🔍 Read the full analysis: Amazon’s Message On AI Metrics: Look Beyond The Tokens on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Amazon published revised software-engineer guidance on October 27, warning that token counts and other AI-use metrics can create perverse incentives when used as targets. The memo says those measures may help identify patterns, but engineers should be evaluated on outcomes and judgment, not how much AI they use.
Amazon published new guidance for all software engineers on October 27, warning that token counts and other measures of AI use can encourage waste when they become performance targets. The memo says teams should focus on customer, business and team outcomes, a policy shift that follows reports of employees using an internal AI tool for unnecessary tasks to raise leaderboard rankings.
The memo was written by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business, and presents a revised role guideline developed by senior engineers. It says the guidance avoids prescribing fixed AI rules because those rules could quickly become outdated. Instead, it centers the engineering role on impact and judgment.
Amazon identifies token usage, AI-tool adoption, code changes and reviews as input measures that can be useful but are “deeply flawed” when treated as goals. The memo argues that pursuing them as targets can raise costs and add complexity. It uses the company’s leadership principles to frame the concern: “tokenmaxxing is at odds with frugality” and measuring lines of code is at odds with “invent and simplify.”
The memo does not say that AI adoption is unimportant. It says input metrics currently “correlate well” with results that matter, and that teams adopting AI most are delivering more, while teams not adopting it are “falling behind.” Its distinction is between tracking usage as a measure and rewarding it as an end in itself. The guidance asks leaders to examine both lower and higher adoption levels, investigate why teams differ, and use AI themselves on production work.
Six months at Amazon
From token leaderboards to an outcomes memo
Amazon spent 2026 learning the same lesson twice: once with how its engineers use AI, once with its own AI models.
- May 2026Token leaderboards backfireFinancial Times (via The Decoder): staff give an internal agent tool pointless tasks to inflate token counts. Reported, anonymous sources.
- Late May 2026KiroRank “deprecated”Amazon spokesperson to Business Insider. SVP Dave Treadwell: “Please don’t use AI just for the sake of using AI.”
- 24 Jul 2026San Francisco AGI Lab site closesConfirmed by Amazon to GeekWire. Frontier model research continues under Pieter Abbeel.
- 28 Jul 2026Most flagship Nova models wound downBusiness Insider, people familiar: Premier, Omni, Reel, Canvas to “keep the lights on”. Amazon: models “remain one of the most important things we’re working on”.
- 14 / 30 Sep 2026Nova Sonic v1, Nova Reel, Nova Canvas reach end of lifeAWS Bedrock model lifecycle page. Requests to them now fail.
- 26 Oct 2026Anthropic files its S-1Amazon notes: $6.2 billion principal, 12% ownership cap that Amazon can waive. Anthropic commits at least $110 billion to Amazon through April 2036.
- 27 Oct 2026The memo: outcomes, not tokenmaxxingUsage and throughput metrics are “deeply flawed” as targets. “Tokenmaxxing is at odds with frugality.”
- 29 Oct 2026Amazon’s third-quarter resultsAnnounced by Amazon. Not yet published when this article went live.
Sources: Amazon engineering memo (27 Oct 2026); Business Insider (29 May and 28 Jul 2026); GeekWire (24 Jul 2026); The Decoder (12 May 2026, summarising the Financial Times); AWS Bedrock model lifecycle (read 28 Oct 2026); Anthropic S-1 (26 Oct 2026).
Powered by Thorsten Meyer AI
A Shift From Usage to Results
The guidance matters because AI-use targets can shape daily work: when employees are rewarded for volume, they may spend time generating more tokens or code without improving the product. That can increase model costs and make software harder to maintain, even if dashboards show rising adoption.
Amazon’s position is not that usage data has no value. Rather, the company is telling managers to use it to find questions worth investigating, not as a direct measure of individual performance. The memo says leaders should look at the 10th percentile of adoption to learn why some teams are not using the tools effectively, and the 90th percentile to check whether heavy use is producing better work.
