Amazon Tells Engineers to Judge AI Work by Long-Term Outcomes, Not Token Use

The memo asks software engineers to stop chasing high AI token consumption and to be more deliberate about which tasks they hand to AI tools. It argues that token usage and code generation volume are weak measures of productivity, and it says the company prefers to measure teams by long-term results. The guidance changes the emphasis inside engineering teams. Engineers are pointed toward choices that carry more weight than raw output, such as whether a change holds up after release and whether it solves the problem it was meant to solve. For founders, the memo is a reminder to set metrics that tie AI spending to outcomes such as shipped features, defect rates and maintenance cost. Usage counts alone can reward waste and tell an engineering leader little about quality, and a clear outcome standard makes it easier to decide where agents belong in a workflow.