Amazon Admits Its AI Metrics Were Flawed. It Kept Them Anyway.

This is the fourth time I’ve written about Amazon and AI tooling. The first three went badly for Amazon:

• February: Kiro decided to “delete and recreate” a production environment. Amazon called it user error (Delete and Recreate).
• March: Amazon.com went down for six hours. About 1,500 engineers wanted Claude Code instead of Kiro (Amazon’s AI Outages Escalated).
• June: a leaderboard ranked engineers by tokens. Staff gamed it until it died (Measure Token Usage, Get Token Usage).

This week Amazon published the memo that walks it back. It is a good memo. It is also a careful one.

What the Memo Says

It is by Terk Terkowitz, a VP and Distinguished Engineer in Amazon Stores. It went to engineers in September and went public on October 27.

If your work does not benefit from current AI tooling, that is a finding, not a failing.

— Amazon engineering leadership memo, Oct 2026

• The role is not the tool. “Neither the developer role nor their output is defined by the use of AI.”
• The metrics are bad. “We are all aware of how deeply flawed these metrics are, primarily because they are disconnected from what matters.”
• Tokenmaxxing costs money. Used as targets, they drive “increasing costs (tokenmaxxing is at odds with frugality).” Lines of code are “at odds with invent and simplify.”
• Leaders must use the tools. They should describe “a task where they helped and a task where they did not.”

That is the right answer. Most of it is what engineers said in February.

What It Leaves Out

• Kiro. The tool Amazon mandated at 80% weekly use does not appear once. No one says the target is gone. No one says it stays.
• The outages. No mention of either one, or of review standards for agent-written code.
• The metrics themselves. They stay. The memo says using them “with care and judgement” for careers and org performance “is appropriate.” Leaders should watch the 10th and 90th percentiles of AI adoption.

Fortune put it plainly: Amazon is “far from abandoning its AI adoption metrics altogether.” A spokesperson told CNBC there are no caps on token usage.

So the target is gone. The meter stays.

The Better Line Came From AWS

Marc Brooker, a VP and distinguished engineer at AWS, gave CNBC a cleaner version. Amazon looks at “feature velocity”: launches and changes that succeed in production.

What we don’t want to see is wasted effort where we are going and spending those tokens solving problems that don’t matter.

— Marc Brooker, AWS, to CNBC

That is an outcome metric. It is what I argued for in June. If Amazon measured that and dropped the adoption percentiles, there would be no post here.

The Engineers Saw It First

The reaction on Hacker News was tired, not angry.

Tokenmaxxing for the win! Where tokens == Lines of code. Yeah, that was proven to be complete and total garbage over a decade ago.

— bradknowles, Hacker News (self-described ex-AWS)

Another commenter showed why Amazon’s memos matter outside Amazon: “At the company where I work as a SWE, management has set a new rule: each developer is now expected to deliver 50% more code because I am using AI.”

Amazon sets the pattern. Other companies copy the pattern, often the bad version.

A Correction

In March I reported a new senior sign-off rule for AI-assisted code. Amazon told Fortune that was “not accurate.”

The Lesson

• Mandates measure obedience. An 80% usage target tells you who complied. It does not tell you who shipped.
• Walk-backs should name the thing. A memo that never says “Kiro” leaves every engineer guessing whether the mandate still stands.
• Measure outcomes. Brooker had it: what shipped, and did it work.
• The job did not change. The memo says the work is outcomes, ownership and the customer. That is the professional part. The tool was never the job.

Amazon got to the right memo eventually. It took two outages, a petition and a dead leaderboard. The next company does not have to pay that much to learn it.