On the first Monday after an AI agent goes wrong, there is always a meeting. It is well attended, well intentioned and almost entirely devoted to a single question that should have been settled months earlier. Whose agent was it?

Operations says it belongs to the product team. The product team says it is a finance workflow. Finance points at legal, because legal signed the contract. Legal points at the vendor, because the vendor wrote the logic. The vendor, when it joins the call, is warm, apologetic and quietly precise: the contract assigns responsibility to the customer’s designated owner. Somebody opens the register of agents. The designated owner is a name that nobody in the room recognises, because the person left the company in the spring.

This scene is not a joke, and it is no longer hypothetical. In September, IBM and Censuswide surveyed 3,221 senior executives across eight EMEA markets. 80% said an AI-related incident had reached their board or C-suite in the past year. IBM defined an incident broadly: an agent failure, a data exposure, an AI-driven decision error or an AI security breach. Four in five senior leaders have already sat through some version of that Monday meeting.

Then the survey asked the question that matters. Who owns the outcome when an AI agent goes wrong? The most popular answer was the CIO or CTO. It attracted 17%. There was no consensus on legal liability either: 20% favoured sharing it between deployer and vendor, 19% said a regulatory framework that does not yet adequately exist should decide, and 14% placed it with the vendor.

Read those numbers as a single sentence. Most large organisations have had an AI incident, and most cannot say who is responsible for the next one. They have built the machine, switched it on and forgotten to write a name on the door.

The question nobody put on the project plan

Every agent deployment has a business case, a vendor, a security review, a pilot and a launch date. Almost none has a person whose job it is to answer for what the agent does once it is live. It is the one line missing from an otherwise meticulous plan, and it is missing for a reason: asking for it slows the launch.

IBM’s survey makes the incentive visible. When governance lags behind deployment, 56% of executives keep deploying within agreed risk limits. Only 33% pause until governance catches up. Speed is rewarded and the delay is blamed on whoever caused it. The person who points out that nobody owns the agent becomes the person who is holding up the launch.

So the box on the org chart stays empty, and the agents multiply. 35% of those executives already run agentic AI in production. 75% expect agents to have some level of autonomy over decisions by 2027. Meanwhile only 37% have a formal, board-reviewed estimate of the financial risk from their governance gaps. A board cannot govern a risk it has never priced.

This edition is about that empty box. Who ought to be in it, why nobody wants to be, and what happens to the people who end up there by accident.

The long history of someone to blame

The idea that an organisation answers for the actions of those working on its behalf is old. Employment law has long held that an employer is responsible for what an employee does in the course of the job. Ships have captains, aircraft have a pilot in command, and hospitals have a named consultant on every patient record. These are not just administrative conventions. They are answers to the question that every complex system eventually has to face: when something breaks, where does the account stop?

Automation has repeatedly strained that arrangement. When a machine takes over part of the work, responsibility does not move with it. It stays with the human who is nominally in charge, even as their real control shrinks. The pilot is still responsible for an aircraft that spends most of its flight on autopilot. The operator is still responsible for a plant whose control room has been automated. The more capable the system becomes, the less the person responsible can actually do, and the more the arrangement relies on that person being a convenient place to put the blame.

The researcher Madeleine Clare Elish gave this pattern a name. In a 2019 paper in Engaging Science, Technology, and Society, she described the moral crumple zone: a situation where responsibility for an automated system’s actions is misattributed to a human actor who had limited control over it. The metaphor comes from car design. A crumple zone is the part of a vehicle engineered to absorb the force of a crash so that the passenger compartment survives. In a moral crumple zone, the human absorbs the impact so that the system, and the institution behind it, keeps its integrity.

The AI debate has moved from adoption to governance and accountability.

Ana Paula Assis, Chair of IBM EMEA and APAC, on the findings of the company’s September 2026 executive survey.

Assis is right, and the shift matters more than the sentence suggests. For three years, the argument inside organisations was about whether to adopt. Governance was a compliance exercise, something to be written down and filed. Agents change that, because an agent acts. It sends the email, approves the refund, changes the record, calls the API. A tool you consult leaves the decision with you. An agent takes the decision and tells you afterwards, if it tells you at all.

The mechanism: how an owner becomes a scapegoat

Nobody sets out to build a moral crumple zone. They emerge from three ordinary decisions, each reasonable on its own.

The first is that deployment is decentralised. A marketing team connects an agent to the customer database. A finance analyst gives an assistant access to the invoicing system. A developer gives a coding agent standing credentials. None of these decisions goes through a central approval, because the tools are cheap, the benefits are immediate and nobody wants to wait six weeks for a review.

Workato and Censuswide surveyed 1,001 US IT decision-makers at VP level or above in May. 62% agreed that most of the AI agents in use at their organisation were not formally approved by IT. 63% said their IT team knows less about AI agents than other employees do. These are the people responsible for the technology estate, telling us that the estate has grown without them.

The second decision is that accountability follows the org chart, not the architecture. When something breaks, the organisation asks whose function it falls under. The answer is nearly always technology, because technology is the function that is nominally responsible for systems. The same survey found that 74% of those leaders expect IT leadership to be blamed if an agent breaks policy, exposes data or harms a customer. That is more than twice the share who expect the blame to fall on the AI vendor.

Put the two findings together. IT did not approve most of the agents. IT knows less about them than the people using them. And IT expects to take the blame. That is the moral crumple zone in a single page of survey data: accountability without approval rights, blame without a veto.

The third decision is that the contract assigns the risk downhill. Vendors, understandably, limit their liability. The customer’s designated owner carries the operational risk, and the customer carries the legal risk. By the time an incident occurs, the chain of responsibility runs from a model provider through a platform vendor through an integration partner to one nervous person in a company who clicked “accept” on the terms.

The consequences for that person are not abstract. 23% of the same IT leaders expect to be removed or to resign after a major AI-caused incident. 36% expect their organisation to try to handle a major incident quietly, without disclosure. A role with that profile does not attract volunteers. It attracts people who did not see it coming, and people with no choice.

What gets lost when nobody owns the agent

An empty owner box does not just create a blame problem after an incident. It quietly damages the organisation before one.

Early warning disappears. An owned system has someone who reads its logs, notices the strange behaviour and raises a hand. An unowned system produces the same warnings and delivers them to nobody. The most unsettling detail in many incident reviews is not that the signals were missing. It is that they were there, in an inbox no one was reading.

Learning does not happen. OneTrust and Sapio Research surveyed 1,200 senior decision-makers in eight markets in June and July. 86% had experienced at least one AI-related incident in the past year. Yet only 5% said coordination and accountability are clear across the AI lifecycle. An incident that belongs to no one cannot be investigated properly, because there is nobody whose job it is to change the system afterwards. The organisation experiences a crisis, issues a statement and returns to the previous arrangement.

Authority drifts away from responsibility. Where an owner exists on paper but cannot approve, change or stop the agent, the arrangement is ceremonial. The name on the register is a name to blame, not a person who can act. This is the gap between a decision-maker and a designated owner, and it is the gap the moral crumple zone lives in.

Responsibility cannot be outsourced to the machine. In 2024, a tribunal in British Columbia dealt with a customer who had been given wrong information about a bereavement fare by Air Canada’s website chatbot. The airline argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal did not accept it, ruling that the airline remained responsible for all the information on its website, whether it came from a static page or a chatbot. The decision concerned a small claim, but the principle is not small. A company cannot create an agent and then disown it when it speaks.

Five archetypes of the empty box

Organisations fill the owner role in recognisable ways. Most are wrong, and each fails differently.

The Named Scapegoat. A single person, usually in IT or security, whose name is on the register. They did not choose the agent, cannot veto it and cannot always see it. Their title implies authority; their actual power is the ability to be fired. The Named Scapegoat is the human crumple zone in its purest form, and the organisation treats the arrangement as a successful governance structure until the first incident.

The Committee Ghost. A governance committee that meets monthly, reviews a slide deck and issues guidance. Everyone is responsible, so no one is. The committee can recommend, but it cannot stop a deployment on a Tuesday afternoon, and it has no one to dismiss when the guidance is ignored. When an incident occurs, the minutes show that it was discussed. They do not show who acted.

The Vendor Shield. The organisation assumes the AI supplier carries the risk, and sometimes writes it into the strategy deck. The contract usually says otherwise. The 74% figure from Workato is a useful corrective: IT leaders expect to be blamed at more than twice the rate they expect the vendor to be. Even those who buy from a vendor know the buck does not stop there.

The Absentee Signatory. The person who signed off on the agent months ago and has since moved teams, left the company or forgotten. The register still carries their name, and so do the credentials the agent runs under. The agent continues to operate with the authority of someone who is no longer there. It is the most dangerous archetype precisely because it looks fine on paper.

The Quiet Owner. The rare good case. A named person with authority over scope, access and shutdown, who reads the logs, holds a review date and has a direct line to the executive who will answer for the loss. The Quiet Owner is boring in the best sense: no one notices them until the agent misbehaves, at which point the incident is contained in hours rather than weeks. They are also the hardest to hire, because the role requires technical judgement, organisational authority and a willingness to say no to a launch date.

The uncomfortable truths

1. Ownership without authority is not ownership. It is a liability transfer. If the person named on the register cannot approve, change or stop the agent, the register is a list of people to blame. The test is simple: ask the owner whether they could switch the agent off today without asking permission. If the answer is no, you have a scapegoat, not an owner.

2. The pressure to adopt is not neutral. It is created by executives, repeated by consultants and enforced by peers. Only 37% of EMEA organisations have a board-reviewed estimate of the financial risk, yet 87% of the leaders OneTrust surveyed encourage agent use. A culture that encourages adoption without pricing the risk is not brave. It is borrowing against someone else’s career.

3. Regulators are coming for the name on the door. On 1 October, Senators Josh Hawley and Chris Murphy introduced the AI Agent Accountability Act. As described, it would create criminal and civil liability under the Computer Fraud and Abuse Act for operators who knowingly run an agent that recklessly causes hacking damage or loss, and for developers who fail to put reasonable safeguards in place. It is a proposal, not law, and at the time of writing its text had not been published, so definitions such as “operator” may change. But the direction of travel is unmistakable. The people who run agents are about to be treated as the people who answer for them.

4. “Nobody told me” will not be a defence. The 36% of IT leaders who expect their company to handle a major incident quietly are describing a world in which the first sign of an incident is concealment. Concealment is not a governance model. It is a delay before a larger disclosure, and it exposes the owner twice: once for the incident and once for the silence.

5. An agent without an owner is not autonomous. It is abandoned. We use the word autonomy as if it were a feature. Autonomy is what we grant to someone who can be held to account. A system that nobody can be held to account for is not autonomous. It is an orphan with a root password.

What to do about it

The fix is not complicated. It is just unglamorous, which is why it has not happened. Here is what it looks like by role.

If you are early in your career. You may be offered the agent. A team will need someone to “look after” a new workflow, and the junior person with the energy and the technical curiosity is the obvious candidate. Before you accept, ask three questions. What can this agent do, in plain language? What can I do to stop it, today, without asking anyone? And who do I report to if it goes wrong? If you cannot get clear answers, you are being offered a crumple zone with a nice title. Decline politely, or ask for the authority first. The people who can answer these questions clearly will be among the most valuable in any organisation over the next five years, because they are the ones who make agents safe to use.

If you are a manager building a team. Treat agent ownership as a real role, not an extra line on an existing job description. It needs time in the week, access to logs, the power to pause a deployment and a route to the executive who will answer to the board. When you hire, look for people who can describe how an agent fails, not only what it can do. Ask candidates for a time they stopped something that was working because they did not trust it. And protect the person who raises the incident. If the first thing that happens to them is a difficult conversation with their line manager, you will not hear about the second incident until it is on the front page.

If you run a function or a business. Start with a register of every agent in use. Each entry needs a named owner, a written scope, an access list, a kill switch, a review date and a designated deputy for when the owner is absent or moves on. Review the register against your joiners, movers and leavers list every quarter, because the Absentee Signatory is created by departures, not by decisions. Then give the owner the authority that matches the name. In practice, that means the right to approve, to change and to stop. If they cannot do all three, someone senior has not delegated, they have deflected.

