In April 2026, The Information reported on a leaderboard that Meta’s employees had built for themselves. It was called Claudeonomics. It ranked the company’s roughly 85,000 workers by how many AI tokens they consumed, published the top 250, and awarded titles to the leaders. Among them were “Token Legend” and “Session Immortal”. In one 30-day window, usage on the board exceeded 60 trillion tokens.
Nobody on that leaderboard was ranked on whether the work was any good. There was no column for errors caught, decisions improved, customers helped or hours genuinely saved. There was only one number, and it measured how much the machine had been asked to do.
A few weeks later, the Financial Times reported that Amazon had removed an internal leaderboard of its own, after employees reportedly used AI for unnecessary tasks to push their scores up. According to that reporting, leadership had said token usage would not be used in performance reviews. The employees did not believe it. One senior vice president, Dave Treadwell, eventually told staff: “Please don’t use AI just for the sake of using AI.”
The industry has a name for this behaviour now. It calls it tokenmaxxing. IBM’s explainer on the subject says its popularity peaked in the spring of 2026 and has since declined, and that is the polite way of saying that several of the world’s best-resourced companies spent months rewarding the wrong thing.
This edition is about why that happened, why it will happen again, and what it reveals about how organisations see their own work. I call it the proxy trap: the moment a number that stands in for value becomes a target, and the organisation starts producing the number instead of the value.
It is not a technology problem. It is the oldest management problem there is, now running at the speed and cost of a data centre.
An old law, a new invoice
The idea behind the trap is more than fifty years old. In 1975, the economist Charles Goodhart made an observation about UK monetary policy: a statistical regularity tends to collapse once pressure is placed on it for control purposes. In 1997, the anthropologist Marilyn Strathern gave the idea the form most people now quote, in a paper about the audit culture spreading through British universities:
When a measure becomes a target, it ceases to be a good measure.
Marilyn Strathern, “Improving ratings”: audit in the British University system, European Review, 1997
It is a short sentence and it carries a whole theory of organisations. A measure is useful because it correlates with something you care about. The moment you attach a reward to the measure, people have a reason to move the number without moving the thing. The correlation that made the number useful is the first casualty.
None of this requires villains. It requires only a sensible person, a visible number and a consequence. What is new about AI is not the law. It is the cost of obeying it. When the target was waiting times, gaming the number was cheap. When the target is tokens, every act of gaming is a metered purchase. The proxy does not merely mislead, it bills.
Consider the scale of the incentives in the air during the spring of 2026. IBM’s account records that in March, Nvidia’s chief executive Jensen Huang said he would be “deeply alarmed” if a developer earning $500,000 did not spend at least $250,000 a year on tokens. A partner at Sequoia Capital told the Wall Street Journal in April that “we all should be tokenmaxxing”. By the end of April, Uber said it had spent its entire 2026 AI budget. Those are reported statements and reported spending, not proof that any of it was wasted. But they describe a climate in which the visible sign of seriousness about AI was how much of it you burned.
In a climate like that, the leaderboard was not an eccentric experiment. It was the natural expression of what leaders were saying out loud.
Why the proxy is so tempting
The real question is whether AI is making the work better. That means defining better, in a role where the output is judgement, writing, analysis or code. It means comparing against a baseline that nobody recorded. It means separating the effect of the tool from the effect of the person, the week and the project. It means waiting.
The proxy answers a different question, whether people are using the thing, and it answers it instantly, for everyone, on a chart. For a leader who has told the board that AI is the priority, a rising line is a relief. It looks like progress. It is legible.
That legibility is the trap. A number that is easy to collect gets reported. A number that gets reported gets discussed. A number that gets discussed gets a target. And a target is where the measure stops being a measure.
The mechanism: how a usage target eats value
It helps to walk through how the damage unfolds, because it is rarely dramatic. It happens in four quiet steps.
Step one: the number is introduced as information. Someone builds a dashboard to see who is using the tools. At this stage it is honest. People use AI where it helps and the number reflects that, roughly.
Step two: the number becomes visible. The dashboard is shared, ranked or discussed in a meeting. Now everyone knows where they stand, and what they believe the number means for them. At Amazon, leaders said it would not count in reviews. Employees reportedly did not take that at face value, which tells you something about how much trust a sentence can carry against an incentive. When a ranking exists, people assume it matters.
Step three: behaviour adapts. The sensible response to a visible ranking is to rise in it. For some people that means using AI more in the places it helps. For others it means using it everywhere, whether it helps or not. Employees at Amazon reportedly ran unnecessary tasks. SUCCESS magazine’s account of the same episode adds a detail that shows how targets stack: according to that report, a company-wide aim required more than 80% of developers to use AI every week. Hit the weekly threshold and the box is ticked. Nobody asks about the week.
