Amazon engineering leadership published updated guidance on October 27, 2026, asking software development engineers (SDEs) and software development managers (SDMs) to judge work by long-term outcomes for customers, the business and teams. The guidance says token use and other work metrics can help when interpreted carefully, but become flawed when treated as goals disconnected from results. Read Amazon’s published guidance.
How Amazon says to use AI productivity metrics
Amazon names AI adoption and token use as input measures, and code changes and code reviews as measures of engineering throughput. The guideline says these measures currently correlate with important outputs, so managers may use them with care and judgment. It warns that turning them into targets can increase costs and complexity, and that none should stand in for long-term outcomes.
The guidance is not a call to stop using AI. Amazon expects AI to be central to software engineering and encourages engineers to learn and apply it. It also recognizes that current AI tools may not help every task. When they do not benefit someone’s work, Amazon asks engineers to document and explain why.
What managers should examine at the 10th and 90th percentiles
Amazon recommends examining teams near both ends of the AI-adoption distribution. The 10th and 90th percentiles are points for managerial review, not reported adoption rates for Amazon teams.
| Adoption review point | What managers should investigate | Purpose |
| 10th percentile | Why teams may not be adopting AI effectively | Understand what is limiting useful adoption |
| 90th percentile | Whether heavy AI use corresponds to delivery increases | Assess whether greater use aligns with improved delivery |
At the lower end, the question is why adoption is not effective. At the higher end, managers should check whether delivery rises alongside heavy use. Neither checkpoint, by itself, establishes that AI caused a change in results.
What Amazon expects from engineering leaders
Amazon asks technical and organizational leaders to use AI tools hands-on in production work. They should be able to explain a task where the tools helped and one where they did not. That expectation puts leaders close enough to the work to discuss where AI is useful, while leaving room for teams to explain when a tool does not suit a task.