If you sit on a board. Ask for a number. IBM found that only 37% of organisations have a board-reviewed quantification of financial risk from AI governance gaps. If your board has never seen a loss estimate for an agent failure, it is not governing the risk, it is hoping. Ask who owns the most consequential agent in the organisation, what they can stop and whether they would tell you about an incident. Then ask the question 98% of the IBM sample say they have answered: who holds board-level responsibility for AI risk? Naming a director is easy. Asking what that director has actually reviewed is the part that tests the arrangement.

The name on the door

It would be easy to read this edition as an argument for slowing down. It is not. The Workato leaders were explicit that none of them wanted to slow the business. What they wanted was to know what the agents were doing, to be able to audit and reverse their actions, and to know who is accountable. 29% named that last point as a top priority. They are not asking for permission to stop. They are asking for a name on the door.

The irony is that ownership is what makes speed possible. A pilot can fly faster in a system with a clear captain, a checklist and a flight recorder. An organisation can deploy more agents, more boldly, once it knows who is watching each one, what that person can do and where they go when things break. Trust in automation has never come from the absence of human responsibility. It has come from the presence of it.

So the next time someone proposes a new agent, do not begin with the business case. Begin with a short, awkward conversation. Whose name goes beside this? What can they switch off? What will they tell the board when it fails? If the room goes quiet, you have found the work.

Because the Monday meeting is coming. The only choice you still have is whether, when it starts, someone in the room already knows the answer to the question on the first slide.

In April 2023, an engineer at Samsung pasted sensitive internal source code into ChatGPT. A few weeks later, staff in one of the company’s biggest divisions received a memo. Generative AI was now banned on company phones, computers and tablets, and over the internal network. Staff were also asked not to share company information with such tools on their personal devices, and the memo warned that breaking the rule could lead to disciplinary action up to and including dismissal.

It was, by every corporate standard, a responsible decision. A leak had happened. A risk had been named. A rule had been written. Samsung was not alone: within months, a string of banks and large companies had banned or restricted ChatGPT for similar reasons, and “we have blocked it” became a respectable answer in boardrooms everywhere.

Three years later, we can measure what those memos achieved. The answer is uncomfortable for anyone who signed one. Blanket bans did not stop people using AI. They stopped leaders seeing it. They converted a visible, governable behaviour into an invisible, ungovernable one, and they did it while giving the people at the top a pleasant feeling of control.

This edition is about that feeling, and about what it costs. My argument is simple: banning AI is not an AI policy. It is the absence of one, dressed up as decisiveness.

We have run this experiment before

On 17 January 1920, the United States made the manufacture, transport and sale of intoxicating liquor illegal. The intention was moral, the law was national and the enforcement was real. Agents raided warehouses, smashed barrels and poured whisky into gutters for the newspapers.

Americans kept drinking. What changed was where. In New York alone, contemporary estimates of the number of speakeasies ran from 30,000 to 100,000. Drinking moved from licensed saloons, where the owner had a name, an address and something to lose, into back rooms, basements and unmarked doors. The alcohol got worse, because nobody could inspect it. The economics got darker, because the only people willing to supply an illegal market were people already comfortable with illegality. By 1929, the city’s own police commissioner, Grover Whalen, was warning Mayor Jimmy Walker about who the trade had pulled in.

“The illegal nature of the business attracts to it the stickup thug, the gangster, the degenerate, the racketeer and the gunman.”

Grover Whalen, New York Police Commissioner, report to the mayor, 1929 (via the NYC Municipal Archives)

The point is not that drinking was harmless. The point is that prohibition did not reduce the harm. It removed the supervision. It took a behaviour that society could see, regulate and tax, and moved it somewhere nobody could inspect the bottles.

The workplace has its own, smaller versions of this history. In the late 2000s, plenty of IT departments refused to let staff connect personal smartphones to company email. Staff forwarded their work mail to personal accounts instead. Eventually the industry gave up and invented a name for the thing it could not stop: bring your own device. The policy that followed was not a ban. It was a set of rules for managing what was already happening.

Every one of these stories has the same shape. A new tool arrives that makes individual work easier. The institution sees the risk before it sees the value. It bans the tool. The tool survives the ban, and the institution loses the one thing it actually needed, which was the ability to see.

The mechanism: demand does not obey memos

To understand why AI bans fail, start with the demand, because the demand is not a fringe phenomenon.

Microsoft and LinkedIn’s 2024 Work Trend Index, based on a survey of 31,000 people in 31 countries, found that 75% of knowledge workers were already using AI at work. More importantly, 78% of those AI users were bringing their own tools, rather than waiting for their employer to provide them. Microsoft gave it a name, bring your own AI, and the name is telling. It describes something done to the organisation, not by it.

Then Software AG asked the question that every banning executive should have asked first. Its survey of 6,000 knowledge workers across the US, UK and Germany found that 46% would keep using their personal AI tools even if their organisation banned them completely. In Germany, the figure was 49%. Those workers cited time savings, easier jobs and higher productivity as their reasons. In other words, they were not rebels. They were people trying to do their jobs well, who had found a tool that helped.

The largest study of all makes the pattern unmistakable. The University of Melbourne and KPMG surveyed more than 48,000 people across 47 countries for their 2025 global study of trust in AI. 57% of employees said they hide their use of AI and present AI-generated work as their own. Almost half admitted to using AI in ways that contravene company policy, including uploading sensitive company information into free public tools.

And the trend has not reversed. In June 2026, PagerDuty published a survey by Wakefield Research of 1,250 office professionals at companies with at least $500 million in revenue. Two thirds said they had used AI tools at work despite believing they were not permitted under company policy. At organisations with 1,500 or more employees, the figure rose to 72%.

Put those findings together and the mechanism is clear. A ban does not change the incentive. The employee still has a report due on Thursday, a backlog of emails and a manager who expects faster turnaround than last year. AI still makes those things easier. The only thing the ban changes is the route the work takes. It goes from the company laptop to the personal phone, from the enterprise account to the free one, from the logged network to the home broadband.

I call this the speakeasy effect: when you prohibit a tool that people find genuinely useful, you do not remove the use, you remove the inspection. The behaviour continues, but every safeguard you could have applied to it disappears.

What the ban actually costs

For a while, the speakeasy effect was a theory. It now has a price tag.

IBM’s 2025 Cost of a Data Breach report was the first edition to look hard at shadow AI, meaning AI tools used without the knowledge or approval of the organisation. It found that one in five breaches in its study involved shadow AI. Organisations with high levels of shadow AI paid an average of $670,000 more per breach than those with low levels or none. Those breaches were also more likely to expose personal data and intellectual property. And among organisations that reported an AI-related breach, 97% lacked proper AI access controls.

Read those numbers alongside the Samsung memo. The original fear was that staff would paste sensitive data into a public tool. The ban’s answer was to forbid the tool. The result, across the economy, is that staff paste sensitive data into public tools on devices the company cannot monitor, and the company finds out during the breach investigation.

But the financial cost is the easiest one to see. The deeper costs are the ones that never appear in a report, because they are, by definition, invisible. I count three.

First, you lose the learning. When AI use is visible, an organisation can discover which uses work. A sales team finds a prompt that halves proposal time. A finance analyst works out which reconciliation tasks the model gets wrong. Those discoveries can be shared, tested and turned into standard practice. When AI use is hidden, every discovery stays private. Every employee runs their own experiment, alone, and the organisation learns nothing from any of them. The KPMG study found that only 47% of employees had received any AI training. In a banning culture, the other half are teaching themselves in secret.

Second, you lose the error signal. The same study found that 66% of employees rely on AI output without evaluating its accuracy, and 56% say they have made mistakes in their work because of AI. In an open culture, those mistakes surface. Someone says, “the model invented that clause”, and the team adds a check. In a hidden culture, nobody admits the source of the error, because admitting the source means admitting the breach of policy. The mistake gets fixed, perhaps, but the cause stays buried, ready to recur.

Third, you lose honesty itself. This is the cost I worry about most. A ban that most people break teaches everyone that the official rules and the real rules are different things. It trains people to present AI-assisted work as unassisted work. It turns ordinary productivity into a small, daily act of concealment. Once a workforce has learned that the safest move is to say nothing, that lesson does not stay neatly confined to AI.

The four ban archetypes

Not every ban looks the same. In the conversations I have had with leaders over the past year, I see four recurring patterns. Each one feels like governance. None of them is.

• The Panic Ban. Something goes wrong, usually a leak or an embarrassing output, and the response arrives within days. The Panic Ban is the purest form of the speakeasy effect, because it is written in fear and enforced with threats. It names the tool, not the risk. It is often the right short-term call and the wrong long-term policy, and the trouble is that nobody ever comes back to replace it.

• The Paper Ban. The policy exists, somewhere, in a document nobody reads. It says AI tools require approval. There is no approval process. There are no approved tools. Everyone knows the rule exists and everyone knows nobody enforces it. The Paper Ban is the most common archetype and the most corrosive, because it gives legal cover to leaders while giving permission to nobody. The KPMG study found that only 40% of employees said their workplace had a policy or guidance on generative AI use. I suspect many of the rest work under a Paper Ban they have never seen.

• The Selective Ban. AI is banned for most staff, but quietly allowed for a favoured few: the leadership team, the innovation lab, the consultants. The Selective Ban teaches the rest of the organisation that AI is a privilege of status rather than a tool of the job. Survey work this year suggests that decision-makers are often more likely to use unsanctioned AI than the staff below them, which makes the Selective Ban not just unfair but faintly comic.

• The Perpetual Pilot. This one does not call itself a ban. It calls itself a pilot. A small group gets access to a sanctioned tool “while we evaluate”. The evaluation never ends. Eighteen months later, the pilot group is still the only group with access, and everyone else has long since found their own tools. The Perpetual Pilot is a ban with a better press office.

What unites all four is that they answer the wrong question. They ask, “should our people use AI?” That question was settled by the people themselves, years ago, without asking permission. The only live question is, “how should they use it, and how will we know?”

The uncomfortable truths

A ban is the easiest policy a leader can write. It requires no understanding of the work, no budget for tools and no difficult conversations about which data matters. It produces a clean memo, a clear audit trail and a clear conscience. That comfort is precisely why it is so popular, and precisely why it should be treated with suspicion.

Your real adoption rate is higher than your licence count. If you have bought seats for 10% of your workforce and banned everything else, you do not have 10% adoption. You have 10% visible adoption and an unknown amount of invisible adoption. On the evidence above, the invisible part is probably the larger one.

The people breaking your ban are often your best people. The workers most likely to route around a ban are the ones under the most pressure to deliver and the most curious about better methods. A ban does not punish your least diligent staff. It punishes the ones trying hardest, or teaches them to hide.

Fear is not a control. Security teams know this better than anyone. A control is something you can test, monitor and audit. A threat of dismissal is none of those things. It only works on people who expect to be caught, and the whole point of shadow AI is that nobody expects to be caught.

The risk you can see is always smaller than the risk you cannot. A sanctioned tool with data controls, logging and a clear owner carries risk. An unsanctioned tool on a personal phone carries the same risk with none of those things. The choice is never between risk and no risk. It is between risk you can manage and risk you cannot even count.

What to do instead

Replacing a ban does not mean replacing it with a free-for-all. It means replacing a single prohibition with a small number of clear answers. A real AI policy answers four questions: which tools are sanctioned and why they beat the free alternative, which data may never leave the building, who owns and checks the output before it reaches a customer, and how people disclose their use without being punished for honesty. Here is what that means depending on where you sit.

If you are early in your career:

• Do not hide your AI use. Hidden use is invisible skill, and invisible skill does not get promoted. The ability to use AI well, check its output and explain how you got to an answer is one of the most valuable things you can demonstrate right now, but only if someone can see it.

