Amazon wants its engineers using AI. It doesn't want them burning through tokens just to look productive.
In an internal memo published publicly Oct. 27, the company warned against “tokenmaxxing,” or excessive consumption of AI tokens — the units of text models process or generate — to boost usage numbers. Engineering leaders cautioned that chasing higher numbers can increase costs without achieving better performance.
Managers now face the challenge of promoting AI adoption without turning usage into a competition.
AI usage metrics still matter — but require judgment
Amazon's revised engineering guidance puts customer and business outcomes ahead of the tools developers use. Long-term results, however, can take time to emerge and may be difficult to connect to individual contributions.
The memo therefore identifies two supporting measures.
- Input metrics track token consumption and adoption of AI coding tools. High usage can indicate experimentation, but it doesn't necessarily establish whether developers are accomplishing more.
- Throughput metrics cover engineering activity, including code changes and reviews. Treating these counts as performance targets can prompt unnecessary work and software complexity.
Leadership considers both indicators deeply flawed when treated as targets. Nevertheless, the company reports a correlation between AI adoption and increased delivery among its most active users, which it calls "frontier teams."
Employees and groups that fail to embrace AI or improve output are falling behind, according to the memo. Amazon says the metrics can inform thinking about careers and organizational performance when used with care and judgment.
Managers told to examine high and low AI adopters
Engineering managers are advised to investigate teams around the 10th percentile of AI adoption to understand why usage remains low. At the 90th percentile, their responsibility is to determine whether intensive use corresponds with increased delivery.
Developers whose work does not benefit from current AI tools are encouraged to document their experiences and discuss them with colleagues and supervisors. Explaining why the technology fails to improve a task demonstrates sound engineering judgment.
Engineering managers are also expected to use AI on production assignments themselves and identify one task where the tools helped and another where they did not. This experience should inform how they evaluate employees' work.
In addition, updated expectations call on managers to give staff time to develop hands-on AI skills. Engineers are asked to strengthen their understanding of the products and customers they support, an area where company leadership has acknowledged some employees have lost touch.
What eWeek found: Token counts can reflect tool design, not just developer activity
Two engineers using the same AI model could incur very different costs for comparable work simply because their coding-agent software operates differently.
In a benchmark comparison published by AWS on Oct. 21, researchers tested 12 AI models across two coding-agent frameworks under matching task and hosting conditions. One model cost approximately $24 per task under the heavier framework, compared with $8 using the lighter alternative, without improving evaluation scores. Additional AI agents consumed more tokens and processing time.
Amazon calls on managers to examine AI activity and engineering output across teams, but AWS's findings expose a potential problem with those comparisons. Higher token counts could reflect software overhead instead of greater individual effort or productivity.
AWS’s benchmark used tasks from one repository and did not assess employee performance. Its results support a narrower conclusion: coding-agent design can change token consumption and costs without a corresponding improvement in evaluated quality. They do not establish how Amazon’s employee metrics work in practice.
For managers, that suggests checking which models and agent configurations employees use before comparing their consumption. Task complexity and review requirements also deserve consideration: a high token count alone cannot explain whether someone is experimenting productively, repeating failed attempts, or running a more resource-intensive tool.
A more useful assessment would pair usage data with evidence about completed work, including quality, review effort, rework, and delivery time. These are proposed checks for interpreting AI activity; AWS’s benchmark did not evaluate employee performance.
For engineers, documenting where AI helps—and where it adds work—can make those discussions more concrete. The broader workplace question is whether employers can reward demonstrated improvements without creating incentives to generate more activity simply because it is easier to count.
Want to learn more AI tips, tricks, and prompting techniques? eWeek readers get free 7-day access to The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work. Browse all lessons →
Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy and learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. Check out all the lessons here →