### What the memo says, in its own words
The central passage is short enough to quote fairly. Amazon's engineering leadership writes that teams have sometimes measured AI work with input metrics, token usage and tool adoption, or with throughput metrics, code changes and reviews. "We are all aware of how deeply flawed these metrics are, primarily because they are disconnected from what matters." Used as a target instead of a measure, the memo says, they push costs up and code complexity with them. Tokenmaxxing, its one use of the term, is described as being at odds with frugality. Counting lines of code gets the same treatment.
The replacement standard is stated plainly: "We believe that it is best to measure, and be measured by, long-term outcomes." The memo concedes those outcomes are delayed and hard to isolate, which is why the input metrics existed in the first place. And it gives engineers explicit cover for the answer most adoption programs punish: "If your work does not benefit from current AI tooling, that is a finding, not a failing."
### The second half of the memo
Coverage that stops at "deeply flawed" misses what Amazon actually decided. The memo says the flawed metrics correlate well with the outputs that matter, and that using them "with care and judgement" is appropriate. It tells leaders to study both ends of the distribution: the 10th percentile, teams barely adopting AI, to understand why; and the 90th percentile, teams using it heavily, to check whether delivery is rising to match. That is a demotion from target to diagnostic. It is not a removal.
Nor is it a retreat from AI. The same document says AI will be at the center of software engineering practice whatever the eventual best practices are, and tells engineers who have not started to start as soon as possible. Its framing device, in the introduction, is the Renaissance Developer: an engineer defined by judgment and domain knowledge, with AI as an input rather than the definition of the output.
### What happened before the memo went public
Two facts around the memo come from CNBC's reporting, not from Amazon's page, and they are the facts that give the memo its weight. CNBC reported that an Amazon spokesperson said the company had not instituted caps on employee token usage. That does not mean usage is unlimited in every setting, or that Amazon has stopped monitoring the costs. Second, in May the company shut down an internal leaderboard after employees burned tokens to improve their rankings. The behavior the memo warns against had already happened inside the company writing the warning.
CNBC also spoke to Marc Brooker, an AWS vice president and distinguished engineer, who described the alternative measure Amazon watches: feature velocity, meaning launches and successful changes reaching production, framed as return on investment. Even there, he allowed room for experiments whose value is not yet clear. The memo and the interview point the same way. The distinction is between spending that supports useful work or experimentation and spending driven by a usage ranking. The memo warns that turning token consumption into a target can reward the latter without improving the former.
### The practical read
An AI adoption dashboard becomes a version of Amazon's abandoned leaderboard when usage numbers are turned into a competition or a proxy for employee performance. Used diagnostically, the same numbers can help managers investigate whether adoption is producing useful results. The memo's usable instruction is narrow: keep the usage numbers, stop presenting them as achievement, and pair them with a delivery measure that is harder to game. A team whose token use doubles without a corresponding improvement in delivery deserves closer examination, not an automatic success label. Amazon's guidance similarly asks leaders to investigate heavy users and check whether their results justify the activity.
Sources
- [1] Amazon, "An update from Amazon engineering leadership on the role of the developer in the age of AI," October 27, 2026Read source
- [2] CNBC, Annie Palmer, Amazon memo on AI usage guidelines and tokenmaxxing, October 27, 2026 (read via AI Weekly)Read source
- [3] Superpower Daily, "Amazon warns engineers not to confuse heavier AI use with better work," October 27, 2026Read source