• Learn your organisation’s data boundaries before you learn its prompts. Know which information is confidential, which is personal and which is regulated. The fastest way to end an early career is to be the person whose upload appears in the breach report.

• If there is no sanctioned tool, ask for one, in writing, with a specific use case. You become the person who surfaced the problem rather than the person caught causing it. That is a much better reputation to have.

If you are hiring:

• Ask candidates how they use AI, and reward honest answers. If your interview treats every mention of AI as cheating, you are selecting for people who conceal, and you will get a workforce of them.

• Test judgement about data, not just fluency with tools. Give candidates a scenario: a deadline, a confidential document, a public chatbot. What would they do? The answer tells you more than any prompt-writing exercise.

• Look for people who can describe a time AI got something wrong and how they caught it. That is the skill your policy depends on.

If you are leading:

• Count before you rule. Before you write or rewrite any AI policy, find out what is already happening. Run an anonymous survey. Look at network traffic. Ask teams directly, and promise no consequences for the answer. You cannot govern a behaviour you have refused to measure.

• Make the sanctioned path easier than the shadow one. People route around friction, not principles. If the approved tool requires a three-week access request and the free one requires a phone, you have already lost. Speed of access is a security control.

• Write rules about data, not about tools. Tools change every quarter. The categories of information you must protect change far more slowly. A policy that says “never paste customer personal data into any tool that is not on this list” will outlive any list of banned websites.

• Create an amnesty. Tell people that disclosing their current AI use will not be punished, and mean it. The information you gain about what your workforce actually needs is worth far more than the satisfaction of enforcing a rule nobody followed.

• Replace “never” with “here is how”. Every blanket ban is a training budget you chose not to spend. Every hour you spend teaching people to use AI safely is an hour they do not spend teaching themselves in secret.

The memo you should have written

Go back to that Samsung memo in 2023. It was a reasonable reaction to a real incident, and I do not think the people who wrote it were foolish. They did what responsible institutions have always done when a new tool arrives faster than their understanding of it. They reached for the lever marked “stop”.

The trouble is that the lever is not connected to anything. Prohibition did not stop New York drinking. Blocking smartphones did not stop people reading work email in bed. Banning AI has not stopped people using AI. Every time, the lever only did one thing: it moved the behaviour to a place where the institution could no longer see it, shape it or learn from it.

The memo that would actually have reduced the risk would have been longer and harder to write. It would have said: here are the tools you may use, here is the data you must never share, here is who checks the output and here is how to tell us what you are doing without fear. It would have required investment, judgement and honesty about how work really gets done.

That memo is still available to any leader willing to write it. The alternative is to keep the ban, keep the clean conscience and keep the comfortable silence that comes with it.

Just remember what that silence means. You did not stop AI. You stopped hearing about it.

The account nobody checks

Imagine a bank that publishes a glowing dividend report every quarter. Payouts up. Returns up. Customers delighted. Then imagine that the bank has stopped publishing its balance sheet, and no regulator, board member or journalist has thought to ask why. The dividends are real money, deposited in real accounts. What nobody can tell you is what they are being paid out of.

That is where most organisations stand with AI today. Every adoption dashboard in every leadership deck shows the same pleasing shape: usage climbing, throughput climbing, time-to-draft collapsing. All of it is true. And all of it measures the yield. Almost none of it measures the principal: the accumulated human capability, built over years of difficult, slightly painful practice, that the yield is quietly being drawn against.

This edition is about that second number. Call the first one the Atrophy Dividend: the short-term performance gain AI pays out today, funded by a long-term skill withdrawal that no instrument in your organisation is built to detect. The unnerving part is not that the withdrawal happens. The unnerving part is that, by design, the people making it feel richer while they do.

A fifth of an edge, gone

In August 2025, The Lancet Gastroenterology & Hepatology published a study that deserves to be taught in every business school, although it is about colonoscopies. Researchers observed 19 experienced endoscopists, each with more than 2,000 procedures behind them, working across four centres in Poland. Before AI-assisted detection was introduced, these doctors found adenomas, the precancerous growths they are paid to find, in 28.4% of the colonoscopies they performed unaided. After a period of routine AI use, the researchers looked at the same doctors performing procedures without the AI. Their detection rate had fallen to 22.4%. That is a six-point absolute drop and roughly a 20% relative decline in a life-and-death skill, in people who were already experts.

Nobody did anything wrong. The AI worked: with the system switched on, detection came in at 25.3%. These were not careless clinicians. They were excellent clinicians who had, without deciding to, started leaning on a very good crutch, and a body of expertise that had taken years to build began to dim in a matter of months. The study’s lead clinician, Dr Marcin Romańczyk, described it as the first to suggest a negative impact of regular AI use on health professionals’ ability to complete a patient-relevant task. An accompanying commentary warned of “the quiet erosion of fundamental skills.”

“You are gaining something in the short term that might result in losses in the long term.”

Mohammad Hossein Jarrahi, University of North Carolina at Chapel Hill, on the cognitive debt of delegating decisions to AI (as reported by HCA Magazine)

Jarrahi’s phrase, cognitive debt, is the right one, with one amendment. A debt implies a statement: a number you can look at, a repayment schedule, a creditor who will eventually call. Skill atrophy has none of these. It is a debt with no ledger, which is exactly why it compounds.

We have seen this film before

None of this is new. It is simply arriving in a new industry. In 1983, the psychologist Lisanne Bainbridge published a short paper in the journal Automatica titled “Ironies of Automation.” Her argument was devastating in its economy. Automate the routine parts of a task, she wrote, and you leave the human with the exceptions: the rare, difficult, unrehearsed cases the machine cannot handle. But the human can no longer practise on the routine cases, because the machine has absorbed them. So the person you most need at the worst moment is also the person who has had the least chance to prepare for it. The better the automation, the crueller the irony.

Aviation learned this by crashing. After a series of accidents in which crews struggled to fly manually when the automation failed or confused them, the US Federal Aviation Administration issued a safety alert in January 2013 encouraging airlines to promote manual flight operations and to give pilots more hands-on time. Pilots now log deliberate manual flying not because autopilots are bad, but because autopilots are so good that human proficiency decays without being protected on purpose.

Notice what aviation did that most workplaces have not. It admitted the decay was structural. It did not blame individual pilots for getting rusty, and it did not ask them to “be more careful.” It built practice into the system. The industry treated skill the way a bank treats capital: as something that must be actively held in reserve, not assumed to be there.

Knowledge work has nothing like it. We handed our drafting, analysing, summarising and coding to a tool with extraordinary range, and we did it faster than any profession has ever adopted anything. We have no equivalent of manual flight hours. We barely have the vocabulary.

The mechanism: supervising is not practising

Why does the decay happen so reliably? Because expertise is not a possession. It is not a file you saved once and can open again. It is a pattern of connections that is maintained by use, and specifically by retrieval and production under difficulty: the moment you stare at the blank page, the awkward dataset, the bug that won’t yield, and have to generate something from what you know.

AI removes precisely that moment. And what it leaves behind looks like work, feels like work, and is billed as work, but is a different cognitive activity. The AI produces; the human reviews. Review is recognition, and recognition is the easy half of expertise. You can approve a thousand well-formed analyses and become steadily worse at writing one, because nothing in the approving requires you to reach.

Three pieces of evidence show how this plays out.

Evidence one: the learning penalty. In January 2026, Judy Hanwen Shen and Alex Tamkin at Anthropic published a randomised trial with 52 mostly junior software engineers, each with at least a year of weekly Python experience, asked to learn an unfamiliar library. Half used an AI assistant; half did not. On a comprehension quiz afterwards, the AI group averaged 50%; the hand-coding group averaged 67%, a gap the authors described as nearly two letter grades (Cohen’s d of 0.738). The speed gain that was supposed to justify the trade was about two minutes and did not reach statistical significance. In other words: a large, measurable loss of understanding, purchased for almost no measurable time.

Evidence two: the scrutiny penalty. A team at Microsoft Research and Carnegie Mellon, led by Hao-Ping Lee, surveyed 319 knowledge workers about 936 real examples of using generative AI at work, and published the results at CHI 2025. Their central finding was a confidence paradox: the more confident someone was in the AI, the less critical thinking they reported applying to its output. Confidence in themselves pointed the other way. Trust in the machine and scrutiny of the machine moved in opposite directions, and scrutiny is where a surprising amount of professional skill actually lives.

Evidence three: the real-world penalty. The Lancet study, above, shows what these laboratory effects look like once they escape the lab: not a quiz score, but a missed cancer.

Three different populations. Three different measures. One direction.

What gets lost: the invisible reserve

If you take these findings seriously, a new category of organisational asset comes into view. Call it the reserve: the portion of your workforce’s capability that is not being used on any given day, but that you are relying on to exist when something goes wrong.

The reserve is the senior analyst who can sniff out that a model’s output is subtly off because she once built forecasts by hand. It is the engineer who knows what the stack trace means because he has read ten thousand of them without help. It is the editor who can tell the draft is hollow because she has written a thousand hollow drafts herself and learned what they feel like from the inside. None of these abilities shows up in a day-to-day output metric, because on a normal day the AI is right and the reserve sits idle. Its entire value is realised in the exception.

And that is the trap. Reserves look like waste to any optimisation process. A finance team looking at a human who “isn’t needed” for routine cases will see overcapacity. A well-meaning manager will see an opportunity to redeploy. The reserve is liquidated, slowly, by a thousand sensible decisions, and the first sign that it ever existed is the day it is needed and found missing.

There is a second, subtler loss: the pipeline. Today’s reserve was built by yesterday’s routine work. The junior who used to write the first draft, run the first reconciliation, debug the first ticket was not merely producing output. She was becoming a senior. Remove the routine cases and you remove the staircase by which novices climb. We explored this in an earlier edition under the name of the apprenticeship problem; the atrophy dividend is its adult cousin. One stops the reserve from forming. The other drains the reserve that already exists.

And there is a third: the calibration. If you no longer do a thing yourself, you lose your sense of how long it should take, what good looks like and where the traps are. You become a poor judge of the very output you are signing off. This is why the Microsoft/CMU finding matters so much. Skill is not only the ability to produce. It is the ability to know when something is wrong. Lose the first and the second goes with it.

A taxonomy of AI users: four ways to hold the tool

Not everyone who uses AI is on the same trajectory. The Anthropic trial is particularly valuable here, because it looked not just at whether people used the assistant but at how. Three interaction patterns preserved learning: generating code and then asking follow-up questions to understand it, asking for explanations alongside code, and asking only conceptual questions while debugging independently. What separates these from the low-scoring patterns is that the human stays cognitively in the loop. Drawing on that, and on the broader literature, it is useful to think of four archetypes.

The Passenger. The Passenger prompts, receives, pastes. The output arrives complete, and the Passenger’s role is to forward it. Passengers are the fastest people in the building and, over any horizon longer than a quarter, the most fragile. They accumulate output but not understanding, and their confidence in the tool, per the Microsoft/CMU data, is highest exactly where their scrutiny is lowest. A Passenger is one outage, one hallucination or one unusual client request away from a very visible problem.

The Proofreader. The Proofreader reads everything the AI produces and corrects the visible flaws. This feels responsible, and it is better than being a Passenger. But proofreading is recognition, not production, and the Proofreader’s underlying skill slowly erodes while their sense of diligence stays intact. Of the four archetypes this is the most dangerous, because it is the most self-deceiving. The Proofreader is working hard and getting weaker.

The Sparring Partner. The Sparring Partner attempts the problem first, or at least frames it, then uses the AI to challenge, extend and explain. They generate and then interrogate. They ask why, not just what. This maps closely onto the interaction patterns that preserved learning in the Anthropic trial. They are slightly slower on any single task and meaningfully faster at improving over time.

The Fire Marshal. The Fire Marshal is a role as much as a type: the person, usually a leader, who treats capability as infrastructure. They ask which skills the team could not perform without the tool, they schedule unassisted work on purpose, and they measure what the team can do with the system switched off. The Fire Marshal is the rarest archetype and the one organisations need most, because individuals cannot solve a structural problem with personal discipline alone.

