable. Same for Translator, Mechanic, Surgeon. The named taxonomy is itself the kindness, because it tells people what to study, what to ship, and what to put on the portfolio.

You owe them three things. First, an honest conversation about where their existing role is going. Second, a specific archetype-shaped path with training, support, and time. Third, a real off-ramp — including financial — for the ones who will not, or should not, make that jump. Most will fall into the third category. That is not a moral failure; it is a structural fact about how much the role has changed.

The uncomfortable truth

Middle management, as a category of work, was a 20th-century technology.

It solved a real problem: in an organisation of more than a hundred humans, information had to be moved up, down, and sideways, and humans were the only entities capable of doing that moving. The middle manager was the routing protocol, the dispatcher, and the people-coordinator, fused into a single role for efficiency.

The 21st-century version of that routing protocol is the agentic stack. The 21st-century version of the dispatcher is the orchestration framework. The 21st-century version of the people coordinator is a smaller fraction of one person’s calendar.

The 21st-century human role that sits alongside this stack is not a smaller middle manager. It is the AI Operator: an architect of workflows, an engineer of failure modes, an owner of throughput across a small human team and a large fleet of agents. And the AI Operator is not one role; it is four — Conductor, Translator, Mechanic, Surgeon — each recognisable, each learnable, each with a clear pre-AI lineage. Most organisations will not promote their way to those four roles.

They will hire from outside, often awkwardly, often at premium comp, and discover painfully over the next 24 months that the internal upgrade path was never going to work in the volume the consulting decks suggested it would. The middle layer is not being upgraded. It is being structurally replaced — by four named roles that the current org chart has nowhere to put.

For the people currently sitting in that layer, the next 18 months will look like a choice that mostly is not theirs to make.

The role they signed up for is being deprecated. Some will make the jump into one of the four. Most will not. The honest organisations will name the four archetypes out loud, offer the specific paths into each, and help the people who will not make the jump land somewhere they can be the version of themselves they want to be. The dishonest organisations will say nothing, retitle a few people “AI Operations Lead”, run a workshop, and quietly discover in 2028 that they have neither the leadership pipeline nor the AI Operator capacity to compete.

Next week: the AI Verifier. The second role that emerges, and the one most strategists are missing entirely — the human counter-weight to a fleet of fluent, plausible, frequently-wrong agents. If the AI Operator runs the stack, the AI Verifier is the reason any of its output can be trusted.

In December 2025, McKinsey released its ninth annual State of AI report. Inside the dense pages of statistics, one number stood out for its quietness. 88% of executives surveyed planned to increase AI-related budgets in 2026. The accompanying breakdown of what they intended to spend that money on was where the strangeness began.

Roughly nine in ten of those budget increases were earmarked for one specific AI capability. Generative — the family of large language models that produce text, summaries, drafts, code, images. The engine everyone has been touching, demonstrating, procuring, and proudly listing on quarterly earnings calls since 2023. The other four engines barely featured.

This piece is the first of a five-part series I’m calling The Organisation of the Future. Over the next five Wednesdays I’ll lay out — across roles, structures, and stack — what the 2030 enterprise actually looks like once multiple AI capabilities are integrated and orchestrated. The pieces stand alone, but they compound. To talk usefully about how AI restructures work, we need to start with the parts.

There are five. Most leaders are treating them as one. That is the most expensive vocabulary mistake of the decade.

“Traditional AI adoption has climbed to 72% over the past eight years, but organisations are adopting agentic AI rapidly, well before they have orchestration strategies in place.” — MIT Sloan Management Review × BCG, 2025

That sentence describes a particular kind of breakage. You have one type of AI everywhere. You bolt on another type of AI without a plan for how the two interact. Then you wonder why the productivity numbers don’t add up.

This piece names the five engines, what they do, where they’re real, where they’re brittle, and — most importantly — why the field’s collective fascination with one of them is the structural mistake that will determine which firms compound advantage by 2030 and which firms get stuck.

A short, honest history of why we got here

The vocabulary collapse — the way “AI” came to mean “generative AI” in mainstream business conversation — has a date. It happened over roughly 18 months between late 2022 and mid-2024. ChatGPT launched. Then GPT-4. Then Claude, Gemini, Mistral, Llama. Boards started asking about AI. The market for “AI strategy” presentations exploded. Vendors learned that the word “AI” sold faster than “machine learning” or “computer vision” or “operations research” had ever sold. The semantic field shrank.

The shrinkage was efficient. It made the technology procurable. A board could be told “we are deploying AI” and understand, instantly, that there would be a chatbot. The procurement department could be told “buy the AI” and know which licences to negotiate. The McKinsey-deck-driven AI strategies of 2023 and 2024 worked because the word had narrowed to a single referent.

What was lost was the rest of the field. Predictive AI — running quietly in fraud detection, credit scoring, demand forecasting since the 90s — became unfashionable, almost embarrassing. Computer-vision projects struggled for budget because they didn’t fit the chatbot shape of the conversation. Reinforcement-learning teams were quietly disbanded at firms that “moved their AI investment to generative”. The legacy AI infrastructure of most enterprises was treated as old, even when it was producing more measurable value than the new generative initiatives.

By 2026 the cost of that shrinkage is becoming visible. The firms that “won” with generative AI are the ones who already had the other engines running and integrated the new one cleanly.

The firms that bet on generative as a single solution have spent two years discovering that a chatbot, on its own, does not transform an operating model.

To get the rest of the decade right, we need to recover the vocabulary. Five engines. Each with a distinct shape of value. Each with a distinct cost of ownership. Each with a distinct failure mode. None of them, on its own, is the strategy.

Engine 1: Generative

What it does. Produces content. Text, code, images, audio, video, structured outputs. By predicting likely continuations of an input.

Where it is real. Knowledge work involving drafting, summarising, translating, restructuring, idea generation. Customer support. Marketing first drafts. Code completion. Internal-search-with-explanation. Anywhere a human used to stare at a blank page or a long document.

Where it is brittle. Anywhere accuracy matters more than fluency. Generative AI does not have a concept of true. It has a concept of likely. Those are different things. The Deloitte fabrication scandals of late 2025 — A$440,000 worth in Australia, a near-million-dollar healthcare report in Canada — are not anomalies. They are what generative AI does when nobody verifies. It produces fluent, plausible, well-structured wrongness.

The deeper truth. Generative AI is a median engine. It pulls work toward typicality (I wrote about this last week in The Originality Tax). That makes it perfect for low-stakes drafting and dangerous for high-stakes thinking. The professionals who treat it as a writing assistant will be fine. The ones who treat it as a thinking partner are quietly outsourcing their judgement to a probability distribution.

