Amazon employees are automating unnecessary work to hit internal AI usage targets, creating exactly the kind of perverse incentive that metrics-obsessed organizations breed.
The company has set a goal for more than 80 percent of developers to use AI each week and began tracking token consumption—the units of data processed by models—on internal leaderboards. The move is part of a broader push by Silicon Valley to justify massive spending on AI infrastructure. Amazon is expected to spend $200 billion in capital expenditure this year, the vast majority directed toward AI and data center infrastructure.
But the adoption pressure has produced an unintended consequence: employees are using Amazon's internal MeshClaw tool—which automates tasks like email triage, code deployments, and Slack interactions—to create busywork that inflates their token consumption. "There is just so much pressure to use these tools," one Amazon employee told the Financial Times. "Some people are just using MeshClaw to maximize their token usage."
The leaderboard visibility is driving competitive behavior. "Managers are looking at it," said another employee. "When they track usage it creates perverse incentives and some people are very competitive about it." Amazon has told staff that token statistics will not be used in performance evaluations and has limited leaderboard access to employees and their direct managers. But the damage is done: employees have already internalized the message that usage matters, whether or not it's formally tied to reviews.
Meta employees have engaged in identical behavior, gaming their own internal leaderboards through "tokenmaxxing." The pattern suggests this is not an Amazon-specific problem but a systemic issue across companies trying to rapidly embed AI into workflows.
There's also a security dimension. Multiple Amazon employees expressed concern about granting an AI agent broad permissions to act autonomously—initiating deployments, monitoring systems, and accessing applications. "The default security posture terrifies me," one engineer said. "I'm not about to let it go off and just do its own thing." An agent making errors or taking unintended actions could have significant consequences in a production environment.
Amazon's statement framed MeshClaw as "empowering teams" to experiment with AI and automate repetitive tasks. More than three dozen employees worked on the tool. But the real experiment here is whether adoption metrics can coexist with genuine productivity gains—and so far, the data suggests they cannot.