It is worth being honest about the incentives. Almost everything in a typical workplace rewards the Passenger: faster output, quicker turnaround, visible adoption. Almost nothing rewards the Sparring Partner, whose benefits are deferred and invisible. If you leave this to individual virtue, the Passengers will win the performance review and the Sparring Partners will quietly outlast them. That is a bet most firms have not knowingly made.

The uncomfortable truths

The first uncomfortable truth: the loss feels like a gain. This is what sets skill atrophy apart from most operational risks. With the tool on, you perform better, and so you feel sharper. The decline only becomes visible when the tool is off, wrong or out of its depth, which is precisely the situation you were hoping never to face. Every individual signal says you are improving. The signal that says otherwise only arrives at the worst possible moment.

The second: your metrics are structurally blind to it. Output per head, turnaround time, tickets closed, drafts shipped: every standard measure improves while the reserve declines. You are optimising a number that is positively correlated with the problem. A dashboard that celebrates AI adoption without measuring unassisted capability is not neutral. It actively rewards the drawdown.

The third: it is irreversible faster than it is visible. Skills that took years to build can erode in months. Jarrahi is blunt about the end state: in many cases, he says, the skill is gone and he doubts it can be redeveloped. Whether or not that is always true, the asymmetry is real. Building expertise is slow. Losing it is fast. And the loss tends to be noticed after the window to prevent it has closed.

The fourth: this is not an argument against AI. The Lancet study found that the AI genuinely helped. Anthropic’s trial found that some ways of using it protected learning. Nothing here says to stop. It says to stop treating the yield as the whole story. A person who takes a dividend while checking the principal is an investor. A person who takes a dividend and never looks is, eventually, an heir to nothing.

Practical implications: running the reserve like capital

The aviation analogy is useful because it points to a solution that is neither Luddite nor naive. The answer is not to take away the autopilot. It is to protect the manual hours deliberately. Here is what that looks like at three levels.

If you are early in your career. The struggle you are tempted to skip is the apprenticeship. The practical rule is simple: attempt before you ask. Write the bad first draft, sketch the approach, form a hypothesis, then bring in the AI to challenge it. When you do ask for code or text, ask it to explain its choices, and do not close the window until you could reproduce the logic. This is the Sparring Partner pattern, and the evidence suggests it is not just virtuous but effective. Treat unassisted practice the way an athlete treats training: something you do precisely because the match does not need it today. You are building the reserve that will make you valuable when the easy cases are all automated.

If you hire and manage. Stop using speed and visible AI usage as proxies for competence. Add an unassisted component to hiring, promotion and review, sometimes, on the hard cases: a live problem-solving session, an annotated walkthrough of how someone reached a conclusion, a task where the tool is deliberately unavailable. Ask people to explain the AI’s output, not just ship it. Notice who can say where the output is likely to be wrong. That person has a functioning reserve. The one who cannot is, however fast, a Passenger. And be careful with workload: the single biggest driver of Passenger behaviour is overload, not laziness. People who are drowning do not practise.

If you lead. Treat capability as a balance-sheet item. Identify the five or six skills your organisation could not survive losing, then ask honestly how many people could perform them today without the tool. Schedule unassisted work: manual-flight hours, in aviation’s terms. That might be a monthly “tools-off” review of a critical process, rotation of juniors through unassisted work so the pipeline keeps forming, or tabletop exercises in which the AI is declared unavailable and the team must recover. Build a continuity plan that humans can actually execute, written by people who have done the work and reviewed by people who could do it again. Then measure it. Add a capability metric next to the productivity one, and watch the pair rather than either alone. If you only chart the dividend, you will keep paying it until the principal runs out.

One more practical point, for all three levels: pay attention to which tasks you give away, not just how many. Some tasks are pure friction and should be automated with a clear conscience. Others are the exercise that keeps the muscle alive. The skill is in telling them apart, and that judgement is itself a thing you only retain by continuing to practise.

The statement that never arrives

The cruellest feature of the atrophy dividend is its timing. Gains arrive immediately, loudly and with a dashboard. Losses arrive late, quietly and without a name. A financial institution that behaved like this would be shut down. A workforce that behaves like this is called “AI-forward.”

The endoscopists in Poland were not careless. They were excellent, and they were given an excellent tool, and in a matter of months a fifth of their unaided edge was gone. Nobody told them. No alarm sounded. The only reason we know is that someone thought to run the unassisted numbers again.

That is the whole assignment for the coming year. Not to resist the tool, and not to celebrate it uncritically, but to find out what you still own when it is switched off. Check the principal. Schedule the manual hours. Reward the people who can still do the thing, not only the people who can get it done.

Because every shortcut is a withdrawal from a skill account you cannot see, and the statement does not come in the post. The overdraft just shows up one day, in a room where the tool does not work and everyone is looking at you.

Denise runs marketing operations at a mid-sized software company, and on a Tuesday morning in late 2025 she got a Slack message from a junior analyst on her team: “competitive landscape deck attached, ready for the leadership review tomorrow!” The deck was seventeen slides, cleanly formatted, with a colour-coded matrix comparing eleven competitors across nine dimensions, complete with a five-bullet “strategic implications” summary on the final slide. At a glance, in the ninety seconds she had before her next meeting, it looked like exactly what she’d asked for.

It took her until nearly midnight to discover it wasn’t. Three of the eleven competitors had been discontinued or acquired more than a year earlier: information that was publicly available and trivially searchable, but which nobody had bothered to check because the deck itself looked too polished to need checking. Two of the “differentiators” listed for her own company’s product were features that had been deprecated in the last release. The “strategic implications” bullets were fluent, confident, and completely disconnected from the actual data in the matrix above them, in the specific way that happens when a document is generated end-to-end rather than built up from verified pieces. She spent that night rebuilding the deck from scratch, because by the time she’d finished checking it, fixing it would have taken longer than starting over.

The analyst hadn’t lied to her, exactly. He’d fed a prompt into a chatbot, gotten back something that looked like a finished competitive analysis, formatted it nicely, and sent it along in genuine good faith, under real time pressure, believing he’d done the job. In his mind, the deck was done. In Denise’s, it had just become a different job entirely: verifying and rebuilding something that presented itself as already trustworthy. Nobody in that exchange thinks of themselves as having produced or received a failure. That’s precisely what makes it so hard to see, and why it took researchers naming the pattern before most organisations noticed they were drowning in it.

Naming the thing

In September 2025, researchers at BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 full-time U.S. desk workers and gave this pattern a name that stuck almost immediately across the business press:

workslop. Their definition is precise, and worth sitting with, because it’s narrower than “bad AI content”: workslop is AI-generated work that looks complete and polished on the surface, but lacks the substance to actually accomplish the task it was meant to accomplish, and in doing so, quietly offloads the real cognitive labour onto whoever receives it next.

The numbers were larger than almost anyone in the study expected going in. Forty percent of employees said they’d received workslop from a colleague in just the past month. Workers estimated that 15.4% of everything they receive from teammates now falls into this category. More than half of managers, 54%, report receiving it regularly, compared with 38.5% of individual contributors, a gap that makes sense once you realise managers sit at the checkpoint where unfinished thinking either gets caught or ships to a client. And more than half of respondents, 53%, admitted to having sent workslop themselves at least once, which is the detail that turns this from a story about a few careless employees into a story about an entire incentive structure.

Stanford’s Jeff Hancock, who co-led the research, described the underlying shift in a way that cuts closer to the mechanism than most coverage of “AI at work” usually gets:

“Now that [the effort] piece is gone, I can generate a lot of useless or unproductive content very easily.”

That sentence is worth reading twice. It isn’t a complaint about AI quality. It’s a description of what happens when the cost of producing the appearance of finished work drops to nearly zero, while the cost of verifying whether it’s actually finished stays exactly where it always was: squarely on another human being’s afternoon.

How we got here: a familiar shape wearing a new name

Every technology that lowers the cost of producing plausible-looking output eventually produces some version of this problem, and it’s worth being honest that workslop isn’t entirely unprecedented. Desktop publishing software in the 1990s produced a wave of beautifully formatted internal memos that said almost nothing: the joke at the time was that PowerPoint had made it possible to look prepared without being prepared. Stock photography and template decks did something similar a decade later: professional-looking scaffolding standing in for actual thinking. What’s different this time is the scale, the speed, and the specific way the deception operates. A 1990s memo took real time to format even if it said nothing. Workslop takes twelve minutes to generate something that would have taken three hours to produce honestly, and it does so with a fluency and internal consistency that makes the absence of substance far harder to spot on a first read than a sloppy old-fashioned memo ever was.

There’s a useful contrast with an adjacent, better-known failure mode: AI hallucination. Hallucination is usually a factual problem: the model states something false with confidence, and once you catch it, the fix is straightforward: correct the fact. Workslop is structurally different and, in a specific sense, more corrosive: nothing in it may be technically false. The competitor names in Denise’s deck were real companies. The formatting was accurate. What was missing wasn’t correctness: it was judgement about what actually mattered, what needed verifying, and what the document was supposed to accomplish for the person receiving it. You can fact-check a hallucination. It’s much harder to fact-check the absence of thought.

The MIT Media Lab’s NANDA initiative added a data point in August 2025 that reads, at first, like a completely separate finding, but fits the workslop pattern almost exactly: 95% of organisations piloting generative AI in business processes report no measurable return on that investment. Adoption metrics across the same organisations were climbing fast: more licences issued, more tools rolled out, more employees “using AI” by any usage-log definition. The value those adoption charts were supposed to represent simply wasn’t showing up in the numbers that mattered. Workslop is a large part of the missing link between those two graphs: usage goes up and to the right, while a meaningful share of what that usage produces creates net-negative work for whoever downstream has to deal with it.

The mechanism: fluency severed from labour

Here is the mechanism in its simplest form. Human communication has always used fluency (clean structure, confident tone, professional formatting) as a rough, imperfect but genuinely useful proxy for effort and thought. A well-organised report was reasonable evidence that someone had done the underlying work, not because fluency and substance are the same thing, but because, historically, producing fluent output without substance took almost as much effort as just doing the work properly. That coupling has now snapped, cleanly and almost overnight, for a huge share of everyday knowledge work.

A generative model can produce fluency (clean structure, confident tone, polished formatting) for close to zero marginal effort, entirely independent of whether any actual thinking happened underneath it. That means the signal receivers have relied on for decades to triage their attention, “does this look put-together?”, has stopped correlating reliably with the thing it was always a proxy for. And because most people haven’t consciously updated their triage instincts to reflect that, workslop routinely sails through the first-glance filter that would have caught a genuinely sloppy, honestly-unfinished piece of work in the pre-AI era. Ironically, the worse and more obviously rushed a document looks, the more scrutiny it tends to get. Workslop is dangerous specifically because it looks finished, which is exactly when people’s guard comes down.

The direction of flow matters too, and the study’s data on it is telling. Workslop moves sideways between peers 40% of the time, up to managers 19% of the time, and down from leaders to teams 16% of the time. It isn’t concentrated at any single level of seniority: it’s a property of the incentive structure itself, appearing wherever someone faces a task, a time constraint, and a tool that can manufacture the appearance of having met the deadline faster than actually meeting it.

There’s also a straightforward asymmetry hiding inside the mechanism, and it’s the part that makes workslop so durable once it starts. Generating fluent, plausible-looking output costs the sender almost nothing: a prompt, a few seconds of model latency, a quick skim for typos. Verifying that same output properly, checking the numbers, confirming the named facts, testing whether the logic actually holds together, costs the recipient real time and real attention, often more than it would have taken to produce the thing honestly from scratch. Any system where the cost of producing a claim is a rounding error compared with the cost of checking it will, predictably, fill up with unverified claims. That isn’t a statement about anyone’s character. It’s closer to a law of incentives, and it would apply to workslop’s human-only predecessors too, if humans could generate boilerplate as cheaply as a model can.