Engine 2: Predictive

What it does. Forecasts an outcome from historical patterns. A churn score, a demand curve, a credit risk, a fraud likelihood, a delivery ETA, a sensor failure window.

Where it is real. Insurance, banking, supply chain, energy, healthcare diagnostics, fraud detection, predictive maintenance. Most of the AI value already in production at Fortune 500 firms is predictive — and has been since long before “AI” meant chatbots. The numbers are quietly enormous: 15-18% reductions in inventory cost from hierarchical reinforcement-learning approaches, 20-30% reductions in working capital from AI-driven demand forecasting and inventory optimisation. Nobody puts those on a slide because they are not interesting.

Where it is brittle. Distribution shifts. The world changes; the model was trained on what came before. The 2020 demand-forecast model that survived COVID is famous because nearly all of them did not. Predictive models also offer a single number with high confidence, which encourages humans to defer instead of think. That is its own failure mode.

The deeper truth. Predictive is the unsexy engine that has paid every bill in enterprise AI for the last twenty years. It does not write you a memo. It tells you what is likely to happen, how confidently, on what basis. It is the engine generative AI conspicuously cannot replace — though many vendors have tried.

Engine 3: Perceptive

What it does. Turns raw sensor data — pixels, audio waveforms, depth maps, vibration signatures, electrocardiograms — into structured states. The defect on the assembly line. The tumour on the scan. The shoplifter at the self-checkout. The fatigue in the operator’s voice.

Where it is real. Manufacturing (32% of computer-vision deployment by 2025, per Markets and Markets), healthcare imaging, automotive driver assist, agriculture, retail loss prevention, security. Visual AI systems can now detect assembly defects in under 200 milliseconds and reduce unplanned downtime by 50%. The market — $23 billion in 2025, projected $63 billion by 2030 — is one of the fastest-growing in enterprise tech, and almost none of that growth is in chatbot-shaped form.

Where it is brittle. Edge cases. The model was trained on what was photographed. The condition that wasn’t photographed because nobody knew it was a condition is the condition the model misses. Perceptive AI also fails silently — a vision system that should detect a defect and doesn’t has no way of telling you it failed. It just confidently says “no defect.”

The deeper truth. Perceptive AI is what turns a software product into an operational one. It connects the digital model of the business to the physical state of the business. The firms that have not yet wired up perceptive AI to their physical operations are running the digital twin of an organisation they cannot see.

Engine 4: Agentic

What it does. Takes actions in the world. Opens browsers. Calls APIs. Writes to databases. Dispatches emails. Escalates tickets. Runs multi-step workflows that previously required a human in the loop at every junction.

Where it is real. Tightly-bounded enterprise workflows: ticket triage, lead enrichment, basic customer-service resolution, internal-search-and-action, meeting prep, scheduling. The number of organisations experimenting with agentic AI is high — 62% by McKinsey’s count. The number actually scaling agents in production is much smaller — 23%. The number running agents reliably enough to take revenue or compliance risk is smaller still.

Where it is brittle. Open-ended environments. Long task chains where errors compound. Anywhere the cost of a mistake is high relative to the cost of the action. Gartner forecasts that 40% of agentic AI projects launched in 2025-2026 will be cancelled by 2027 — not because the technology fails, but because the orchestration around it fails. Most enterprises do not yet have the verification, recovery, or escalation patterns required to deploy agents safely.

The deeper truth. Agentic AI is the youngest, most fragile, and most over-promised of the five engines. It is also, by 2030, likely to be the most consequential. The firms that learn to deploy agents safely — meaning with clear boundaries, fast fallback, and human verification at the high-stakes nodes — will reset what one mid-level employee can accomplish in a day. The firms that deploy agents naively will appear in postmortems.

Engine 5: Optimisation

What it does. Decides — given constraints, objectives, and dynamic state — what action minimises cost or maximises return. Reinforcement learning, mathematical optimisation, dynamic pricing, route planning, scheduling, bid optimisation, network management.

Where it is real. Logistics (route and inventory), pricing (e-commerce, ride-share, airlines), advertising (real-time bidding), energy (grid management), telecoms (network optimisation). Hierarchical reinforcement learning approaches now reduce inventory costs by 15-18% on real industrial deployments (OpenReview, 2025). Distributed AI-enabled control systems integrating IoT sensors, deep-learning forecasts, and RL decision policies are moving from research papers into operational pipelines.

Where it is brittle. Reward design. The optimisation engine maximises whatever you tell it to maximise. If the metric is wrong, the optimisation makes the wrong thing happen faster and at scale. Optimisation is also brittle to environments it has not seen — a pricing model that has never seen a recession is a model that has never seen a recession.

The deeper truth. Optimisation is invisible because it lives upstream of the user. Its outputs are decisions, not artifacts. You cannot demo optimisation in five minutes. The board cannot see what it does. Which is why it is consistently underbought in firms whose AI strategy is procurement-driven — and consistently the engine that produces the highest measurable ROI in the firms that have it running.

Why one engine is not a strategy

Most enterprises in 2026 are running a generative-AI strategy. They have a Copilot. They have a custom GPT. They have a chatbot or three. They have, in McKinsey’s framing, “scaled an agentic system” — meaning they have one workflow that calls a model. That is one engine.

A one-engine stack is a stack that can do one thing. It can write but cannot see. It can summarise but cannot predict. It can suggest but cannot decide. It can draft a fraud report — but it cannot detect the fraud. It can summarise a maintenance log — but it cannot tell you the bearing is two weeks from failure. It can write a customer email — but it cannot decide which customer to email first.

The firms that are compounding real advantage in 2026 are the ones running multiple engines in choreography. A perceptive model identifies the defect on the line. A predictive model estimates downstream production impact and supply implications. An optimisation model reschedules production around it. A generative model drafts the customer-comms package. An agentic model dispatches the field engineer with the right parts.

None of those engines is the headline story. The choreography is the story. And the choreography is the thing most enterprises in 2026 do not yet have a vocabulary for, much less a role to own.

What this means

For early-career people: stop building deep skill on one engine. The professionals who will compound value through 2030 will know how to interrogate at least three. Pick one to specialise in deeply — but be conversant in all five. The career risk of being “the prompt engineer” five years from now is identical to the risk of being “the Excel macro expert” was in 2002 — when Excel macros were still cool.

For hiring managers: the most undervalued profile in 2026 is the candidate fluent across multiple engines. They are rare because the market hasn’t named the role yet. Look for engineers who have shipped both an ML model and an LLM workflow. Look for product people who can talk fluently about both confidence intervals and prompt evals. They cost more. Hire them anyway. They will be the operators and workflow designers I describe in the next four pieces.