For engineers, the document also gives room to report that a tool has not helped a particular task. The memo calls that “a finding, not a failing,” and asks employees to explain and share their reasoning. That stance could influence how other employers evaluate AI productivity, though the guidance alone does not establish how Amazon will apply it in performance reviews.
Popular right now
Samsung 42-Inch S90H OLED 4K Smart TV Bundle
Vitamix Propel Series 510 Professional-Grade Blender, 3 Programs, Black
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Leaderboards to Formal Guidance
The memo follows earlier reporting about internal AI-use tracking. In May, the Financial Times reported that some Amazon employees gave the MeshClaw agent unnecessary tasks to increase token counts. That account, summarized by The Decoder, also described a target for more than 80% of developers to use AI weekly and internal token leaderboards. Those details rely on anonymous employee accounts and are reported claims, not independently confirmed in the memo.
Amazon did confirm a separate detail to Business Insider at the end of May: a leaderboard called KiroRank had been deprecated. The company described it as an informal tracker created by employees and said it was never intended to promote AI use for its own sake. Senior vice president Dave Treadwell also told staff, “Please don’t use AI just for the sake of using AI.” The October guidance now puts a related principle into the company’s formal engineering role guideline.
The source article also compares the memo with Amazon’s broader AI investments, but those are separate developments rather than evidence that prompted the guidance. Business Insider reported in July that most flagship Nova models were moved to maintenance, while AWS lifecycle information lists several Nova products as retired. Amazon’s filings and the source article describe a large investment in Anthropic; the memo itself does not connect that investment or Nova’s product status to its measurement policy.
“Tokenmaxxing is at odds with frugality.”
— Amazon engineering memo
How Managers Will Apply the Guidance
The memo sets out principles, but it does not specify how Amazon will measure individual performance, whether existing targets will be removed across every team, or how managers will weigh adoption data against longer-term outcomes. It also acknowledges that outcomes take time to measure, leaving teams to exercise judgment about which short-term signals are useful.
The earlier accounts of employees inflating token counts came from anonymous sources. Amazon’s confirmation that KiroRank was deprecated does not independently verify every detail reported about MeshClaw or a company-wide weekly adoption target. It is also not yet clear whether the new guideline will change the behavior of teams that rely on AI-use dashboards.
The source article says Amazon is due to report third-quarter results on October 29, but those results had not yet been released at the time of the October 28 article. The memo provides no financial figures or evidence that its policy has already changed AI-related costs or productivity.
Amazon’s Results and Policy in Practice
The next near-term company event is Amazon’s third-quarter earnings report, scheduled for October 29. Those results may provide information about the company’s overall performance, but the guidance does not predict what Amazon will report or establish how its AI investments are performing financially.
For the engineering policy, the practical test will be how leaders use the revised guideline: whether teams stop treating usage as a score, examine why adoption differs, and assess whether AI-assisted work improves outcomes. The memo asks leaders to use the tools on production tasks and compare cases where they help with cases where they do not. Amazon has not announced a timetable for evaluating the guideline’s effects.
Key Questions
What did Amazon announce?
Amazon published revised software-engineering role guidance on October 27, 2026. It says engineers should be judged on impact and outcomes rather than token counts or AI usage alone.
Is Amazon telling employees to stop using AI?
No. The memo says AI adoption is associated with results that matter, while warning against using adoption or token volume as a goal in itself. It asks teams to examine when AI helps and when it does not.
What does “tokenmaxxing” mean here?
It refers to maximizing AI token use to improve a usage ranking, even when the extra work is unnecessary. The memo says that behavior conflicts with the company’s frugality principle.
Did Amazon confirm employees inflated token counts?
The source article cites May reporting based on anonymous employee accounts about unnecessary tasks and token leaderboards. Amazon separately confirmed that KiroRank, an informal employee-built leaderboard, had been deprecated; that statement does not verify every reported detail.
What remains unknown about the policy?
Amazon has not detailed how the guidance will change performance reviews, which usage measures teams will continue tracking, or whether the policy will produce measurable changes in costs or productivity.
Source: ThorstenMeyerAI.com