What gets lost

The most visible cost is the time. The average piece of workslop costs its recipient one hour and fifty-one minutes to identify, unpick, and either fix or redo: about twenty minutes longer, on average, than it would have taken the original sender to simply do the work properly the first time. Multiply that across a large organisation and BetterUp’s researchers arrive at roughly $186 in wasted time per affected employee per month, or north of $9 million a year in a company of 10,000 people. That number exists nowhere on any organisation’s financial statements, because there is no line item for “hours spent decoding a colleague’s chatbot output,” even though the hours are entirely real and entirely lost.

The second cost is relational, and the study’s data here is sharper than most coverage of it acknowledges. When people receive workslop, 54% report feeling annoyed, 46% frustrated, 38% confused, and 22% describe themselves as outright offended: a striking word to use about a piece of office correspondence, and a sign of how personally people take being handed something that wastes their time while pretending not to. More consequentially, more than half of recipients say the experience made them view the sender as less creative, less capable, and less reliable. Over 40% see the sender as less trustworthy going forward. Nearly a third say they’re less willing to collaborate with that person again. The person who sent the workslop, in other words, thought they were being efficient. Their colleague now quietly trusts them less: a cost that never shows up in any productivity metric, but shows up very clearly the next time that person needs a favour, a reference, or the benefit of the doubt.

The third cost is the hardest to quantify and, in the long run, probably the most important: an erosion of the norm that finished work should actually be finished before it leaves your hands. Once workslop becomes common enough inside a team, the entire distribution of trust shifts. People start reflexively re-verifying things that used to be safe to take at face value, which is its own tax, levied on everyone, including the honest senders who never produced a slop deck in their lives.

Four ways this shows up

Not everyone producing or receiving workslop is doing the same thing for the same reason, and collapsing all of it into “people being lazy with AI” misses most of what’s actually happening. Four recognisable patterns show up across the research and across most teams once you start looking for them.

The Overwhelmed Corner-Cutter. Genuinely underwater (too many deliverables, not enough hours) and reaching for AI as the only way to hit a deadline that was unrealistic before AI ever entered the picture. This person isn’t being dishonest so much as making a rational trade under bad constraints: something delivered on time beats something correct delivered late, in a culture that measures the former far more visibly than the latter.

The Mandate Follower. Was told, often vaguely, to “use AI more” by leadership eager to show adoption numbers, with no accompanying guidance on where AI output still needs human judgement layered on top. This person is complying with an instruction as literally as it was given. The resulting workslop isn’t a personal failure so much as a predictable outcome of an unclear directive handed down without the quality guardrails to go with it.

The Confident Skimmer. Uses AI output as a first draft with every intention of reviewing it properly, then runs out of time, attention, or diligence before the review actually happens, and ships the draft anyway, genuinely believing it’s closer to finished than it is. This is arguably the largest and most sympathetic category, and the hardest to fix with policy alone, because the failure happens silently, inside someone’s own judgement about their own remaining time.

The Downstream Absorber. The recipient-side archetype: the manager, teammate, or client-facing employee who catches workslop before it does real damage, at direct personal cost: the extra two hours, the late night, the quiet erosion of trust in a colleague they still have to work with tomorrow. This is the person actually paying the tax the other three archetypes generate, and the one whose cost is least visible in any usage dashboard, because absorbing workslop successfully looks, from the outside, exactly like normal work getting done.

Most teams contain some blend of all four simultaneously, which is exactly why a single fix (a training session, a new AI usage policy, a stern all-hands email) rarely solves the whole problem on its own. Solve for the Mandate Follower with clearer guidance and you’ll still have Overwhelmed Corner-Cutters reaching for the fastest available option under deadline pressure the guidance never addressed, and Confident Skimmers whose failure happens entirely inside their own private judgement about how much review time they actually have left. A policy aimed at one archetype is, at best, a partial fix for the other three.

What actually holds up

The researchers behind the original study were careful to frame their findings as a management problem rather than an individual-behaviour problem, and the practical implications follow from taking that framing seriously.

For individual contributors: treat AI-generated first drafts exactly like drafts, not like finished deliverables, regardless of how polished they look on screen. The specific discipline that helps most is verifying the parts of any AI output that are cheapest to fake convincingly (named facts, comparative claims, anything presented with confident specificity) because those are precisely the details that read as most credible while requiring the least underlying effort to fabricate.

For managers: the instinct to reward speed and visible AI adoption needs a counterweight that explicitly values verified, load-bearing work over merely fast-looking work. A few managers interviewed alongside the original study started asking a simple diagnostic question before accepting any AI-assisted deliverable: “walk me through the one part of this you’d bet your reputation on.” Workslop tends to fail that question immediately, because there usually isn’t a specific part the sender actually stands behind: the whole document was generated with roughly uniform, unverified confidence, cover to cover.

For leaders: an AI adoption mandate that measures itself purely by usage (seats activated, queries run, tools rolled out) is optimising for exactly the wrong variable, and workslop is the fully predictable output of that specific measurement choice. The fix isn’t slowing AI adoption down. It’s pairing every mandate to “use AI” with explicit, concrete guidance on where human judgement still has to sit on top of the output, and building review capacity into timelines rather than assuming AI adoption makes that step optional. Leadership research on this consistently points to the same root causes: unclear mandates from the top, and overwhelmed teams with no safe way to admit they don’t have time to do the verification properly. Both of those are organisational design problems, not individual character problems, and both are entirely within a leadership team’s power to fix, which is, depending on how you look at it, either the most hopeful or the most damning part of the entire finding.

The closing thought

Denise never told her analyst how much of that night she spent rebuilding his deck from scratch. She didn’t think it would land the way she needed it to: he genuinely believed he’d delivered on the ask, and in a narrow sense, measured against the instruction he was actually given, he had. What she quietly changed instead was her own definition of “done” for anything AI-assisted crossing her desk: not delivered, not polished, not confident-sounding, but specifically verified, on the parts that would actually cost something if they turned out to be wrong.

That’s a small, individual-level fix to a problem the research is clear is mostly structural, and it won’t scale to an entire organisation on its own. But it points at the actual shape of what’s being lost underneath the adoption charts and the usage dashboards, all trending confidently up and to the right: not effort, exactly, which AI has always been honest about reducing, but the quiet, unglamorous labour of deciding what’s actually true and what actually matters before you hand something to somebody else. That labour hasn’t gotten any cheaper. It’s just gotten a great deal easier to skip, and, for now, almost entirely someone else’s problem to notice.

The organisations that get ahead of this won’t be the ones that slow down AI adoption or ban the tools outright; the numbers make clear that ship has sailed and isn’t coming back. They’ll be the ones that figure out, deliberately and early, how to keep paying for judgement even as the cost of fluency collapses to zero, because judgement was always the scarce, valuable half of the work. Fluency was just the part that used to be a reasonably reliable clue that judgement had happened underneath it. It no longer is, and pretending otherwise is how a $9 million-a-year tax quietly becomes a permanent line in the budget nobody ever agreed to add.

The interview that was too clean

Priya has screened engineering candidates for eleven years, long enough to develop the thing every senior recruiter eventually develops: a private, unscientific but rarely-wrong sense for when an interview is going too well. The candidate in front of her that Tuesday in November 2025 was going too well. Thirty-five minutes into a live system-design round, he had produced a clean, well-structured answer to a deliberately ambiguous scaling problem, complete with the kind of trade-off language, “it depends on your consistency requirements here”, that usually takes people years to acquire, not months. He paused for exactly the right amount of time before each point, the way someone pauses when they’re reading rather than recalling. When Priya asked one small follow-up, off-script, about a detail he hadn’t mentioned, there was a half-second delay that hadn’t been there before, and then a perfectly composed answer to a question she hadn’t actually thought was answerable that cleanly.

She didn’t flag him in the moment. She wasn’t sure enough, and being wrong about this kind of accusation carries its own cost. She let the interview finish, gave him a strong internal score, and only two weeks later, reading a report her company’s newly deployed AI-detection tool had generated on the recorded session, did she learn what her instinct had been circling: micro-patterns in his response timing, consistent with reading rather than generating in real time, flagged at 89% confidence. He had, in all likelihood, been fed the answer through a hidden overlay, the kind marketed today as “invisible to screen-share”: Cluely, Final Round AI, LockedIn AI, and a dozen less famous competitors, all built for exactly this moment, all doing a brisk business.

He didn’t get the job. But Priya’s story isn’t really about him. It’s about the fact that she almost missed it, that her company only caught it because they’d bought expensive detection software specifically built for this exact scenario, and that, by the numbers now coming out of the recruiting industry, for every candidate like him who gets caught, a comparable number don’t. Across 19,368 live technical interviews conducted between July 2025 and January 2026, 38.5% of candidates were flagged for AI-assisted cheating behaviour, with the flag rate for technical roles alone climbing to 48%. And of the candidates flagged, 61% scored above the passing threshold and received an offer anyway. The detection worked. The consequences, for most of them, did not follow.

How we got here: the short history of a proxy

The job interview has never been a perfect measure of anything. It has always been a proxy: a stand-in, chosen because it was practical, for a thing employers actually want to know and can’t directly observe: can this person do the work, reliably, under real conditions, for years. Different eras picked different proxies and believed in them with roughly the same misplaced confidence.

The whiteboard-algorithm era, which peaked in tech hiring through the 2010s, bet that the ability to invert a binary tree under time pressure correlated with the ability to ship production software. It didn’t correlate especially well: a body of research eventually made that case convincingly enough that companies started phasing whiteboard rounds out even before AI entered the picture, partly because AI coding assistants had already made memorising algorithms less relevant to the actual job. The behavioural-interview era that layered on top of it bet that a well-told story about a past conflict predicted how someone would handle a future one. That proxy had its own well-documented flaws (coachability, rehearsal, the simple fact that articulate storytellers aren’t always the strongest performers) but it survived for decades because, flawed as it was, nobody had a cheaper or faster substitute.

What’s different now isn’t that the proxy is imperfect. Proxies have always been imperfect. What’s different is that the proxy has become gameable at scale, cheaply, invisibly, by anyone with an internet connection and twenty dollars a month, and gameable in a way that produces a result which is, to a human evaluator watching in real time, functionally indistinguishable from the real thing. Previous generations of interview gaming (a coached answer, a rehearsed story, a friend’s practice questions) still required the candidate to actually absorb and reproduce the material themselves. The AI-overlay era removes that last requirement entirely. The candidate doesn’t need to know the answer. They need to be willing to read it, fluently, off a screen nobody else can see, in real time, while maintaining eye contact and a confident tone. That is a categorically easier skill to acquire than the thing the interview was originally built to measure, and it produces results that score, on every visible metric, just as well.

There’s a useful parallel in how academic testing handled a similar shock a few years earlier. When large language models first made take-home essays trivially fakeable, some institutions responded by banning AI outright and trying to detect its use after the fact: an approach that produced a steady stream of false accusations against students who simply write in a clean, structured style, and that degraded almost as fast as the detection tools it depended on. Others responded by redesigning the assessment itself: oral defences, in-class components, work that had to be produced and explained live, in the room, with no opportunity to outsource the thinking beforehand. The second approach held up considerably better, for a simple reason: it stopped trying to detect the presence of AI and started making AI assistance structurally useless for the specific thing being measured. Hiring is now living through the identical decision point, several years behind education and with far higher stakes attached to getting it wrong.

“The candidates most equipped to beat this system aren’t the most competent ones. They’re the ones most comfortable performing competence they don’t have — which is, if you sit with it for a moment, close to the exact opposite of what a hiring process is supposed to select for.”

The mechanism: an arms race with no ceiling

The tools themselves have moved fast enough that describing them risks going stale within a quarter, but the shape of the arms race is stable enough to describe with some confidence. On the candidate side: browser overlays that sit invisibly on top of a video-call window, capturing the interviewer’s question via audio or screen-share, running it through a language model, and displaying a generated answer in a window that doesn’t appear in anything the interviewer can see, because it’s rendered outside the shared application, or because it exploits how screen-share capture works at a technical level. Some candidates pair this with a Bluetooth earpiece and real-time audio coaching instead of a visual overlay, reading a script fed to them word for word. A smaller, more aggressive tier uses deepfake video tools to swap out who’s actually appearing on camera entirely, letting a more experienced person sit the interview on someone else’s behalf while their face is rendered convincingly onto the candidate’s video feed. Tools originally built for the wave of smart-glasses cheating that got them banned from the SAT are, unsurprisingly, being repurposed for the exact same function in a Zoom call.