Step four: the number detaches. The usage figure keeps climbing, because it is being produced for its own sake. The relationship between the figure and the value it was meant to track has quietly broken, but the dashboard cannot show that. The chart looks wonderful at precisely the moment it has stopped meaning anything.
By step four, the organisation is in a worse position than if it had measured nothing. A company with no data knows it is flying blind. A company with confident, rising, meaningless data believes it can see.
The evidence that we cannot feel it either
You might think that people close to the work would sense when AI is helping and when it is not, and that the right answer is simply to ask them. The best evidence we have suggests this is less reliable than we would like.
In 2025, the research group METR ran a randomised controlled trial with 16 experienced open-source developers working on 246 real issues in repositories they knew well. Each issue was randomly assigned to allow or forbid AI tools. Before starting, the developers forecast that AI would make them 24% faster. The result went the other way: with AI allowed, the issues took 19% longer. Even after the study, the developers still believed AI had sped them up by 20%.
Take care with what this does and does not say. It is a small study of a specific group, on a specific kind of work, using the tools of early 2025. METR itself has noted that those results are dated and has published updated data since. Nobody should read it as a verdict that AI makes developers slower today. Tools improve, and so do the people using them.
But the finding that matters for this argument survives that caution. The gap was not between the tool and the clock. It was between the clock and the people. Skilled professionals, close to their own work, were off by nearly forty percentage points in their sense of what had happened to them.
If the person doing the task cannot reliably feel whether AI helped, then a count of how often they invoked it is not a weaker version of the truth. It is an answer to a different question entirely. Asking people how useful it feels and counting how often they use it are both proxies, and the research gives us a specific reason to distrust the first as well as the second.
What gets lost
The visible cost of the proxy trap is the invoice. The hidden costs are larger.
You lose signal about where AI genuinely works. If everyone is nudged to use AI everywhere, you can no longer see where it earns its keep. The places where it transforms the work and the places where it merely adds noise look identical on the dashboard. The most valuable organisational knowledge about AI, which is where to use it and where not to, is precisely the knowledge a usage target destroys.
You lose honest reporting. Once a number is tied to standing, people stop telling you the truth about it. The employee who thinks a task is better done by hand has a reason to stay quiet, or to run the tool anyway for the score. The sceptics, whose judgement you most need, go silent first.
You lose the craft of restraint. Good professionals know when not to use a tool. A surgeon’s skill includes the operations she declines. A usage target treats every refusal as underperformance. Over time you train the restraint out of your best people.
You lose trust. Amazon’s leaders said the numbers would not count and, reportedly, employees did not believe them. That is a quiet disaster. Every future statement of intent from that leadership now has to compete with an assumption that the real rules are elsewhere.
The governance gap beneath the dashboard
There is a further irony. The organisations most eager to count AI activity are often the least equipped to say who is accountable for it.
Grant Thornton’s 2026 AI Impact Survey asked 950 senior business leaders about AI at their organisations. As reported by SUCCESS, 78% of them said they could not pass an independent AI governance audit. The same survey found that organisations with AI fully integrated into their operations were nearly four times as likely to report AI-driven revenue growth as those still piloting it: 58% against 15%.
Put those findings next to the leaderboards and a picture forms. Value shows up where AI is embedded into how the work is actually done. Activity shows up wherever it is easiest to count. They are not the same place, and a company can pour enormous effort into the second while neglecting the first. A firm that cannot explain who owns its AI decisions, but can tell you to the trillion how many tokens it burned last month, has built an excellent speedometer for a car with no steering.
Five people the leaderboard creates
Every target produces a cast. Here are the five I see most often when organisations start counting AI usage.
The Burner. The Burner has understood the game completely. They run automated jobs, ask for summaries of summaries, and schedule tasks that exist only to be counted. They are often the most capable operators in the building, which is the sad part. Their talent is real and it is being pointed at a number. The Burner is not cheating by their own account. They are doing what they were asked.
The Performer. The Performer does not burn tokens so much as narrate them. They share screenshots of their usage, post about their workflows and attend every demonstration. Their skill is visibility. The work may be good or it may not, but it is always well lit. In an organisation that rewards the look of adoption, the Performer is promoted for the lighting.
The Sceptic. The Sceptic uses AI selectively and says so. They have tried the tool on their work, found it helps in two places and hurts in three, and use it accordingly. On a usage dashboard, the Sceptic is at the bottom. In reality, the Sceptic is the person who knows what the tool is for. Their knowledge is worth more than anyone’s score, and it is the first thing a leaderboard punishes.
