Amazon tells engineers to stop ‘tokenmaxxing’ to look productive

Amazon wants its engineers to quit chasing token counts, warning that heavy AI usage does not automatically mean better work. In updated engineering guidelines published on October 27, the company laid out how developers and managers should gauge AI’s real impact — and the takeaway is blunt: consuming more tokens is not the same as delivering more value. Amazon is not backing away from the technology, though. “We are certain that whatever future best practices are, AI will be at the center of software engineering practice,” the company wrote.

The guidance targets ‘tokenmaxxing,’ a practice where employees deliberately inflate their AI usage to look productive on internal dashboards. According to the memo, metrics such as AI adoption rates, token consumption, code changes, and code reviews become ‘deeply flawed’ when they are disconnected from actual results. Turning them into targets, Amazon says, encourages wasteful behavior that clashes with its emphasis on ‘frugality and simplicity’ and can raise both compute bills and software complexity.

The company’s preferred yardstick is long-term results. “We believe that it is best to measure, and be measured by, long-term outcomes,” Amazon wrote. That does not mean managers should ignore usage data altogether, as input metrics remain useful when examined ‘with care and judgment’ alongside other indicators. Amazon suggests looking at both the teams using AI the least and those using it the most to determine whether adoption is translating into better delivery.

The push comes as Amazon, under CEO Andy Jassy, has been weaving AI deeply into its sprawling operations. Jassy said earlier this year that the technology will ‘change the scope and speed’ of everyday work, and he has suggested that automation could shrink the company’s corporate workforce over the next few years — though he also expects it to create ‘other jobs.’

The memo also pushes back on the idea that every task needs an AI tool. If current tooling does not benefit someone’s work, Amazon says, that is ‘a finding, not a failing.’ Engineers who can explain where today’s tools fall short are demonstrating judgment rather than reluctance to adopt AI.

In May, it emerged that Amazon employees were using AI unnecessarily to boost internal usage scores, amid pressure to use the company's tools. Workers reportedly used MeshClaw, an internal AI agent capable of handling tasks such as deploying code and interacting with Slack, to inflate their token counts.

Notably, Amazon is not the only company rethinking how it spends on AI. Uber president and COO Andrew Macdonald said in May that higher token usage has yet to translate into useful consumer features. “That link is not there yet,” he said. The remark stands out because Uber’s CTO previously revealed that the company had burned through its entire 2026 Claude Code budget by April.

Accenture appears to also be wrestling with the same problem on behalf of its clients. Leaked audio obtained by 404 Media in June captures executives acknowledging that token spending is climbing fast as companies move from simple chatbots to agentic workflows, and that nobody can reliably predict what a given task will cost or whether it was worth the money. Accenture’s agentic AI strategy lead described it as ‘not a niche problem,’ and the firm is reportedly building an advisory tool called ‘Token IQ’ to help customers manage the bill. OpenAI CEO Sam Altman has previously acknowledged that AI token costs had become 'a huge issue', as companies faced mounting bills from heavy AI use.

Amazon, however, says it has not imposed token caps on its employees. According to CNBC, however, it shut down an internal leaderboard in May after workers were found gaming it to climb the rankings — an early example of the behavior the new guidelines are meant to discourage.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.