On the employer side: the response has been to fight AI with AI. 61% of companies surveyed in Greenhouse’s 2026 AI in Hiring Report now use dedicated detection software during live interviews: analysing response-timing patterns, eye-gaze consistency, behavioural biometrics, and the subtle statistical fingerprints that distinguish someone reading generated text from someone generating an answer in real time as they speak. Another meaningful share of companies are actively evaluating similar tools. It is, in the most literal sense, an arms race: an entire specialised market has emerged on both sides of the interview table simultaneously, each side selling tools designed to defeat the other side’s tools, with candidates and recruiters as the actual customers on either end.

Nobody involved in this arms race describes it as stable or winning. 62% of hiring professionals surveyed openly admit that candidates are getting better at faking their way through AI-assisted deception faster than recruiters are getting better at catching it. Human accuracy at spotting deepfaked video in an interview setting was measured at just 55.5%, barely better than a coin flip, which is a genuinely remarkable thing to discover about a task hiring managers have spent their careers believing they were good at. And the deception isn’t limited to live interviews. 91% of recruiters in the same Greenhouse survey report encountering some form of candidate deception in the broader hiring process this year: 63% report resume exaggeration, 48% report fabricated references, and a smaller but rapidly growing share report prompt injections: hidden text embedded in a résumé file, invisible to a human reader, engineered specifically to manipulate an AI resume-screening tool into ranking the candidate higher regardless of their actual qualifications, before a human recruiter ever opens the file.

What gets lost

The most obvious cost of all this is the one that gets the headlines: a company might hire someone who genuinely cannot do the job they were hired for, discovering it only once real stakes are attached and the scaffolding is gone. That cost is real, but it’s not actually the largest one, and treating it as the whole story misses what’s happening underneath it.

The first hidden cost is speed and cost, distributed across every candidate rather than concentrated on the dishonest ones. 68% of hiring managers report being more personally involved in hiring than they were a year ago: more manual screening, more live problem-solving exercises, more in-person final rounds specifically reintroduced to strip away the conditions that make AI-assisted cheating possible. Google, Cisco, and McKinsey have all shifted at least their final interview rounds back to in-person formats for exactly this reason. That’s a rational institutional response to a real problem. It’s also a tax on everyone in the pipeline, honest and dishonest alike, in the form of longer processes, more logistical friction, and more expensive hiring per head: a cost that didn’t exist five years ago and now exists because the interview itself can no longer be trusted at face value.

The second hidden cost is a trust collapse that runs in both directions simultaneously, and is arguably more corrosive than either side’s individual behaviour. 70% of hiring managers say they now trust AI tools to make faster, better hiring decisions than they could make unassisted. Only 8% of job seekers describe AI-driven hiring processes as fair. Neither number is really about AI’s actual accuracy. Both numbers are about a hiring relationship where each side has become convinced, with substantial justification, that the other side is gaming the system: hiring managers assuming candidates are cheating, candidates assuming AI screening tools are unfairly filtering them out before a human ever looks at their application. Trust, once it degrades this broadly across an entire labour-market relationship, doesn’t come back because one detection tool gets slightly more accurate. It comes back, if it comes back at all, only once both sides believe the underlying process is worth trusting again, and right now, neither side does.

The third and least visible cost falls specifically on honest candidates, and it’s the one that should worry anyone who cares about fairness in hiring more than the headline fraud statistics do. The candidate who spent genuine hours preparing, who answers a beat slower because they’re actually generating the answer live rather than reading it, who occasionally says “let me think about that for a second” because they are, in fact, thinking, that candidate is now being scored, in real time, against a fluency benchmark partly set by people who aren’t actually thinking at all. Some of those honest candidates are starting to reach for the same tools themselves, not because they want to deceive anyone, but because refusing to use an AI overlay increasingly feels less like integrity and more like showing up unarmed to a competition where the rules have already changed for everyone else.

Four faces of the interview theatre

The people inside this system aren’t uniform, and treating “AI cheating in interviews” as one behaviour obscures more than it reveals. In practice, four recognisable archetypes show up on both sides of the table.

The Augmented Performer. Uses an AI overlay not because they lack any relevant knowledge, but because the tool removes risk from an already-stressful, high-stakes performance: reading a slightly better-structured version of an answer they could have produced themselves, just slower and less cleanly. This is the hardest case morally and the hardest case to detect, because the underlying competence often genuinely exists; what’s fabricated is the polish, not the substance.

The Panicked Improviser. Has little relevant knowledge for the specific role, discovered the tool out of desperation in a brutal hiring market, and is functionally borrowing an entire skill set for the length of one video call. This is the case the fraud statistics are mostly measuring, and the one most likely to produce a painful collapse once the job actually starts and the overlay is gone.

The Honest Straggler. Refuses to use any AI assistance during interviews, on principle or simply because they haven’t thought to, and is now competing directly against a fluency bar partly set by candidates in the first two categories: absorbing, without realising it, most of the real cost this entire dynamic imposes on the labour market.

The Detector’s Blind Spot. The hiring-side archetype: a recruiter or hiring manager who trusts their own in-person read of a candidate more than they trust detection software, unaware that human accuracy at spotting this specific category of deception sits close to chance. Not naive, exactly, just operating on an instinct that used to be reliable and now, quietly, often isn’t.

Most real interviews contain some blend of these on both sides of the table simultaneously, and the blend is precisely what makes any single detection method (human intuition, statistical timing analysis, in-person format) insufficient on its own. A company that solves for the Panicked Improviser with better timing-analysis software will still walk directly into the Augmented Performer, because the statistical fingerprint of “reading a slightly better version of an answer I could have produced myself” is far subtler than the fingerprint of “reading an answer to a question I have never encountered before.” Solving for one archetype at a time is how a detection strategy ends up looking effective in a vendor’s own case studies while quietly missing most of the actual problem.

What actually holds up

None of this is an argument that hiring is doomed, or that every candidate using AI assistance is acting in bad faith, or that detection software is a solved problem. It’s an argument for being honest about what an interview can and can’t currently tell you, and adjusting accordingly on both sides of the table.

For candidates: the signal that used to differentiate you, sounding sharp and fluent under pressure, is now available to anyone with a hidden browser tab, which means it differentiates you less than it used to, whether or not you personally use any AI assistance. What still differentiates you is depth that survives unscripted follow-up: a body of real, specific, defensible work you can discuss for fifteen minutes past the point any overlay would have run dry, in detail no generated script would have anticipated. Build toward being interrogated, not toward being fluent for exactly as long as a scripted answer lasts.

For hiring teams: stop optimising interview scoring for fluency, because fluency is now the cheapest thing in the entire process to fake. Score the follow-up chain instead: five or six specific, unscripted “why” and “what if” questions that force a candidate to reason live, off any script, about their own stated answer. Overlay tools can sustain a first answer convincingly. They struggle badly, and increasingly visibly, past the third layer of genuine improvisation, because genuine improvisation is precisely the thing they were never built to generate under real-time constraint.

For leaders: if you can’t currently describe, with some confidence, how your last ten hires actually got hired (not how fast, not how cheaply, but on what evidence you now trust) you’re running an organisation built partly on signal you can no longer verify. That gap doesn’t show up in an onboarding week. It shows up the first time someone on your team faces a genuinely novel problem with no script, no earpiece, and no panel watching their confidence instead of their actual output, and by then, the cost of finding out has moved from a bad interview to a bad quarter.

The closing thought

Priya still can’t fully explain, months later, what tipped her off about that candidate before the detection report confirmed it. She describes it as a feeling that the interview was “too finished”, that every answer arrived complete, in a way real thinking rarely does, without the small hesitations and self-corrections and half-formed detours that usually mark someone actually working through a problem live. She has started listening for those detours deliberately now, in every interview she runs, treating the mess of real thinking as a signal worth scoring rather than a flaw worth penalising.

That’s a small, human-scale adjustment to an arms race that neither candidates nor employers are close to winning outright. It won’t stop the next overlay tool, and it won’t restore the interview to whatever imperfect authority it briefly had. But it points at the thing actually worth protecting underneath all of this: not the appearance of fluency, which has become nearly worthless as a signal, but the visible, occasionally clumsy evidence of someone actually thinking, live, with nowhere to hide it and no script to hide behind. That evidence is still real. It’s just gotten a great deal harder to find in a room full of people who’ve learned exactly how to look like they have it.

Renee has been at the same regional insurance company for four years: three as an underwriting analyst, and one, technically, in the exact same seat, waiting for a promotion that got frozen indefinitely eleven months ago in a two-line Slack message from HR: “pausing internal role changes while we finalise headcount planning.”

She still opens LinkedIn most mornings, out of habit more than hope. She still gets recruiter messages, two or three a month, for roles that would pay maybe 4% more than she makes now, in exchange for an unknown manager, an unknown culture, and a probation period she’d be serving during the least stable hiring market in a decade. She closes the tab. She has done this, by her own count, more than thirty times this year.

Renee isn’t lazy, and she isn’t disengaged in the way her company’s annual survey would flag her. She would tell you, if you asked, that she’s being smart. And on the numbers, she’s right: 57% of workers now describe themselves the same way Renee would describe herself — a “job hugger,” someone staying in a role not out of enthusiasm but out of a calculated read on how dangerous the alternative currently looks. Five months earlier, when ResumeBuilder ran the identical survey, that number was 45%. Renee’s read on the market didn’t change in five months. The market did.

That twelve-point jump is the real story here, and it isn’t really a story about Renee, or about any individual’s tolerance for a job they’ve outgrown. It’s a story about what happens to an entire workforce when the two escape routes people have always used to get out of a stalled career — leaving, or moving up where they are — both quietly seize up in the same eighteen-month window.

What job hugging replaced

It’s worth remembering how recently the opposite was true. In 2021 and 2022, at the peak of what got called the Great Resignation — and, more accurately for what actually happened, the Great Reshuffle — quitting your job wasn’t a risk. It was frequently the single most reliable way to get a raise. Workers who switched employers during that window saw median wage growth of roughly 16%, compared to single digits for people who stayed — an 8-to-10 percentage-point gap that made job-hopping, for a couple of years, close to a rational financial obligation. Company loyalty, as a concept, briefly looked almost quaint. LinkedIn profiles updated themselves constantly. Counteroffers became a genre of their own.

That era didn’t end with a crash. It ended with a slow freeze, and the freeze has a name now, and it isn’t “the Great Resignation” running in reverse. “Job hugging” describes something categorically different from simple loyalty or contentment: it’s staying put specifically because the calculated risk of leaving has gone up, not because the reward of staying has gone up. The distinction matters enormously, because it means the underlying dissatisfaction that used to drive people toward the exit hasn’t disappeared. It has simply been priced out of acting on itself.

This isn’t the first time the labor market has swung this hard in the other direction. After the 2008 financial crisis, a similar freeze produced what economists at the time called “job lock” — people staying in roles they disliked because unemployment felt catastrophic and re-entry was slow. But job lock in 2009 was mostly about the absence of open positions. Job hugging in 2026 is stranger, because it’s happening while AI vendors and executives are simultaneously promising a productivity boom, while job postings for AI-adjacent roles are, in some categories, actually growing, and while unemployment itself remains historically unremarkable. The freeze isn’t a shortage of jobs in the aggregate. It’s a shortage of jobs anyone currently trusts enough to jump toward — a subtler, more psychological kind of lock, built less on scarcity than on uncertainty about which roles AI reshapes next, and when.