For leaders: audit your AI strategy for engine balance. If your line items are 90% generative, you have a chatbot strategy, not an AI strategy. Insist that at least two of the other four engines be on the roadmap inside twelve months. The competitor that ships a perceptive-and-predictive workflow before you ship your second chatbot is the one taking your margin in 2027.

The uncomfortable truth

Generative AI was easy to buy because it was easy to demo. Type something. Get something. The procurement decision became trivial. The success criteria became “did the demo work.” The strategy became “deploy a chatbot.”

The other four engines are hard to buy because they are hard to demo. They require integration with the messy parts of the business — sensors, logs, ERPs, OMS, dispatch, finance systems. They cannot be demonstrated in five minutes. They cannot be wrapped in a chat interface. They require, in the most literal sense, doing the work.

Most enterprises in 2026 will not do the work. They will buy more generative licences and call it a strategy. The boards will be told that AI investment has tripled. The press releases will land. The Copilot deployment numbers will be cited at the next earnings call. None of that constitutes an AI strategy. It constitutes a generative-AI strategy.

The firms that build the full stack — the ones running all five engines in choreography by 2028 — will, by 2030, be running operating models the rest of the field cannot match. The gap will not look closeable, because the rest of the field will still be staffing for a one-engine world. They will keep hiring prompt engineers when they need verifiers. They will keep buying chatbots when they need optimisation pipelines. They will keep procuring AI as a product when they should be building it as a stack.

The organisation of the future is not a one-engine organisation. It is not a chatbot wrapped in a corporate brand. It is a coordinated stack of generative, predictive, perceptive, agentic, and optimisation capabilities, run by people whose roles do not yet exist on most org charts, supervised by structures most boards have not yet drawn.

This series, over the next four weeks, is about who those people are, what they do, and how the org around them takes shape.

Start with the vocabulary. Five engines. Most leaders are treating them as one.

The strategic question of the next 36 months is not “which AI vendor.” It is: how many engines are you actually willing to run?

In 2024, two researchers — Anil Doshi and Oliver Hauser — published the most quietly unsettling study of the AI-and-creativity literature so far. They asked one group of writers to produce short stories with no help. They asked another group to use a generative AI to brainstorm. Then they had thousands of independent readers rate the results.

The headline finding was the optimistic one. AI helped. Stories produced with AI assistance were rated as more enjoyable, better written, and more creative — especially when the writer wasn’t a particularly creative person to begin with. Generative AI, the paper concluded, “enhances individual creativity.”

The footnote finding was the dangerous one. The AI-assisted stories, when compared to one another, were significantly more similar than the unassisted stories. Better individual outputs, narrower collective idea space. A social dilemma, the authors called it: “writers are individually better off, but collectively a narrower scope of novel content is produced.”

That single graph — individual creativity up, collective diversity down — explains something most of us have felt in the last eighteen months but couldn’t quite name. The pitch decks all say the same things in the same voice. The strategy memos arrive at the same recommendations. The think-pieces converge on the same takes. The LinkedIn posts have started to share a faint, unmistakable family resemblance.

We have more content than ever, and it sounds more alike than ever.

This is a piece about the invisible cost of producing that: the cognitive tax that every professional now pays, whether they notice it or not, whenever they sit down to do work that is supposed to sound like them.

I’m going to call it the originality tax. It is the new, real, unmeasured cost of producing work that doesn’t sound like everything else.

What “median machine” means, structurally

Let’s start with a piece of mathematics that almost nobody discusses honestly.

A large language model is, in its mechanics, a probability distribution over what comes next given what came before. It was trained on enormous quantities of human-written text, and it learned to produce the most likely continuation. Not the most original. Not the most surprising. The most likely. That is the loss function. That is the thing the model was optimised to do.

The implications are easy to miss because they sound innocuous. The model is “helpful”. It “smooths your prose”. It “polishes your work”. But what it is actually doing — at the level of the cost function it was trained against — is moving your text toward a predicted average of how text like this is usually written. That’s the only thing it can do. It cannot give you a continuation more original than its training distribution permits, because its training distribution defines what original even means to it.

The first three suggestions any AI tool gives you are, almost by definition, the three most predictable continuations. Not the most useful. Not the most precise. The most predictable. The probability mass concentrates near the centre of the distribution, and that is what arrives in your editor as a polite suggestion.

Once you see this, you cannot unsee it. Every “make it more professional” pass is a step toward the median professional voice. Every “shorter and clearer” suggestion is a step toward the median form of brevity. Every “tighten the introduction” recommendation is a step toward the median introduction. Each step is small. Each step is an obvious improvement, when judged locally. None of them feels like a loss.

You only notice the loss collectively, at the level of the field — in the Doshi & Hauser graph, in the indistinguishable consulting reports, in the strange feeling of reading three professionals in your industry and being unable to tell them apart.

The mechanism is not malicious. It is built into the architecture. And there is no version of using AI for writing that escapes it.

The autocomplete tilt: where it shows up first

If the median pull were only an aesthetic problem, we could shrug and move on. Everyone uses the same Microsoft Word. Most professional writing was already mediocre.

The deeper problem, surfaced by a series of papers between 2023 and 2025, is that the median pull does not stop at prose. It reaches into thought.

The clearest evidence comes from a 2023 Cornell study by Maurice Jakesch and colleagues, titled — with characteristic academic dryness — “Co-writing with opinionated language models affects users’ views”. The setup was simple. Participants were asked to write short essays about whether social media is good or bad for society. Some used a writing assistant tuned to gently suggest pro-social-media completions. Others used one tuned to suggest anti-social-media completions. The participants were not told.

The result was unsettling. After a single short writing session, participants exposed to the pro-leaning model expressed measurably more positive opinions about social media — and the participants exposed to the anti-leaning model expressed measurably more negative opinions. They reported these views as their own. Asked whether the assistant had influenced them, they said no.

“AI suggestions don’t just speed up writing. They shape what people think. Because the user is the one who chooses to accept the suggestion, the brain internalises it as an original thought.” — Jakesch et al., extended interpretation in PsyPost, 2024

The mechanism here is the bit that should keep professional services partners awake at night. When AI offers you a sentence and you accept it, you are not just borrowing a phrase. You are absorbing a frame. You think you wrote it because you typed it, but the typing was the surface action — the deeper cognitive act, the framing of the thought, came from the model. The user’s sense of authorship survives. The actual authorship moves elsewhere.

Multiply that by every professional who uses AI to draft anything — emails, briefs, decks, op-eds, strategy memos, performance reviews — and you start to understand why the field is flattening. It isn’t that everyone is being lazy. It’s that everyone is being subtly shaped by the same shaping force, and they cannot feel the shaping because the shaping happens at the level of what they take to be their own thoughts.

