Amazon workers game AI metrics

Amazon employees have been using an internal AI tool for unnecessary or low-value work to raise their usage numbers, according to a Financial Times report cited by Digital Trends. The report said some workers used a system called MeshClaw to generate extra AI activity because Amazon had set adoption targets and tracked token consumption on internal leaderboards. Amazon told employees the token statistics would not be used in performance evaluations, but workers told the Financial Times that managers still appeared to watch the numbers. The episode offers a clear example of how companies can get distorted signals when they measure tool adoption by raw activity instead of completed work. Amazon has been pushing generative AI across its business, with CEO Andy Jassy telling employees in June 2025 that the company was using AI in “virtually every corner” of Amazon and that the technology would change how work gets done. Amazon has also promoted its own AI products, including Amazon Q for software development and enterprise tasks. (digitaltrends.com) ### What were Amazon employees reportedly doing to raise the numbers? Three Amazon employees told the Financial Times that internal AI usage metrics were likely inflated because some colleagues were running non-essential tasks through MeshClaw, according to follow-on reports that summarized the paper’s findings. MeshClaw is described as an internal tool that lets employees build AI agents that can interact with workplace software, triage emails, trigger code deployments and work with tools such as Slack. (aboutamazon.com) One employee told the Financial Times there was “so much pressure” to use the tools, and another described people using MeshClaw to maximize token usage rather than useful output, according to those reports. The practice has been described as “tokenmaxxing,” a term used in several accounts of the episode. ### Which metric appears to have driven the behavior? (techspot.com) Amazon set a goal for more than 80% of developers to use AI each week, according to the Financial Times reporting summarized by other outlets. The company also began tracking token consumption on internal leaderboards earlier this year, creating a visible count of how much employees were using the tools. (the-decoder.com) That matters because token counts measure activity, not whether work was completed faster or better. The reports do not say Amazon publicly tied those counts to compensation, but employees told the Financial Times that managers appeared to monitor them even after staff were told the data would not be used in reviews. ### What has Amazon said about the tool? (letsdatascience.com) Amazon said MeshClaw has enabled “thousands of Amazonians to automate repetitive tasks each day,” according to a company statement quoted by Retail Gazette. The company described the product as part of its effort to let teams experiment with generative AI in a “safe, secure and responsible” way. (letsdatascience.com) Andy Jassy has made AI adoption a central theme of Amazon’s public messaging. In Amazon’s 2025 shareholder letter, published April 9, 2026, Jassy said AI would alter customer experience and company operations, while a June 17, 2025 employee message said generative AI and agents would change work and eventually reduce the need for some existing roles. (retailgazette.co.uk) ### Why does this matter beyond one internal leaderboard? Amazon’s case shows how a company can get a misleading picture of adoption when the easiest number to collect is usage volume. If employees are rewarded informally for higher invocation counts, the metric can reflect compliance behavior rather than whether the tool improved code quality, shortened cycle times or eliminated manual work. The Financial Times report, as relayed by Digital Trends and others, does not show how widespread the behavior was across Amazon. (aboutamazon.com) But it does identify a specific mechanism: visible token leaderboards, weekly targets and employee concern about how managers would read the numbers. Those ingredients can turn a measurement system into a participation contest instead of a productivity test. ### What should readers watch next? Amazon has not announced a public change to how it measures internal AI adoption, based on the material reviewed here. The next concrete marker will be whether Amazon, Andy Jassy or AWS executives begin talking more specifically about outcomes such as speed, quality or cost savings — rather than weekly usage counts — in future company updates and product disclosures. (aboutamazon.com) (digitaltrends.com)