Amazon staff engage in 'tokenmaxxing' to boost internal AI usage metrics

Amazon employees are increasingly engaging in "tokenmaxxing" by utilising the internal MeshClaw AI tool to automate non-essential tasks and generate excessive token consumption. This behaviour is driven by internal pressure to meet a target requiring more than 80 per cent of developers to use AI weekly, coupled with the tracking of usage on internal leader boards. Despite company assurances that these statistics would not factor into performance evaluations, staff report feeling compelled to boost metrics to demonstrate frequent usage of the technology.

The MeshClaw product, developed by more than three dozen Amazon employees and inspired by the external viral tool OpenClaw, allows users to create AI agents capable of connecting to workplace software to perform actions on their behalf. Internal documents describe the tool as being able to "dream overnight to consolidate what it learned" and monitor deployments while employees are in meetings. Specific agents are reported to initiate code deployments, triage emails, and interact with applications such as Slack, creating a mechanism for staff to artificially inflate usage data.

Silicon Valley groups are broadly pushing to increase generative AI adoption to demonstrate returns on vast infrastructure spending and embed technology into daily work. Amazon is expected to spend $200 billion this year on capital expenditure, the vast majority of which is directed toward AI and data centre infrastructure. In an effort to show that this investment is yielding results, the e-commerce giant has introduced metrics to track adoption, though the intensity of the pressure has led to unintended consequences among the workforce.

While Amazon has recently limited access to team-wide AI usage statistics so that only employees and managers can view them, the perception of surveillance remains a key concern. Several current employees believe managers are actively monitoring the data, creating an environment where token consumption becomes a competitive metric. One staff member noted that when usage is tracked, it creates perverse incentives, leading some individuals to be very competitive about their standing on the leader boards.

The concept of tokenmaxxing refers to the practice of using AI tools to consume tokens, which are units of data processed by models, unnecessarily to boost usage metrics. Meta employees have similarly engaged in this behaviour to improve their standing on internal leader boards, highlighting a wider trend within the sector. Amazon has stated that the tool enabled "thousands of Amazonians to automate repetitive tasks each day" and is an example of empowering teams to experiment, but the current usage patterns suggest the focus has shifted toward metric maximisation rather than genuine productivity gains.

Concerns regarding the security implications of granting AI agents permission to act on a user's behalf have also been raised by staff. One employee expressed fear over the default security posture, noting the risk of agents making errors or undertaking unintended actions. While the company maintains that the data will not be used in performance evaluations, the gap between official policy and on-the-ground reality continues to fuel the behaviour.