Amazon's AI memo tells engineers non-use is fine. Its performance process says otherwise.

Amazon circulated an internal memo to software engineers and their managers in September 2026, written by Terk Terkowitz, a VP and Distinguished Engineer on the company's Independent Contributor Technical Career Track. The company published the document publicly this week. It addresses a narrow question: what should happen when an engineer tries an AI coding tool and the tool makes the work worse.

The memo's answer, quoted directly: "If your work does not benefit from current AI tooling, that is a finding, not a failing." The instruction is to document the result and bring it to a manager. That line is the whole policy. It does not appear in Amazon's published performance-review criteria, promotion rubrics, or compensation guidelines, because those documents exist separately and the memo doesn't amend them.

Writer Corey Quinn flagged a gap between that sentence and two others in the same memo. The document calls input metrics, meaning AI token usage and tool-adoption rates, "deeply flawed," then says to use them anyway, carefully, then warns that engineers who don't adopt the tools are falling behind. Three statements, three different instructions. The memo doesn't say which one controls when a manager sits down to write a review.

Nothing in the text enforces the first sentence over the other two. A manager who wants to protect an engineer who skipped a broken AI tool can point to "a finding, not a failing." A manager who wants to document someone for a performance improvement plan can point to "falling behind." Both citations match the memo. The memo gives no tiebreaker, and no separate HR policy fills that gap.

The memo does include one concrete measurement proposal: look at the 90th percentile of AI adoption inside a team, the group burning the most tokens, and check whether its delivery output rose by a comparable amount. That test is specific and falsifiable. It is also, as written, a recommendation inside an internal memo, not a documented input to calibration committees or promotion packets, so there's no mechanism tying a team's result on that test to any individual's review.

The memo also singles out "tokenmaxxing," meaning running up AI usage numbers for their own sake, as incompatible with Amazon's internal frugality principle. Amazon shut down an internal product called KiroRank in May 2026 after employees used it to rank each other by token consumption and then generated junk tasks to climb the leaderboard. Meta disabled a comparable internal tracker, referred to as Claudenomics, the same week, affecting a workforce of roughly 85,000 employees.

The frugality argument applies only inside Amazon's own cost accounting. Amazon Web Services sells Bedrock and Kiro to outside customers on a per-token billing model, so from AWS's revenue side, higher token consumption is the business plan, not a violation of it. The memo's frugality language governs Amazon's internal engineering organization. It says nothing about, and has no jurisdiction over, the commercial terms AWS sells to its own customers.

Amazon also confirmed roughly 250 job cuts this week, separate from the memo and not addressed in it. The memo doesn't connect the cuts to AI adoption rates or performance scores, and Amazon hasn't published a document linking the two. The memo and the layoffs are two separate corporate actions that landed on employees in the same week, read against each other by the people who received both.