A 2025 CHI paper from Agarwal and colleagues found something even more pointed: Indian users of AI writing assistants progressively adopted Western prose styles, even when writing in their own language about their own culture. The tool has a default voice, drawn from its training distribution, which is overwhelmingly English-language and Western-coded. The default voice wins. Not just what is written — how it is written.

This is the originality tax, levied early and quietly. By the time the polished draft lands, the cost has already been paid in invisible currency — in the texture of how the writer thinks.

Three taxes, named

The originality tax shows up in three distinct forms. They are paid in different currencies, by different people, at different stages of the work.

• The autocomplete tax. The cost paid when a tool finishes your sentence and you accept the finish. You save time. You lose the thinking that would have happened in the moment of completing the sentence yourself. Over thousands of small acceptances, the texture of your prose drifts toward the model’s prose. Most professionals pay this tax constantly and unconsciously. They are not aware it is being levied.

• The polish tax. The cost paid when a finished draft is run through a “make this better” pass. The output is more competent — and less distinctive. Specific word choices that signalled a personality get replaced with generic high-frequency synonyms. Idiosyncratic structures get smoothed into standard ones. The piece is more publishable, and less recognisable as the writer’s. This tax is paid deliberately, but most writers don’t notice they’re paying it because the result looks “more professional.”

• The brainstorm tax. The cost paid when a writer asks AI to suggest angles, arguments, or framings. The angles offered are, by construction, the angles closest to what other people have already written about this topic. The writer feels generative — three options! — when in fact they are picking from three samples of the median. The writing that follows is competent and forgettable. This is the most expensive tax, because it is paid at the structural level of what the work even is.

You can ship a piece without paying the autocomplete or polish tax. It takes effort. You can write the awkward sentence the autocomplete keeps trying to smooth out. You can refuse the polish pass on the parts that are intentionally rough. Both are inefficient, both are slower, and both produce output that is more recognisable.

The brainstorm tax is harder to evade. To brainstorm without the model is to face the blank page — slow, frustrating, often unproductive in a single sitting. The model offers you something for the discomfort. Most writers, under deadline, accept.

What an over-taxed market looks like

If you wanted a single image of what an over-taxed market looks like, the recent series of consulting-firm AI debacles is the cleanest one available.

In October 2025, the Australian government revealed that a A$440,000 Deloitte report on welfare compliance contained AI-generated fabricated citations, hallucinated quotes attributed to a real federal-court judge, and several invented academic references. Deloitte refunded part of the fee. The story went global. A few weeks later, a Canadian provincial government discovered the same pattern in a near-million-dollar Deloitte healthcare report. Both clients had paid premium rates for premium expertise. They had received, in part, AI-tinted output that nobody had bothered to verify.

The temptation is to treat these as scandals about laziness. They are not. They are scandals about taxation. Inside Big-Four firms, the originality tax has been quietly accepted as a cost-of-doing-business. Internal accounts published in 2025 — including a leaked McKinsey post-mortem on its “Lilli” tool — describe consultants discarding up to half of AI-generated slides because the suggested frameworks were too generic; up to 25% of AI-drafted deliverables requiring substantial rewriting before they reached partner review; senior partners spending two to three additional days per engagement on quality control they didn’t used to do.

These are not “AI productivity gains.” They are the visible, measurable surface of the tax. The output is faster. The verification cost is enormous. The originality of the final product is, by most accounts, lower than what the same firms used to produce ten years ago.

And yet most professional services firms still report AI as a productivity win. Why? Because the tax is paid in a currency the income statement doesn’t track: distinctiveness. Voice. Edge. The things that, in the long run, make a firm worth hiring instead of any other firm. None of those show up in quarterly numbers. The savings do.

What gets lost when distinctiveness goes

There is a temptation here to romanticise pre-AI prose as if it were always good. It wasn’t.

Most writing has always been mediocre. Most consulting reports have always read like other consulting reports. The originality tax did not create the average; it deepened it.

What’s at risk is the stratum above the average. The professionals and brands whose work was, for whatever reason, recognisable. The firm whose memos felt different from the four other firms. The analyst whose voice you’d recognise blind. The designer whose work could be picked out of a deck without the byline. These were always rare. They were also, in a real sense, what made the market a market — the differentiating stratum that gave clients meaningful choices.

When the tax pulls everyone toward the median, the differentiating stratum thins. Not because those people stop existing — but because their work, run through the same tools as everyone else’s, comes out sounding more like everyone else’s. The signal weakens. Clients can no longer tell the firms apart.

The market response to this, eventually, is to start paying a premium for the people who somehow still sound like themselves. We are at the early edge of that. The boutiques that beat bigger firms in 2026 increasingly do so on voice. The newsletters that grow against the big institutional outlets win on voice. The individual contributors who get hired against teams win on voice. None of those are accidents. They are early arbitrage on the originality premium.

Practical implications

• For early-career people: the most valuable thing you can do this year is pay the tax deliberately. Write your first drafts without the AI. Notice your actual voice — what words you keep choosing, what rhythm your sentences have, what kinds of metaphors come naturally. Then use the tool to sharpen specific bits, never to speak for you. The delta between your unassisted writing and your AI-assisted writing is the most precise diagnostic you have of how much of your voice is still yours. Watch it.

• For hiring managers: stop treating speed of output as a signal. AI broke the relationship between speed and skill. The signal that matters now is the delta between someone’s AI-assisted output and their unassisted output. If they’re identical, you’re hiring the model. The juniors who will become your differentiated seniors in 2032 are the ones who can articulate, when asked, what they refused to let AI do for them.

• For leaders: every workflow that prioritises throughput over distinctiveness is taxing your firm’s originality whether you measure it or not. Your competitors using the same tools are converging on the same shape. Clients are starting to notice — slowly, then suddenly. The strategic question is which budgets you’re willing to spend protecting the people who still produce work that sounds like nothing else. They are your differentiation. They are also, currently, more expensive to retain than to lose.

The uncomfortable truth

Originality used to be free. Everyone faced a blank page. The blank page was democratic — it pulled nothing out of you, suggested nothing, finished no sentence. Whatever appeared, however clumsy, was yours.

The AI-assisted page is not blank. It is suggesting, finishing, polishing, every time. Producing something that isn’t the median requires effort the blank page never demanded. The tax is real. It compounds. It is paid invisibly, in cognitive currency, by every professional who uses these tools without explicitly resisting them.

The most uncomfortable part is not that the tax exists. It’s that most people paying it don’t know they’re paying it. They think the smoothed-out, AI-tinted version was their own voice all along. The mechanism by which AI shapes thought — the autocomplete that nudges, the polish that flattens, the brainstorm that median-tilts — operates beneath the level of conscious noticing. By the time you can feel it, your reference point has already moved.

