Token Use per Employee: What the Count Tells You and What It Does Not

In April 2026, Fortune reported that a Meta employee had built a leaderboard called “Claudeonomics” that showed the top 250 AI token users among the company’s more than 85,000 employees. A token is the unit AI vendors bill by, roughly a piece of a word. According to the article, which relayed reporting by The Information, total use on the dashboard passed 60 trillion tokens in 30 days, and the highest-ranked user averaged 281 billion. Meta told Fortune that the employee took the dashboard down “at their discretion” and that “Meta did not request this action.”

Similar stories followed about Amazon, Uber and Microsoft, and the habit of ranking people by token use acquired a name, “tokenmaxxing.” By October, Amazon’s own engineering leadership had written that such measures are “deeply flawed.” This article sets out what was reported and who said it, then what a token count can and cannot tell a CFO or CHRO.

What was reported in 2026, and by whom?

The sourcing matters when these stories are repeated in a board meeting. The Amazon leaderboard account rests on Financial Times reporting that we read only as relayed by Fortune. That Fortune article says Amazon reportedly told staff that tokenmaxxing would not be a factor in performance reviews, and that several employees told the Financial Times they worried managers watched the numbers anyway. Neither point is an on-record Amazon statement.

Not every executive argued against heavy use. The same Fortune article relays remarks by Meta’s chief technology officer, Andrew Bosworth, to Forbes: his best engineer was spending about as much on tokens as the engineer’s salary and was “5x to 10x more productive.” That is one executive’s description of one person. It was not presented as a measurement.

What does a token count tell you?

A token count answers two questions reliably. The first is cost. Vendors bill by the token, so volume multiplied by the price of the model used gives the spend. Fortune’s April article did this arithmetic for the top Meta user and arrived at more than $1.4 million at a rate of $5 per million tokens. That figure is the reporter’s estimate, and Meta did not supply it.

Many companies still lack this view. In the Deloitte survey of 3,500 business and IT leaders, 71% said they track total AI inference and API spend, 51% track token volume and 44% track per-unit inference cost. Costi Perricos of Deloitte said companies are getting better at watching the total bill, but “far fewer understand the unit cost underneath it.”

The second question is adoption: who has started using the tools and who has not. Ankit Jain, founder of The Hangar, told LeadDev that the mistake was “letting an onboarding metric be used for performance-management.” Amazon’s memo takes a similar line. It tells leaders to look at the teams with the lowest AI adoption and understand why, and to check whether the teams with the heaviest use are seeing delivery increases to match. The memo also says these input measures correlate with useful outputs at present, so Amazon has not discarded them.

What does the count leave out?

It leaves out the work. A token count is the same whether the request drafted a contract clause or kept an idle agent running. Once people know the count is being watched, some of them raise it. Gil Luria, an analyst at D.A. Davidson, told Fortune, “You get the behavior that you create the incentive for.” Fortune’s AI editor, Jeremy Kahn, described the pattern with Goodhart’s Law, the observation that a measure stops being a good measure once it becomes a target.

A count also says nothing about whether the model was the right size for the task. The same count can carry very different costs, because the price per token differs from one model to the next. Salesforce chief executive Marc Benioff told Business Insider that he wished for a “smart router” to send simpler queries to cheaper models, according to the same Fortune piece.

Part of the company is missing from the count as well, because vendors report usage in different units. Anthropic’s Enterprise Analytics API returns per-user token usage and cost for usage-based Enterprise plans. Microsoft’s usage report for Microsoft 365 Copilot returns the last date each user was active in each app and has no token field. A token leaderboard therefore covers only the tools that expose tokens.

The largest gap is the result. Research shows that the effect of AI on output varies widely by person and task, and that people misjudge it. In a randomized trial by METR published in July 2025, 16 experienced open-source developers took 19% longer on their tasks when AI tools were allowed, having expected a 24% speedup. METR has since said those results are out of date and, in a February 2026 update, called its newer data “only very weak evidence” of how much has changed. In customer support, a study of 5,179 agents found that an AI assistant raised issues resolved per hour by 14% on average, with a 34% gain for novice and low-skilled workers and minimal effect for experienced ones. In both studies the finding came from measuring output directly, which a usage count cannot do.

What should sit beside the count?

Two things belong beside it, both at the level of a role. The first is a description of the role’s work, task by task, with a view of how much of it current AI can do. This is what tells you whether high token use in a team is expected or odd. Heavy use in a role where few tasks suit AI is a reason to question the spend. Light use in a role where many tasks suit AI may point to training and rollout that has not happened yet. Amazon’s memo makes the first point from the employee’s side. If a person’s work does not benefit from current AI tooling, it says, “that is a finding, not a failing.” JobRoute publishes how it scores AI exposure by task, and other public task data can serve the same purpose.

The second is the role’s own output measure, read from the system where the work happens: issues resolved for a support team, for example, or days to close for a finance team. Where token use is high and the work suits AI, this is the check that output actually moved.

Several products already measure AI spend or usage, and a buyer should look at them. Larridin traces license and token costs to teams and agents. Rippling’s AI Spend Console ties AI spend to employee attributes such as department and role. Microsoft’s Copilot Business Impact report relates Copilot usage to outcome data the customer uploads, and notes that the differences it shows have not been tested for statistical significance. We read these pages on November 11, 2026. None of the three pages describes a task-by-task baseline of what AI can do in each role, though a feature missing from a public page may still exist in the product.

Why measure roles and teams instead of individuals?

A token count is not a performance measure, and the leaderboard accounts relayed by Fortune show what can happen when it is treated as one. Reading it by role and team gives no individual a personal number to raise, and it matches the question a CFO is asking, which is where the spend belongs.

There is a legal reason as well. The EU AI Act lists AI systems used to monitor and evaluate the performance and behaviour of workers as high-risk in Annex III, point 4(b), and Regulation (EU) 2026/1744 sets December 2, 2027 as the date those rules apply. Whether a given usage dashboard falls under that definition depends on how it is built and used. Our 2027 compliance calendar lists the related United States rules. This is general information and not legal advice.

JobRoute joins the three for each role, token use, the role’s tasks and the role’s output measure, on its AI spend by role page.