The Quiet Craftsperson. Related to the Sceptic, but without the opinions. They just do excellent work, using AI wherever it quietly improves it and nowhere else. They do not talk about it. Their usage is modest, their output is sound, and their line on the chart is flat. When the company reviews the numbers, they look like someone who has not got the message.
The Auditor. The Auditor is the person who asks what the number is for. They are rare and often unwelcome, because the question arrives after the dashboard has been presented to the board. If the Auditor is listened to, the organisation avoids most of what follows. If not, they are the person who later says they warned everyone, and is quietly right.
Look at who each system promotes. In a usage-ranked organisation, the Burner and the Performer rise. The Sceptic, the Quiet Craftsperson and the Auditor, the three people holding real knowledge about what works, drift towards the exit or the margins. The target does not just fail to measure capability. It selects against it.
The uncomfortable truths
A usage target manufactures adoption, it does not measure it. The number goes up because you are paying people to move it. That tells you about your incentives, not your capability.
Saying the number does not count is not enough. People read the structure, not the speech. If a ranking exists and anyone senior can see it, it counts. The only way to make a number not matter is not to publish it as a competition.
The feeling of progress is not evidence. The METR developers believed they were faster while being slower. Leaders watching a rising line believe they are winning for the same reason. In both cases the feeling is sincere and the data is elsewhere.
The cheapest metric is the most expensive decision. Tokens are easy to count, so they get counted. Outcomes are hard to count, so they get deferred. The difficulty is not a reason to settle for the easy number. It is the reason the hard number is worth having.
What to do instead
The fix is not to stop measuring. It is to measure the thing you care about, accept that it will be messier, and keep activity numbers in their place as diagnostics rather than goals.
The Amazon episode offers a hint. India Today’s reporting, relayed by Let’s Data Science, says Amazon moved to a measure it calls normalised deployments, intended to capture useful AI-driven work rather than raw token counts. I cannot verify how well that works, and one company’s replacement metric is not a template. But the direction is right. The measure sits at the far end of the work, where something was delivered, rather than at the near end, where something was typed.
If you are early career
Keep your own record of what AI does for your work. Not how often you used it, but what changed: the draft that took an hour instead of three, the error it caught, the one time it confidently produced nonsense and you caught that. This is your evidence, and it will outlast any dashboard.
If you are ranked on usage, do not game the number, and do not stay silent about it either. Say, calmly and with an example, that usage and value are different things. You are far more persuasive with a story about one task than with an argument about measurement theory.
If you are hiring
Ask candidates where they chose not to use AI, and why. The answer reveals more judgement than any list of tools. A person who can explain a refusal understands the task, the tool and the trade-off.
Treat “I use it for everything” as a prompt for questions, not a credential. And be wary of candidates who can quote their usage but not their results. The same trap that catches organisations catches individuals who have learned to measure themselves the same way.
If you are leading
Replace activity metrics with outcome metrics, even clumsy ones. Quality of output, cycle time, cost per result, customer impact. Keep usage data as a diagnostic: it can tell you where to look, never whether you succeeded.
Never publish a leaderboard for a number you would not want maximised. Before you build any ranking, ask what the most extreme version of the behaviour would look like, and whether you would want it. If the honest answer is “an invoice with too many zeros”, you have your answer.
Say plainly, and early, what will not be used in performance reviews, then keep your word. Trust is the only mechanism that makes a disclaimer work, and it is very cheap to lose.
Finally, fund the governance work before the rollout. If 78% of leaders cannot pass an independent audit of their AI governance, then the next dashboard to build is not a usage chart. It is a map of who owns which AI decision, and who answers when it goes wrong.
The thing under the number
Every proxy is a small act of faith. We trust that the thing we can count tracks the thing we cannot. Most of the time the faith is justified, which is why proxies are useful at all. The danger is not in using them. It is in forgetting that they are faith, and then paying people to prove it.
The organisations that come through this period well will not be the ones that burned the most. They will be the ones that quietly asked, early, what exactly they wanted more of. Better decisions. Fewer errors. Faster answers to customers. Work that a person can look at and be proud of.
None of those appears on a token count, which is the point. The things worth wanting are rarely the things that are easiest to count.
The easiest way to look busy has always been to measure busyness. The hardest part of the AI era may be learning to ask for something harder.



The same trap waits in every app built to keep someone company. Its easiest number is time spent, and its leaderboard is retention.
In counselling we measure the other way round: the work is going well when the person needs you a little less each month. No dashboard I’ve seen has a column for that.