The professionals and brands willing — and resourced — to pay the tax will become rare and valuable. The ones who can’t, won’t. They will sound, increasingly, like everyone else. Their work will be perfectly competent. It will also be functionally interchangeable.

In ten years, when the field has flattened further, the question worth asking will not be “did you use AI.” Everyone will have. The question will be: did your work still sound like you, after?

Most people will not be able to answer.

The few who can will own a market the rest will be too tired to compete in.

Software developers between the ages of 22 and 25 saw their employment fall by nearly 20% from 2024, even as headcount for older developers in the same firms continued to grow. That number, buried in a Stanford Digital Economy Lab paper from late 2025 with the deceptively cosy title “Canaries in the Coal Mine?”, is the most consequential statistic in the entire AI-and-work conversation right now.

It is also the one almost nobody is reading correctly.

The popular framing is robots-take-jobs: AI is eating the entry-level rung of the labour market, and a generation of young professionals is being locked out. That framing isn’t wrong, exactly. It’s just narrow. The deeper story isn’t that early-career people are losing jobs. It’s that we are unwittingly dismantling the system that has, for a century, produced the senior people we depend on.

The apprenticeship — the messy, slow, half-broken, economically inefficient way humans have always trained the next generation of professionals — is imploding. And we don’t have a replacement.

This is a piece about that implosion: where it shows up first, why it’s almost invisible to the firms causing it, and what it costs us when we cash in the seed corn.

A century-old contract, broken in eighteen months

Every knowledge profession runs on the same hidden contract, and it has barely changed since the modern firm was invented in the early 1900s.

A junior shows up. They are bright, ambitious, and largely useless. They are paid relatively little to do work that is, by design, beneath the seniors: the document review, the literature search, the first-pass code, the financial model that gets thrown away, the deck nobody will see. In exchange, they get exposure — to clients, to problems, to seniors thinking out loud. Slowly, over thousands of unglamorous hours, they absorb the tacit knowledge that turns information into judgement. The senior gets cheap leverage. The firm gets work done. The junior gets a career.

It is, viewed from the outside, an absurdly inefficient arrangement. We pay seniors enormous salaries to spend a meaningful chunk of their day giving feedback on first drafts that the senior could have written better in a third the time. We tolerate juniors making mistakes that anyone with five years of experience would not. We invest in people for years before they generate net positive value.

And it works. It worked for medicine, for law, for accounting, for engineering, for journalism, for architecture, for consulting, for software, for design, for almost every profession that depends on seasoned judgement. The inefficiency was the feature. We were paying for the production of expertise.

That contract just got broken — not on the senior side, but on the labour side. When AI can produce a passable version of a first-year associate’s deliverable in seconds, the economic case for hiring that associate quietly evaporates. And the market is doing exactly what economic theory says it should.

Entry-level postings in software development and data analysis dropped by roughly 60% between 2022 and 2024. A 2025 LeadDev survey of engineering leaders found 54% planned to hire fewer juniors specifically because AI copilots were letting their seniors handle more. Salesforce announced it would hire “no new engineers” in 2025. The Stanford team, looking at the most AI-exposed jobs across the entire US economy, found a 13% relative decline in employment for early-career workers in those occupations. Wages held; jobs disappeared. The labour market is adjusting through hiring decisions, not pay cuts — which is precisely why the adjustment is invisible quarter to quarter and catastrophic decade to decade.

Here is where almost everyone stops. The story is presented as a Gen Z problem, a tech-bubble correction, a passing storm. The bigger story is what comes next.

The pipeline you cannot see on the income statement

Senior expertise is not manufactured in a classroom. It is the residue of thousands of hours of low-stakes mistakes in low-stakes work. The associate who has read four thousand contracts develops a sixth sense for the one with a buried liability clause. The accountant who has reconciled a thousand inter-company ledgers can spot the kind of error that doesn’t even show up as an exception. The engineer who has debugged two hundred 3am production incidents can read a stack trace like sheet music. None of that comes from reading about it. None of it comes from watching someone else do it. It comes from doing it badly, many times, with consequences low enough to survive.

Harvard Kennedy School’s Project on Workforce calls this dynamic the “expertise upheaval,” and they argue that AI’s effect on the learning curve is its most important and least appreciated impact. When AI compresses the curve, it doesn’t just speed up training. It removes the substrate training was built on. You cannot develop judgement about an AI’s output if you have never done that work yourself. The judgement is what you get from doing the work. The work is the school.

“If the current generation of juniors never grapples with low-level problems because AI solves them automatically, they may never develop the deep intuition and tacit knowledge required for senior roles. By 2030, the industry may face a catastrophic shortage of true senior engineers and leaders — those capable of understanding the system below the AI abstraction layer”.

That block quote is from a synthesis of 2024–2026 employment data published by Rezi’s research team in early 2026. It is, as far as I can tell, the most honest thing anyone has written on the subject. We have no plan for how senior people will exist in 2034 if we stop investing in juniors in 2026.

The firms making this decision are not stupid. They are responding to incentives. A junior costs roughly £75,000–£120,000 per year fully loaded for the first three years before they generate meaningful value. A licence to a frontier coding assistant costs a few hundred dollars a month. The numbers, on a one-year planning horizon, are not close. Cutting junior hires this year is good for the income statement this year. It is also liquidating the capital asset — the apprenticeship system itself — that produced every senior currently sitting in the firm.

That asset doesn’t appear on the balance sheet. So nobody books the loss when it depreciates. Until, one day, they look around and the bench is empty.

The verification asymmetry: where the bottleneck moves

There’s a second, subtler dynamic worth naming. Even when juniors are hired into AI-augmented teams, what they’re being asked to do has changed shape — and the new shape is brutal.

The classical apprenticeship asked juniors to produce mediocre work and gave them a senior’s feedback to improve. The AI-augmented version asks juniors to evaluate AI-produced work that already looks polished, in volumes that previously would have taken weeks to generate. The skill demanded is verification, not production. And verification is harder than production, not easier.

To know that a contract clause is wrong, you have to have written enough contracts to feel the wrongness. To know that an AI-generated chart is misleading, you have to have built enough charts to recognise the lie. To know that an AI’s code is subtly off, you have to have written enough code to have intuitions about what good looks like. None of this comes free. And asking a 23-year-old to verify the output of a system designed to sound authoritative on every topic — including topics they have never personally touched — is asking them to build a kind of expertise we do not yet know how to teach in the absence of doing.

