Amazon Employees Say AI Usage Targets Drive Unnecessary Token Consumption

Amazon rolled out its internal AI agent builder, MeshClaw, to help workers automate repetitive tasks. Earlier this year the company began tracking how many tokens—units of data processed by the AI model—each team consumes, displaying the numbers on internal dashboards that some describe as leaderboards. Over 80% of developers received a weekly target for AI use, intended to encourage adoption of the technology.

One employee told reporters that managers monitor the token data, which creates perverse incentives and fuels competition among staff. Amazon set a weekly AI usage target for more than 80% of its developers. The company started tracking AI token consumption earlier this year using internal leaderboards or dashboards.

The tracking system, while framed as a cost‑efficiency measure, is being interpreted by workers as a performance signal. As a result, some employees report using MeshClaw to automate extra, non‑essential tasks solely to raise their token totals. This behavior mirrors patterns seen in other workplaces where quantitative targets unintentionally encourage gaming the system. Amazon maintains that the data are not used in performance reviews and that the goal is to understand cost and efficiency, not to rank individuals. Some workers say the pressure feels like a covert productivity metric, even though leadership insists the numbers are only for internal analytics. This disconnect has sparked discussions in internal forums about transparency and the proper role of AI targets in engineering workflows. Meanwhile, external observers note that similar token‑based incentives have appeared at other tech firms experimenting with generative AI tools.