Amazon's anti-tokenmaxxing memo says engineers should be pickier about using AI

Amazon has put a clearer line around how it wants engineers to use AI, and the message is basically: stop using the tools just to prove you’re using them. In an internal memo circulated in September and later posted on Tuesday, the company warned staff against “tokenmaxxing,” the habit of leaning on AI as much as possible with little concern for whether the work actually improves anything.

The memo also updates how Amazon thinks about software engineers and managers now that AI is creeping into everyday development work. The company says this is still early days for AI in software engineering, but it’s already confident about one thing: AI will sit at the center of future practice. That’s a strong claim, even as Amazon admits the best way to work with these tools is still being figured out.

The timing matters. Under CEO Andy Jassy, Amazon has pushed AI across its business, while Jassy has said the technology could change the pace and scope of work and may reduce the corporate workforce in the next few years, even as it creates other jobs. At the same time, companies everywhere are wrestling with rising AI costs and the harder question of whether the tools are actually paying off.

Amazon says metrics such as token usage, adoption rates and code changes are “deeply flawed” ways to judge teams or organizations, because they can drive up costs and complexity. But it is not throwing those measures away entirely. Instead, it says they should be weighed carefully alongside other signals, with the real goal being long-term outcomes. Marc Brooker, a vice president and distinguished engineer at Amazon Web Services, said the company looks at things like feature velocity — how many launches happen and how many changes make it into production.

And Amazon is drawing a line in the other direction too. The memo says engineers should not force AI into tasks where it doesn’t help. If AI isn’t useful for a piece of work, that isn’t a failure, the company said. It’s a judgment call. In May, Amazon even shut down an internal leaderboard after finding employees were tokenmaxxing to climb it, which is a pretty neat summary of why these rules exist in the first place.