The result is a quiet but significant shift in where the work piles up. AI generates fluently and confidently. Juniors pass it through with whatever skepticism they can muster. Errors compound. Eventually a senior catches them — but the senior is now spending more time reviewing than they used to spend producing, and they are reviewing things they did not produce themselves. The bottleneck moves up the org chart. Which is, of course, exactly the bottleneck that “AI productivity” was supposed to remove.

This pattern shows up in domain after domain. Senior engineers report spending more time reviewing AI-generated PRs than they ever spent reviewing human ones. Senior consultants describe doing twice the QA work on decks their juniors built with AI. Senior writers find themselves rewriting more, not less. The narrative says AI lets seniors focus on high-leverage work. The reality, in many cases, is that AI lets seniors do verification at scale — which is necessary work, but it is not high-leverage work, and it is not what we were promised.

A taxonomy of responses

Look at any given firm and you will see one of four postures emerging in response to all this. They are not equally good.

• The Liquidator. Cuts junior hiring aggressively, books the savings, claims AI productivity gains, ignores the long-term pipeline question entirely. Common in firms with short executive tenures and quarterly earnings pressure. The 2024–2025 wave of Big Tech layoffs hit early-career engineers disproportionately, and most of those firms have not announced any structural plan to rebuild the pipeline. They are betting either that AI will keep getting better fast enough that no junior pipeline is needed, or — more cynically — that this is the next CEO’s problem.

• The Pretender. Continues to hire juniors at roughly the same rate, but reduces investment in their training because “AI will teach them.” This is the worst posture of the four. The juniors arrive, find no senior willing to spend mentoring time, fail to develop, and leave or are quietly let go after eighteen months. The firm congratulates itself on still hiring, while producing exactly zero new seniors.

• The Restructurer. Recognises the apprenticeship is broken and tries to rebuild it explicitly. Ropes & Gray’s late-2025 “TrAIlblazers” pilot is the most public example: first-year associates are encouraged to spend 20% of their billable target — roughly 400 hours a year — on AI training and experimentation, with those hours counting toward their internal evaluations. It is an honest admission that the old model is dying and that someone has to pay to build the new one. Whether 400 hours of AI training a year produces the kind of judgement that 1,500 hours of document review used to is a separate question. But the firm is at least showing up for the conversation.

• The Inverter. The most interesting and rarest response. A small number of firms — generally smaller, founder-led, long-horizon — are doubling down on junior hiring precisely because they expect a senior shortage in eight to ten years and intend to be the ones who have the people. They treat the senior bench as a strategic asset and the apprenticeship as their moat. If the Stanford data is even directionally correct, these firms will look extraordinary in 2034.

Most firms reading this will recognise themselves in one of the first two. The third is hard. The fourth is rarer still. But the choice is being made — actively, by inaction, every quarter that goes by without a deliberate position.

What gets lost when the apprenticeship goes

There is a temptation, when describing the apprenticeship, to romanticise it. It was often miserable. Junior bankers worked themselves into hospital beds. Junior associates billed eighty-hour weeks doing soul-crushing work. Junior consultants flew home Friday nights only to fly back Monday morning. Some of what is being eliminated is genuinely worth eliminating.

But the apprenticeship was never just labour extraction. It was, at its best, a transmission system. It transmitted technical skill, of course — but also taste, ethics, professional norms, the unwritten rules of how to handle a difficult client, how to push back on a partner, how to know when something is wrong before you can articulate why. It transmitted judgement. None of that comes through in a textbook, and very little of it comes through in a six-week onboarding. It comes through in the proximity of doing real work, watching real seniors handle real situations, and slowly — over years — developing the same intuitions.

The thing we are at risk of losing is not the document review. We can lose document review. We should lose document review. The thing we are at risk of losing is everything that used to come with document review — the proximity, the watching, the slow soaking-up of how this profession actually works in the parts that aren’t written down.

You can replace the labour with AI. You cannot replace the proximity with AI. Or rather, you can try, but the people who emerge from a fully AI-mediated training process will be a different kind of professional than the ones who came before — and we are about to find out, at scale, whether they are good enough.

Practical implications

• For early-career people: stop waiting for an employer to invest in your judgement. They have less incentive than they have ever had. Build the verification skill explicitly. Pick problems where you do the slow, manual version and the AI version, and notice the delta — that delta is where your future taste lives. Find seniors and pay for their time if you have to. Mentorship is now a market good. Treat it like one.

• For hiring managers: stop screening for the skills AI now does well. Screen for taste, judgement under uncertainty, and the willingness to do hard reps. The juniors who will be your seniors in 2032 do not look like the ones who became your seniors in 2022. They are people who can articulate why an AI output is wrong without immediately being able to fix it — that is the verification muscle, and it is the most important hire you can make right now.

• For leaders: your future senior bench is a balance-sheet item that does not appear on your balance sheet. Cutting junior hires this year saves money this year. It also liquidates the apprenticeship that produced every senior you currently depend on. Rebuilding that — formally, with budget, with senior time explicitly allocated to mentorship the way Ropes & Gray has allocated billable hours to AI training — is the most strategic move available to you in 2026. If you do not make it, your competitors will. And in 2034, they will have the only people who can actually run the work.

  ---

The uncomfortable truth

We are running an experiment we have not consented to. We are removing the entry-level rung from every knowledge profession at the same time, on the unspoken assumption that AI will somehow produce its own seniors. It will not. AI gets better at what AI does. Humans get better at judgement by doing the work — including, especially, the work AI now does.

The apprenticeship was never inefficient. It was a transmission system. We are scrapping the transmission and hoping the wheels still turn. They will, for a while. They are turning right now, on the senior expertise we built up before all this started, and that expertise has perhaps a decade of inertia in it. Then it runs out.

The question for the next ten years is not whether AI is taking entry-level jobs. The question is who will be the senior partner, the staff engineer, the principal designer in 2034 — and whether anyone is willing to pay, today, for the slow, unglamorous, economically inefficient work that produces them.

If everyone waits for someone else to train the next generation, no one will.

We will look around in eight years and find the bench empty. We will be very surprised. We should not be.

The workslop economy

In September 2025, Harvard Business Review published a number that should have terrified every knowledge-work organisation in the western economy: 40% of desk workers had received AI-generated “workslop” in the previous month. Content that looked polished but lacked substance. Decks padded to look complete. Memos with the shape of analysis but no spine. Reports where the conclusions had been generated before the evidence.

The average worker spent 3.4 hours per month cleaning it up — triangulating sources, re-running numbers, rewriting sections that only sounded finished. For a 10,000-person company, HBR calculated the cost at $8.1 million per year. Two months later, Merriam-Webster quietly marked the moment by naming “slop” its 2025 word of the year: “low-quality AI-generated content flooding online spaces”.