The mechanism: why leaving got this expensive

Three numbers explain most of what’s happened. The first is the vanishing wage premium for switching jobs. By February 2026, the Federal Reserve Bank of Atlanta’s Wage Growth Tracker showed job switchers earning median wage growth of 4.7%, against 3.6% for people who stayed in place — a gap of about one percentage point, down from 8 to 10 points at the 2022 peak. The single biggest financial argument for leaving a job has, for most workers, all but disappeared.

The second is the price of getting it wrong. The average worker’s reservation wage — the lowest offer they say they’d actually accept to switch jobs — hit a record $88,387 in July 2026. Reservation wages rise when the perceived risk of unemployment rises; workers are effectively pricing in the cost of landing somewhere unstable, in a market where getting laid off within your first year at a new employer carries much higher stakes than it did when hiring was healthy.

The third is the hiring and layoff data underneath both of those numbers. Planned U.S. hires fell to 507,647 in 2025 — the lowest level since 2010, and a 34% drop from the year before. In the same year, employers announced 1,206,374 job cuts, up 58% year-over-year, with AI cited explicitly as the reason for 54,836 of them. Seventy percent of workers now say they’re worried AI could threaten their own job within six months. And underneath all of it, the U.S. quits rate — the cleanest available measure of how many people feel safe enough to walk out the door — has sat at or below 2.0% for seven consecutive months, near its lowest reading since 2016, excluding the initial COVID-19 lockdown.

Put simply: the market spent three years teaching workers that moving pays. It has spent the last year and a half teaching them the opposite lesson, just as convincingly, and workers are behaving exactly as rationally now as they were behaving then.

None of this required anyone to announce a new policy or coordinate a response. It’s an emergent property of thousands of individual, entirely sensible decisions, made simultaneously, by people who’ve done the math and concluded that the devil they know is currently underpriced compared to the devil they don’t.

The escalator was supposed to be the backup plan

Here’s the part of the story that should worry leaders more than the quit-rate headline. When the external market freezes, the textbook answer is internal mobility — let people move up, sideways, or into a new team without ever leaving the building, and you retain the ambition that would otherwise walk out the door. LinkedIn’s own February 2026 Workforce Report shows why that backup plan isn’t holding: national hiring is down 5.7% year over year and employers are taking longer to fill open roles, but internal movement isn’t absorbing the slack the way it’s supposed to. Fifty-one percent of employees say they’re simply unaware of what internal openings exist at their own company. And sixty percent of high-potential employees name their own direct manager as the primary obstacle standing between them and an internal move.

That second number is the more revealing one, because it points to a second-order effect that’s easy to miss. In a hiring freeze, a manager who loses a strong performer to an internal transfer frequently can’t backfill the seat — headcount approvals are frozen right alongside external hiring. So the same freeze that’s making external moves feel dangerous for employees is quietly giving their managers a private incentive to make internal moves harder, too, whether or not that’s ever said out loud in a one-on-one. Nobody designed this as a retention strategy. It emerged as one anyway, and it’s a strategy built entirely on friction rather than on anything an employee would recognize as genuine investment in them.

The result: the two ways people have always advanced — leaving, or moving up where they are — have both seized in the same window. Nobody’s climbing. Nobody’s transferring. The escalator that was supposed to be the pressure-release valve for a frozen job market is itself frozen, and everyone standing on it is calling the stillness “stability” because the alternative explanation is more uncomfortable to sit with.

What gets lost

The direct cost of job hugging is invisible almost by design — it shows up as a number that isn’t there, a resignation letter that was drafted and never sent, a recruiter call that ended in “I’m happy where I am.” The indirect costs are not invisible at all, once you know where to look, and they show up in three places.

The first is engagement, and the data here is stark. Fifty-four percent of employees report experiencing some degree of “quiet cracking” — a slow erosion of morale and discretionary effort distinct from quiet quitting, because it isn’t a decision to disengage so much as a gradual, often unconscious drift into disengagement under sustained insecurity. One in five describe it as frequent or constant. Gallup’s most recent workforce data puts active engagement at just 31% of U.S. employees — meaning most retention dashboards showing record-low attrition are sitting directly on top of engagement numbers that tell almost the opposite story. Globally, Gallup estimates disengagement costs the world economy roughly $8.9 trillion a year in lost productivity — a bill that doesn’t appear on any single company’s P&L, but is being paid all the same, distributed invisibly across millions of Renees quietly doing the minimum required to stay safely unremarkable.

The second is skill development, and it compounds the first. Employees who haven’t received any training in the past year are 140% more likely to report feeling job insecure — which means the workers least equipped to move quickly if the market does thaw are disproportionately the same ones sitting still right now, and the ones least likely to be actively developed while they wait. Lateral moves, stretch assignments, and new-employer onboarding have always been quiet but powerful skill-building mechanisms; when all three stall simultaneously, an entire cohort’s skill growth stalls along with them, at precisely the moment AI is raising the bar on what “current” skills need to look like.

The third is harder to quantify but arguably the most expensive: the erosion of what “good work” is understood to require. When advancement stops being connected to performance — because there’s simply nowhere to advance to — the signal that used to link effort and outcome gets fainter. People don’t stop caring overnight. But caring without any visible mechanism for it to matter is exhausting to sustain, and exhaustion, given enough time, looks identical to indifference from the outside.

Four faces of the stalled worker

Job hugging isn’t one uniform posture, and treating it as one is how well-meaning retention initiatives end up solving the wrong problem — a generic “we value you” email does nothing for someone who’s actually being structurally blocked by their own manager, and a mobility portal does nothing for someone who’s already decided no internal move is worth the risk of losing their current, at least-known, seat. In practice, job hugging tends to show up as one of four recognizable types.

The Loyalist by Necessity. Would genuinely leave if the math worked, and knows exactly what the math would need to look like. Runs the numbers every few months, usually after a recruiter call, and closes the tab in the same place every time — the risk-adjusted return still isn’t there. Not disengaged; extremely aware.

The Waiting List. Was told, explicitly or implicitly, that a promotion or transfer is coming — “when budget opens back up,” “next planning cycle” — and has now been told that for multiple cycles running. Still believes it, mostly, because believing it is easier than reassessing everything else about the decision to stay.

The Ghost Employee. Present, competent, quietly checked out. This is quiet cracking in its purest form: doing the job well enough that nothing triggers a conversation, while investing nothing beyond what’s required, because investing more hasn’t paid off in longer than they can remember.

The Handcuffed Manager’s Report. Wants to move internally, has the skills and the manager’s private acknowledgment that they’d thrive elsewhere in the org — and is blocked by that same manager, not out of malice, but because losing them means losing a headcount line nobody can promise will be refilled.

Most job huggers are some blend of at least two of these at once, and the blend shifts month to month as recruiter calls come in, promotion cycles pass, and the broader hiring data either firms up or doesn’t.

The compounding problem

The uncomfortable part isn’t the current state. It’s what happens when it ends. Seventy-one percent of current job huggers expect to keep hugging their job for at least six more months; 44% think it will take a year or longer before they feel secure enough to move again. That’s not a workforce settling into a new equilibrium — it’s a workforce under sustained pressure, with the release valve deliberately held shut, for a duration long enough that the pressure itself becomes the dominant fact of the arrangement rather than a temporary condition of it.

Pressure held that long doesn’t release gently when it finally does release. It tends to release all at once, in a rush that looks nothing like the gradual, individually-timed departures a healthy labor market produces. Organizations that spent 2026 quietly congratulating themselves on record-low attrition, without doing anything to address the engagement and mobility numbers sitting underneath it, are the ones most exposed to that rush — because low attrition built on fear rather than genuine investment doesn’t convert into loyalty the moment fear subsides. It converts into an exit queue that’s been forming, silently, the entire time.

Some sectors are more exposed than others. Industries where AI has already visibly reshaped day-to-day roles — customer service, entry-level analysis, first-draft creative and marketing work — are sitting on the largest hidden queues, because the workers inside them are simultaneously the most anxious about staying and the most anxious about leaving, unsure whether a new employer’s version of their role even looks like the one they currently hold. Industries where AI’s effect has been slower or less visible — skilled trades, in-person healthcare delivery, parts of regulated finance — are seeing comparatively less hugging behavior, not because those workers are less rational, but because their risk calculation simply has fewer unknowns in it. The size of the eventual thaw, in other words, will not be evenly distributed. It will land hardest on exactly the functions leadership is currently most confident are “stable,” because stability was never the honest description of what was happening there.

What actually holds up

None of this is an argument that workers should quit recklessly, or that leaders caused a labor market they didn’t design. It’s an argument for treating the current calm as exactly what the data says it is — a function of fear, not of health — and building accordingly.

For individuals: staying put right now can look identical to being loyal and be, underneath, just fear wearing a nicer outfit. That’s not a criticism — it’s frequently the financially sound decision — but it only pays off long-term if the time spent staying is actively used, not just endured. The job huggers who benefit most when hiring thaws will be the ones who spent this period building a visible track record where they already are: new skills, cross-team visibility, a wider internal network, a portfolio of finished work that speaks for them the moment the market opens back up. Waiting and building are not the same activity, even though they can look identical from the outside for months at a time.

For hiring managers: a candidate with an unusually long, static tenure right now is not automatically a red flag, and it’s not automatically a loyalty story either. The honest question is what happened during that stretch. Someone who kept growing in place — new responsibilities, new skills, visible initiative — is a very different hire from someone who was simply present. That distinction rarely shows up on a résumé. It shows up when you ask, directly, what they built or learned during the period they didn’t move, and listen for specifics rather than a general sense of busyness.

For leaders: a low attrition number is not, by itself, a result you’ve earned. Before presenting it as evidence of a healthy culture, check it against engagement data, training investment, and internal mobility rates for the same period. If quiet cracking is running high underneath a green retention dashboard, the honest read is that you’re not looking at a stable team — you’re looking at a stalled one, held in place by external conditions rather than by anything your organization actively did. Stalled things, when the pressure holding them still finally lifts, tend to move all at once, and rarely in the direction leadership was hoping for.

Renee still hasn’t sent an application. She still closes the recruiter messages, still tells herself she’s being smart, and by any reasonable financial calculation, she still is. But she’s also started keeping a private document — nothing dramatic, just a running list of what she’s actually built this year, updated every few weeks, in case the math ever changes and she needs to move fast when it does.

That document is the real difference between staying and stalling. An escalator that isn’t moving can still be climbed, one deliberate step at a time, by anyone willing to do the work the machine was supposed to do automatically. Most job huggers aren’t doing that. They’re standing still, waiting for the machine to start again, and calling the wait a strategy. It might be exactly that, for a while. But it’s worth being honest, at least privately, about which one you’re actually doing — because the market that eventually thaws won’t ask which explanation you preferred. It will just ask what you built while you were waiting.

Priya opens the shared doc at 4:50 p.m. on a Thursday. Her colleague finished the competitive analysis an hour ago — clean headers, a tidy comparison table, five confidently stated market-share figures, a closing paragraph that reads like it was written by someone who has thought about this problem for a decade.

She skims it, relieved. Then she notices one figure doesn’t match what she remembers from last quarter’s board deck. She checks. It’s wrong — not wildly wrong, just wrong enough to have been believable. So she checks the next number. Also wrong. By 6:40 p.m., she has rebuilt half the document from scratch, and she still isn’t sure which parts of the other half she can trust.

Nothing about this document looked unfinished. That was the problem. It looked, by every visual and structural signal available to Priya, exactly like good work. Harvard Business Review gave this phenomenon a name in September 2025 — workslop — and the researchers at BetterUp Labs and Stanford’s Social Media Lab who coined it found Priya’s Thursday evening is now startlingly common: 40% of 1,150 U.S. full-time employees surveyed said they’d received AI-generated work like this in the past month, costing them nearly two hours of rework, on average, per instance. 53% said it left them annoyed. 42% said it made them trust the sender less. A third said they’d rather not work with that colleague again.

What’s actually being measured here isn’t an AI capability gap. It’s a mirage, and mirages are interesting precisely because the failure isn’t in what you’re looking at. It’s in your own perceptual machinery.

