Companies pull back from AI token contests as costs and ROI doubts rise
Employers that tracked AI token use as a productivity signal are rethinking the metric after higher bills and weak links to business results.
By Daniel Okafor · Business Editor
3 min read
Some technology employers are backing away from internal contests that rewarded workers for heavy use of AI tokens, Fortune’s Jeremy Kahn reported. The retreat matters because token counts became a rough management signal for AI adoption, even as companies struggled to connect that spending to product output or companywide returns.
The practice, often called “tokenmaxxing,” treated high token consumption as evidence that employees were using AI agents aggressively. Fortune described tokens as the data units processed by AI models, with one token equal to about a word and a half of English text.
According to Fortune, Meta, Amazon, OpenAI and other companies had formal or informal leaderboards that let engineers and developers compare token use over set periods. The idea was to encourage experimentation with AI agents and identify workers using the tools most actively.
Costs and incentives collided
The metric created behavior that did not necessarily match business goals. The Financial Times reported that some Amazon employees ran AI agents on pointless or unnecessary work to keep their token statistics high after managers began using the figures in performance assessments.
AI tokens also carry real costs, especially when workers use advanced models from companies such as Anthropic and OpenAI. Fortune reported that some employers have responded by limiting access to third-party AI agents or restricting use of the most capable models inside agent tools.
Meta removed an informal token-use leaderboard that employees had created, according to Fortune. The Verge reported that Microsoft ended Claude Code subscriptions for employees in several major product groups.
Uber Chief Operating Officer Andrew Macdonald said on a podcast that the company had used up its full 2026 token budget in the first four months of the year, Fortune reported. The company’s heavy use of Claude Code contributed to that spending, according to Fortune.
Salesforce CEO Marc Benioff has said the company’s Anthropic bill will be about $300 million this year, according to Business Insider. Benioff also said he wanted a “smart router” that could send only the hardest requests to the most expensive models while directing simpler tasks to cheaper systems.
Executives want clearer returns
Macdonald said Uber has had difficulty tying productivity gains for individual workers to broader business results. He said token spending is harder to defend when a company cannot connect it directly to useful features and functionality shipped to users.
Fortune’s Kahn tied the problem to Goodhart’s Law, the idea that a measure loses value when it becomes a target. In this case, token use may show activity, but it does not by itself show whether AI is improving products, speeding delivery or lifting financial performance.
Azeem Azhar, author of the Exponential View newsletter, argued that companies may be facing a familiar pattern with general-purpose technologies. He described an AI productivity “J-curve,” in which firms spend time and money experimenting before gains show up in results.
Azhar compared the current period to the early adoption of electricity in factories. He wrote that manufacturers first used electricity for lighting and then to replace steam engines, but major productivity gains arrived only after companies redesigned factory layouts and powered individual machines.
Fortune’s Kahn wrote that many companies remain in the early stages of that process with AI. Counting tokens is easier than redesigning workflows, he argued, while the larger gains may require companies to rethink products, services and business models.
This story draws on original reporting from Fortune.