But the dollar figure understates the problem. The cost isn’t the hours. It’s what those hours required.

To clean up workslop, you need taste.

You need the ability to look at a shiny-looking output and feel — in the prose, in the structure, in the argument — where it has gone wrong. You need a calibrated sense of what a good memo reads like, what a coherent deck argues, what a deliverable that earns trust actually contains.

And here is the trouble: most organisations are operating with less taste than they had five years ago. Not more.

This is the most important thing happening in knowledge work right now, and it barely has a name.

Let’s call it the Taste Gap.

The abundance flip

For every generation of knowledge workers before this one, the scarce resource was production. Could you write this quickly enough, with enough polish, at the required volume? Could you design it, code it, model it, illustrate it? Time, skill, and raw output capacity were the bottleneck.

That bottleneck is gone.

In 2026, a moderately-skilled practitioner with the right tools can generate, in a single afternoon, what used to be a fortnight of work from a well-staffed team. Decks, research briefs, first-draft strategies, landing pages, prototype code, visual identities, internal memos, data summaries — all close to free. Not perfect, but close enough that the variance between “good” and “mediocre” is no longer bridged by doing the work. It’s bridged by knowing what good looks like.

This is what I’ve come to call the abundance flip. When you can generate anything, the only remaining question is what’s worth committing to. And that’s a taste question. Not a production question.

The designer and writer at Designative put the shift crisply:

“Taste is the judgement that operates when options are abundant — when many solutions are technically viable, data-backed, and defensible. It’s what allows teams to discriminate between them, to explain why one direction deserves commitment while others do not.”

For a workforce trained primarily to produce, this is an unexpected pivot. We optimised for output for a century. Taste was the thing you picked up informally — from the partner who’d return your draft covered in red ink, from the VP who’d kill your concept and tell you why, from long hours in review comparing five takes to pick the one. Taste was a side-effect of production. An apprenticeship dividend. It was never the job itself.

Now it’s the job. And most of us are underqualified.

What taste actually is

Before we go further, it’s worth being precise, because “taste” is a word that has absorbed too much mystification.

Taste is not a vibe. It isn’t subjective. It isn’t “knowing what you like.”

Taste a learnt sensitivity to context, audience, and consequence, developed through prolonged exposure, critique, and revision.

Nielsen Norman Group calls it a decision-making skill. Anders Ericsson, the psychologist who founded the field of deliberate practice, would have recognised it as the output of thousands of cycles: attempt, feedback, reflection, refinement. The research on expert performance is clear: experts aren’t born with taste; they’re built through mentored repetition in high-feedback environments.

You can split taste into four working varieties, each at a different stage of decay:

• Contextual taste — knowing what’s right for this audience, this organisation, this moment. The instinct to recognise that a deck that would slay in Amsterdam will die in Tokyo; that the Friday-afternoon email wants a different register from the Monday-morning one.

• Editorial taste — structural judgement. Knowing what to cut, what to emphasise, what to reorder. Feeling when an argument has a hollow middle, or when the second paragraph is doing the work the first should have done.

• Aesthetic taste — sensory judgement. Knowing what reads right, what sounds right, what looks right. Not “pretty” but calibrated. The reason two versions of the same deck provoke different reactions even when the content is identical.

• Strategic taste — discernment about what’s worth doing in the first place. Which problems are actual problems. Which questions are worth asking. Which bets are worth making. This is the highest-stakes form of taste, and the one AI has least access to, because it’s fundamentally a question of what matters, and AI has no stake in what matters.

All four are degrading as we outsource the practice that built them.

The apprenticeship vacuum

Here’s the most uncomfortable part.

Ira Glass — the radio producer — famously articulated what he called “the gap” for creative beginners: people enter a field because their taste is already sophisticated. They can tell good work from bad. Their problem is that their output doesn’t yet match their taste. That’s the gap.

His advice was the only advice that has ever worked: do a huge volume of work. Put yourself on deadlines. Accept the discomfort of producing things you know aren’t yet good. Eventually, your output catches up to your taste.

Two decades later, we are watching an inversion of that problem unfold in real time. AI is doing the production. Beginners don’t have to sit in the gap any more. They don’t have to push through the discomfort. They don’t have to produce ten bad decks in order to internalise, viscerally, what a bad deck is and why.

This sounds like progress. It is a catastrophe for taste formation.

Taste does not form by consumption alone. You don’t get it by reading great work; you get it by trying to make great work, failing, comparing your output to the best of the field, and feeling — physically, uncomfortably — where your attempt fell short. You get it through a ten-year cycle of production-feedback-revision-production. The work itself is the training set.

And the work itself is exactly what we are liquidating:

• The junior analyst who used to spend eighteen months pattern-matching across hundreds of client decks? AI drafts the deck now. She never sees the hundred decks.

• The associate designer who used to generate fifty variations of every logo mark? AI does it in thirty seconds. He never develops a feel for the shape of what works.

• The editorial assistant who used to read two thousand submissions to find forty good ones? AI pre-filters. She never builds the eye.

• The new partner who used to sit in every pitch meeting, absorbing how senior partners chose and cut and defended? Those meetings are now abbreviated or auto-summarised. He never sees the cuts that mattered.

We’ve eliminated the apprenticeship without naming what we’ve eliminated.

The production work was never just production. It was the scaffolding on which taste was built. Remove the scaffolding and you don’t get taste more quickly. You get taste not at all.

Call this the Apprenticeship Vacuum. It is the defining risk of the next decade of knowledge work, and almost no one is managing for it.

The calibration crisis

A second, quieter problem runs parallel to the first: we are losing our sense of what “good” even means.

When every deck looks competent, competence loses its signal. When every email reads polished, polish becomes noise. The reference points that knowledge workers once used to calibrate their own standard — that deck from a senior partner, that memo from the CEO, that essay you remembered a decade later — are drowning in a sea of adequately-produced everything.

This is what the “AI slop” discourse is really about. It’s not that AI output is uniformly terrible. Most of it is mediocre-to-decent. The problem is that mediocre-to-decent is now the ambient baseline. Our sense of “great” is eroding because we can no longer easily find the edge cases that used to anchor it. The peaks look lower because the valleys have risen.

Europol has projected that by the end of 2026, as much as 90% of online content may be synthetically generated. Even if you discount that number significantly, the directional truth holds: we are about to live in a world where most of what we read, see, and evaluate at work was produced by systems with no stake in any of it. Calibration under those conditions is not automatic. It requires effort.