Where the mirage comes from

Long before anyone trained a transformer, psychologists had already mapped the terrain AI would eventually colonise. In the 1990s, researchers including Norbert Schwarz and Rolf Reber began documenting what they called processing fluency: the ease with which the brain processes a piece of information becomes, all by itself, a signal the brain uses to judge whether that information is true. Statements printed in a clearer font were rated as more truthful than the identical statements printed in a harder-to-read one. Words that were easier to pronounce were judged more trustworthy. None of this had anything to do with the content. It was entirely about the packaging.

This is not a flaw we can shame ourselves out of. It’s a load-bearing shortcut.

Evaluating every claim on its actual merits, from scratch, every time, would be cognitively unaffordable — so the brain outsources a huge portion of its trust calibration to surface cues: confidence, fluency, structure, specificity.

For most of human history, that shortcut worked reasonably well, because producing fluent, confident, well-structured, specific-sounding content was actually hard. It required expertise, rehearsal, or both. Fluency was, imperfectly but usefully, correlated with competence. Con artists and skilled salespeople have always understood how to break that correlation deliberately, but they were rare, and rarity was itself a kind of natural defence.

Generative AI didn’t invent the fluency heuristic. It did something more consequential: it made fluency free, infinite, and instant, for anyone, on any topic, at any hour, completely severing the correlation the heuristic was built to exploit. The packaging that used to be expensive to produce, and therefore a semi-reliable signal, is now the cheapest part of the entire pipeline. Only the substance remains expensive. And substance is exactly what the fluency heuristic was never designed to check.

A related finding from the same era of research makes the picture sharper still. The illusory truth effect, first documented by Lynn Hasher, David Goldstein, and Thomas Toppino in the late 1970s, showed that people rate a statement as more likely true simply because they’ve encountered it before, independent of whether it was true the first time either. Repetition alone manufactures a feeling of familiarity, and familiarity gets misread as validity. Put fluency and repetition together and you get something close to a complete account of how confident nonsense survives contact with a smart, careful reader: it sounds right, it feels familiar, and both of those sensations arrive before any actual verification has had the chance to begin.

AI-generated text is unusually good at triggering both at once: fluent by construction, and often repeating, in slightly different words, the same plausible-sounding claim across a document, which makes that claim feel more corroborated with every restatement, even though it has exactly one source: the model’s own prior sentence.

The mechanism, made visible

It’s worth being precise about why this happens, because “AI sounds confident” undersells the mechanism. A large language model is not making a judgement about whether it knows something and then choosing how to express that judgement. It is predicting, token by token, the most statistically plausible continuation of a prompt, shaped by training and reinforcement to produce answers that human raters found satisfying. Human raters, being human, are subject to the same fluency heuristic as everyone else — they tend to rate confident, well-structured, elaborate answers more highly than hedged, uncertain, terse ones, regardless of accuracy. The result is a system trained, at a deep level, to optimise for the appearance of competence, because that’s what the training signal actually rewarded.

Researchers have now measured the downstream consequence directly. In a set of controlled experiments, participants were shown AI-generated answers that varied in length and elaborateness. Longer, more detailed explanations reliably increased participants’ confidence in the answer but did not improve their ability to tell correct answers from incorrect ones. Only 26% of participants could accurately perceive when an AI response was overconfident relative to its actual reliability. Verbosity was functioning as a proxy for expertise: length and elaborateness were triggering what researchers describe as an epistemic authority heuristic, one that bypasses genuine comprehension checking almost entirely.

Across these studies, a consistent pattern holds: the more elaborate and confident an AI’s explanation, the more people trust it, and the less that trust tracks whether the explanation is actually correct.

This is the mechanism underneath every piece of workslop Priya has ever received. It’s not that the model is lying with unusual skill. It’s that the model never learned to signal uncertainty the way a careful human expert instinctively does — the pause, the hedge, the “I’d want to double-check this before you use it,” the visible discomfort of a person who knows the difference between confidence and correctness. An LLM has no discomfort to display. It presents a fabricated statistic with exactly the same tonal certainty as a verified one, because tonal certainty was never connected to truth in the first place. It was connected to what got rewarded during training.

What gets lost

The direct cost of the fluency mirage is the two hours Priya spent on Thursday. The indirect cost is larger and slower-moving, and it shows up in three places.

The first is trust between colleagues. The workslop data on this point is unambiguous. Handing someone fluent-but-hollow work doesn’t just cost them time; it changes how they see you. Forty-two percent of workslop recipients said it made them view the sender as less trustworthy; roughly half rated the sender as less capable, less creative, or less reliable than before. AI is being used, by well-intentioned people trying to move faster, in ways that are quietly corroding the interpersonal trust that actual collaboration depends on.

The second is what happens to judgement itself when it goes unpracticed. Verification is a skill, and like any skill, it atrophies without use. If the default posture toward AI output becomes acceptance rather than scrutiny — because scrutiny is effortful and the output looks so finished — then the muscle required to catch the next, more consequential error weakens exactly when the stakes are rising.

This is not hypothetical. MIT Sloan Management Review’s 2026 research on AI verification makes the trajectory explicit: as AI systems are given longer, higher-risk, more autonomous tasks, checking whether the work was actually done correctly becomes harder and slower — not easier — even as organisational reliance on that work accelerates. The gap between how much we trust AI output and how well we can verify it is widening, not narrowing, at precisely the moment the stakes attached to that output are growing. A parallel 2026 global executive survey found 86% of leadership teams now consider AI central to strategic priority-setting — meaning the quality of human oversight has quietly become a strategic variable, not just an operational nicety.

The third is subtler: the mirage doesn’t just produce bad individual decisions. It changes what “good work” is perceived to look like across an entire organisation, because the visual and structural markers of quality — polish, structure, confident phrasing, comprehensive-seeming coverage — have become disconnected from the underlying rigor those markers used to signal. Teams start optimizing, often unconsciously, for the appearance that gets rewarded rather than the substance that was supposed to produce it. That’s a much harder problem to reverse than a single bad quarter, because it’s not a mistake — it’s a slowly shifting incentive structure that nobody explicitly chose.

Four faces of the mirage

The fluency mirage doesn’t show up as one uniform failure. In practice, it tends to take one of four recognisable shapes.

• The Instant Expert. An AI-generated answer arrives with the tone and structure of someone who has spent a career in the domain — assured phrasing, domain vocabulary used correctly, a confident framing of trade-offs. Nobody asks how the model would know this, because it doesn’t sound like a guess. It sounds like expertise. The tell isn’t in what’s said; it’s in the complete absence of the specific, hard-won caveats a genuine expert would volunteer unprompted.

• The Verbose Alibi. Length gets mistaken for rigour. A three-paragraph answer feels more trustworthy than a three-sentence one, even when the extra length is padding, restatement, or elaboration on an unverified premise. This is the mechanism the verbosity research captured precisely: elaborateness functions as a proxy for diligence, whether or not any additional diligence actually occurred.

• The Certainty Engine. The response never hedges, never flags uncertainty, never says “I’m not fully sure about this specific figure.” Because the model’s tone doesn’t shift between confident-and-correct and confident-and-fabricated, the human reading it loses the single cue — hesitation — that would normally prompt a second look.

• The Downstream Debtor. This is workslop’s actual delivery mechanism. The polished, fluent, unverified output gets passed along — to a colleague, a manager, a client — and the obligation to notice the gap between fluency and substance transfers with it, usually without anyone naming that a transfer has occurred. The person who generated it experiences speed. The person who receives it experiences the bill.

  ---

The compounding problem

All of this was true when the worst an AI system could do was write one wrong paragraph. It matters more now, because the unit of AI-produced work is no longer a paragraph — it’s a multi-step task, executed with growing autonomy across research, drafting, coding, and decision-support, each step building on the fluent-sounding output of the step before it. MIT Sloan’s 2026 research on AI verification names this directly: as systems take on longer, higher-risk, more autonomous tasks, checking whether the work was actually done correctly becomes harder and slower, not easier, because there are more intermediate steps, each one a fresh opportunity for a plausible-sounding error to get quietly folded into the next stage before anyone looks closely. A wrong number in a single answer is a two-hour problem, à la Priya. A wrong assumption embedded at step two of a fourteen-step agentic workflow is a problem nobody notices until it surfaces three departments away, dressed in the confident, finished-looking output of everything built on top of it. The mirage doesn’t stay contained to the document it started in. It propagates, fluently, at every layer it touches — and each layer adds its own varnish of polish, making the original error progressively harder to spot the further downstream it travels.

This is also why the comfortable assumption — that organisations will simply get better at catching this over time, the way they got better at spotting spam email or phishing attempts — doesn’t hold up well under scrutiny. Phishing emails get caught because their fluency is usually imperfect: a slightly wrong logo, an odd turn of phrase, a domain that’s almost right. The fluency mirage has no equivalent tell. It is, definitionally, the case where the packaging is flawless. There is no typo to notice, no broken image, no uncanny phrasing — just a confident, well-structured, entirely plausible claim that happens to be wrong, sitting in a document that looks exactly like every other document that was actually right.

What actually holds up

None of this is an argument against using AI for real work — it’s an argument for pricing verification back into how that work gets planned, evaluated, and rewarded, deliberately, rather than assuming it will happen for free because it always used to.

• For individuals, that starts with treating fluency as a neutral signal rather than a positive one. A polished, confident answer earns exactly zero epistemic credit until it’s been checked; the smoother it is, the more consciously worth checking it becomes, precisely because smoothness is the cheapest thing about it now. One useful ritual: before passing along any AI-assisted output, name out loud — even just to yourself — the single number, claim, or assumption in it you are least sure about, and check that one thing before you check anything else. It’s a small habit, but it directly counters the illusory-truth trap, because it forces a specific claim to be evaluated on its own merits rather than absorbed into the general glow of a well-formatted document. Building this instinct is uncomfortable early on — it means being the person who says “let me verify that” in a meeting full of people nodding along, which can feel like friction for its own sake. It isn’t. It’s the only mechanism in the room actually distinguishing fluent from true.

• For hiring managers, this reframes what a strong candidate looks like. The ability to produce fluent, well-structured work is no longer a differentiator — it’s table stakes, available to anyone with an API key. What’s scarce, and getting scarcer, is the ability to interrogate fluent output and articulate specifically why it might be wrong: to name the missing caveat, spot the plausible-but-unverified number, ask the question the confident tone was designed to make you forget to ask. That capability shows up far more reliably in a structured conversation — “here’s a polished answer, tell me what’s wrong with it” — than anywhere in a résumé, and it’s worth building directly into how candidates are evaluated rather than assuming it will surface on its own.

• For leaders, the implication is structural. If output volume is being measured without measuring downstream verification cost, the organisation is tracking a number that workslop is subsidising — quietly, and usually with the time of the most conscientious people on the team, the ones who can’t bring themselves to pass along something they haven’t checked. Making verification visible — treating it as legitimate, budgeted work rather than an invisible tax someone absorbs after the fact — is not a productivity drag. It’s the correction that lets the productivity number mean anything at all. That might mean building a verification checkpoint into a workflow the same way a code review checkpoint already exists; it might mean explicitly rewarding the person who caught the error rather than only the person who moved fast. Either way, the goal is the same: make the invisible cost visible before it becomes someone’s Thursday evening.

Priya finishes rebuilding the document a little after 7. Nobody will ever see the two hours it took her — not in the deliverable, which now looks exactly as polished as it did before she started, and not in any metric her organisation tracks. That invisibility is the entire mechanism of the fluency mirage in miniature: the correction disappears into the same seamless-looking output that necessitated it, and next week, someone hands her something else that looks finished.

The uncomfortable truth isn’t that AI produces bad work. It’s that AI produces work indistinguishable, on the surface, from good work, and the surface is the only thing most of us have time to check. Fluency was never supposed to be the whole test. For most of history, it was a reasonable proxy, because faking it convincingly was expensive. That’s no longer true, for anyone, about anything. The proxy broke. What’s left, the only thing left, is the slower, less comfortable, entirely human work of actually checking — and organisations are only beginning to notice that they stopped budgeting for it right around the time they started needing it most.

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!