How Amazon Engineers Fake AI Usage to Maximize Token KPIs

Amazon’s Seattle headquarters recently introduced a sweeping KPI that requires more than 80% of developers to use AI tools each week, accompanied by an internal "AI Token Consumption" leaderboard. Although the company publicly assures that these token statistics will not affect performance reviews, employees quickly realized the metric could influence promotions and job security.

Faced with this pressure, many engineers adopted a workaround known in the industry as tokenmaxxing —inflating their AI‑tool usage to climb the leaderboard. The core of this workaround is the internal AI product called MeshClaw , originally inspired by the open‑source OpenClaw project. MeshClaw can automatically deploy code, sort emails, and even converse on Slack, making it an ideal tool for generating apparent AI activity without genuine work.

One anonymous Amazon employee explained, “We’re under huge pressure; some colleagues simply run the tool to ‘刷’ (刷 =刷榜, i.e., boost) their token count.” Another current employee warned, “Managers are watching the numbers closely, and the incentive structure is now twisted—people are competing fiercely for token scores.”

The practice quickly spread beyond Amazon. Reports indicate that Meta employees are also engaging in tokenmaxxing to climb their own internal rankings.

Beyond the KPI distortion, tokenmaxxing introduces a significant security risk. To automate token generation, employees grant MeshClaw extensive permissions, allowing it to execute code, access emails, and interact with internal systems autonomously. A developer expressed concern: “Its default security settings genuinely frighten me; I would never let it run unchecked in our systems.”

Recognizing the absurdity of a “token‑only” metric, Amazon has quietly restricted access to the token data, limiting visibility to the individual employee and their direct manager, and officially discouraging the use of these numbers for performance evaluation. However, the practice has already taken root, and the cultural momentum is hard to reverse.

This episode illustrates a broader Silicon Valley anxiety: massive AI and data‑center investments—Amazon expects to spend about $200 billion this year—create pressure to demonstrate immediate returns, often by mandating universal tool adoption. The resulting “tokenmaxxing” phenomenon turns sophisticated AI agents into a new form of meaningless repetitive labor: pretending to work hard with AI.