Organisations used to run on implicit calibration. Reviews, edits, critiques — these transmitted, week by week, what the house standard was. When that process is automated or abbreviated — “the AI can redraft it” — the calibration stops happening. Teams drift. Standards don’t fall all at once. They fall one unreviewed deliverable at a time, for years, until one day a senior leader opens a deck and doesn’t understand why it feels so hollow, even though every box is ticked. By then, the people who would have told them why are five years gone.

A counter-argument, honestly considered

“Every new tool triggered this panic,” the sensible person says. “Photography was supposed to kill painting. Calculators were supposed to kill arithmetic. Spell-check was supposed to kill spelling. None of it happened. People adapted. Taste migrated. Why should this be different?”

It’s a fair challenge and worth answering directly.

The honest answer is: the earlier tools removed discrete, bounded capacities. A calculator does long division. A spell-check checks a word. Each replaced one small layer of cognitive work, leaving the surrounding judgement largely intact. You still had to decide which equation to set up, which sentence to write, which argument to make.

Generative AI is different in kind, not degree. It removes the whole surface between initial intent and finished artifact — including most of the middle-skill judgement calls where taste is forged. A junior designer using Photoshop in 2010 made hundreds of micro-choices per day: font weights, kerning, colour relationships, negative space, hierarchy. A junior “designer” using a generative tool in 2026 may make a handful of prompt-level choices and pick from four options. The volume of calibration reps per day has collapsed by an order of magnitude — and it’s the reps, not the output, that built the designer.

That is what makes this particular substitution dangerous in a way that calculators never were. We are not removing a tool. We are removing a gym.

The discernment dividend

There is, however, a bright side hidden inside this — and the organisations that find it first will own the next decade.

Taste is getting scarcer, and scarcity prices value. The Discernment Dividend is the compounding economic premium accruing to people and organisations with calibrated judgement in a world where everyone else can produce but fewer can discriminate.

Signs of it are already visible. Editors are being paid more, not less, in AI-saturated publishing. Senior designers command higher multiples over juniors than they did in 2022. “Curator” roles — people whose sole job is to choose and defend — are appearing in product, publishing, and learning organisations. The creator economy is quietly bifurcating between high-volume generators (low margin, low defensibility) and taste-driven brands (high margin, fiercely defensible).

This is the Discernment Dividend starting to show up in pay packets. It will accelerate.

Practical implications

• For early career: your production ability no longer differentiates you. Your taste does. Treat taste-building as the core of your first decade, not a by-product of it.

Consume excellent work constantly — not passively, but analytically. Why is this piece good? What decisions did the writer make? Where would a worse version have drifted? Keep a private file of work that moved you, and revisit it. Make notes on what specifically landed.

Seek critics. Find the person in your organisation whose taste you most respect and ask them to shred one piece of your work every month. Do the work AI can’t yet: original hypotheses, unexpected framings, critique that takes a risk.

And do some work by hand, sometimes. You will be slower. You will be right less often. You will learn what it feels like to struggle with a problem — which is the only way taste gets installed.

• For mid-career: you are at the most dangerous inflection of your career. Your taste is partially built. Your role is being restructured to lean harder on AI. You will be tempted to coast on the taste you already have while AI handles the execution.

Don’t.

Taste is a muscle. It atrophies. The professionals who will matter in 2035 are not the ones who optimised for AI-assisted output in 2026. They are the ones who kept showing up to the work where taste is tested — live critiques, genuine disagreements, decisions under real stakes. Resist the drift toward being a “reviewer of AI drafts.” You will degrade into it if you’re not careful.

• For hiring: stop screening for production skills. Everyone’s writing samples look good now. Everyone’s portfolio is polished. Screen for discernment. Show candidates three pieces of AI-generated work and ask them to rank and defend. Present a flawed strategy and ask what they’d cut and why. The person who can articulate why one version is better — and can do it in a way that changes how you see the work — is worth five who cannot.

Interview for critique, not composition.

• For leaders: you are running a taste-development programme whether you named it that or not. Every review is a training signal. Every “ship it” teaches your team what good means to you. If you outsource your reviews to AI summaries, you have stopped teaching taste in your organisation. Full stop.

Consider actively protecting apprenticeship work. Keep some decks hand-drafted. Keep some critiques human. Make exposure to your best people’s reasoning a formal benefit of working at your company, not an accident. The companies that do this will quietly collect the strongest talent — because good people want to get better, and they can only get better somewhere that still teaches taste.

• For organisations: audit your AI investment. For every dollar you spend on production tools, how much are you spending on taste development — on critiques, on reviews, on exposure to excellent work, on the meetings and moments that transmit standards? If the ratio is 100:1 in favour of production, you are over-indexed on the thing that has become commodity and under-indexed on the thing that has become moat.

Name “taste” as a strategic capability. Measure it — not with vanity metrics, but with what your best reviewers say about the quality of the work shipping across the org, month over month. Appoint senior people to its cultivation. Build it into hiring, promotion, and performance review. The same rigor you bring to AI adoption, bring to discernment cultivation.

And consider protecting the humble, unglamorous rituals that actually build taste: the weekly deck review where someone says “this section is wrong and here’s why”; the portfolio critique; the editor who line-edits a draft in front of its author; the post-mortem where “what did we almost ship?” is asked as seriously as “what did we ship?” These rituals look like overhead on an efficiency dashboard. They are the only reason your organisation will have taste ten years from now.

Most organisations in 2026 are investing heavily in AI tools to increase production. Almost none are investing, deliberately and at scale, in taste.

That is exactly backwards.

Production is the new commodity. Taste is the new moat.

And unlike AI capability — which compounds in weeks — taste compounds slowly, across years of deliberate practice in environments that reward judgement. By the time you realise you need it, it’s a decade too late to build.

We are living through a once-in-a-generation inversion of what’s scarce. The organisations that recognise it will get quieter about productivity gains and louder about standards. They’ll pay more for discernment than for output. They’ll protect apprenticeship even when it looks inefficient. They’ll treat every senior-junior review as strategically important, because it is.

The organisations that miss it will generate more than ever and land less. They’ll wonder why their output feels hollow, why their best people keep leaving, why the work doesn’t cut through anymore. They’ll blame the market, the economy, the competition.

The real answer will be simpler and harder.

They lost their taste. And they did it in a way that felt, every single quarter, like they were winning — more output, more decks, more campaigns, more content shipped per headcount than ever before. Which is exactly why almost no one will notice until the damage is too compounded to reverse.

The window to act is short.

Taste that’s already built can still be deepened. Taste that isn’t yet built can still — for another few years — be installed through apprenticeship, if we choose to protect it.

Past that, we are rearing a generation of knowledge workers who have never once had to stare at a bad draft of their own work and feel what it meant. And no amount of AI will teach them what we decided, through efficiency, to stop teaching ourselves.