In short
- In a note to engineers, Amazon called AI-usage metrics flawed because, used as targets, they can raise costs and complexity.
- What matters is results for customers, not the number of tokens used.
- If AI does not help a given job, the note says to record it as a finding, not a failing.
In short: Amazon published a note for its software engineers saying that AI-usage metrics, such as the number of tokens, are flawed as targets. The Register described it on 28 October 2026 as an attempt to curb tokenmaxxing.
What the note says
- The engineer’s role does not change: what counts is impact on customers, the business and the team, even though the tools change very fast.
- Input metrics (tokens, adoption of AI tools, volume of code changes) can, when used as targets, raise costs and needless complexity. The note allows them as a signal, used with judgement.
- Leaders should look at the extremes: teams that use AI poorly and those that use it most, and check whether it turns into results.
- Anyone who finds AI does not help their work should record it and discuss it; the note calls this a finding, not a failing.
What this means for you
Whatever your company, you can ask two questions: what did the customer get thanks to AI, and what did it cost? I write more about adopting AI at work in the AI in business hub.
How to read it
This is one company’s internal note, published by the company itself; it does not prove that AI fails. The Register also notes that Microsoft took a similar stance in August. The primary source gives no publication date, so I attribute the date to the press report.