🔍 Read the full analysis: From Tokenmaxxing To Outcomes: The AI Measurement Question Amazon Is Raising on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Amazon published revised software-engineering guidance on October 27, warning that token counts and other AI-use metrics can become costly, misleading targets. The memo says those measures may still help identify patterns, but engineers should be judged by outcomes and use of professional judgment.
Amazon’s engineering leadership published revised guidance on October 27 telling software teams not to treat token counts or other AI-use figures as performance goals. The memo says such metrics can encourage waste and complexity, while arguing that engineers should be assessed on customer, business and team outcomes—a message that arrives as companies face pressure to show what their AI investments deliver.
The memo was published on Amazon’s own news site and was written by Terk Terkowitz, a vice president and distinguished engineer in the company’s stores business. It introduces a revised role guideline for software engineers, drafted by senior engineers. Rather than prescribe specific AI tools or rules, it centers the role on impact. The memo says a detailed definition of AI use could become outdated quickly.
Amazon calls token use, AI-tool adoption, code changes and reviews “deeply flawed metrics” when used as targets. These figures can be useful as indicators, the memo says, and currently correlate well with outputs that matter. But when teams reward them directly, usage can rise without a corresponding benefit. It cites tokenmaxxing—using an AI tool for unnecessary work to increase token counts—as contrary to Amazon’s frugality principle, and warns that targeting lines of code conflicts with simplifying systems.
The guidance asks leaders to examine both low and high adoption. They should investigate why some teams use AI less, and check whether heavy users are producing better results. The memo also asks leaders to use AI tools on their own production work and record cases where the tools helped and did not. Engineers who find that current AI tools do not help a task should document that finding, rather than be treated as failing to adopt.
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
From Usage Counts to Work Outcomes
Amazon’s guidance addresses a measurement problem facing employers adopting generative AI: activity is easier to count than value. Tokens consumed, tools opened and code produced are visible, but do not by themselves show whether a customer received a better service, a task took less time or a system became easier to maintain. The memo’s distinction matters to engineers whose performance could otherwise be tied to metrics they can increase without improving the work.
The company is not saying AI adoption has no value. Its memo says adoption metrics currently correlate well with meaningful outputs and describes teams using AI more as delivering more. But it does not provide figures, define those outputs or establish that AI use caused the difference. The practical message is to use usage data to locate questions for investigation, then assess the work itself. That approach may limit incentives to run up costs merely to rank higher.
The policy also reflects a wider question for investors and customers: what counts as a successful AI investment? A company can report adoption or model activity without showing business returns. Amazon’s own AI portfolio offers a high-profile backdrop, though the source material does not establish that the engineering memo was caused by the performance of any particular model or investment.
Products Worth Considering
Jackery Solar Generator 1000 v2 with 200W Solar Panel
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
As an affiliate, we earn on qualifying purchases.
Amazon’s Earlier AI Usage Push
The memo follows reports of employee tracking and informal competition around AI use. In May, the Financial Times reported that Amazon staff gave an internal agent tool called MeshClaw unnecessary tasks to increase token counts. That account, as summarized by The Decoder, also described a weekly AI-adoption target and internal usage leaderboards. Those details are reported accounts, not confirmed facts in the supplied source.
Amazon did confirm to Business Insider at the end of May that a leaderboard called KiroRank had been deprecated. The company described it as an informal tracker created by a group of employees and said it was not intended to promote AI use for its own sake. The source also quotes senior vice president Dave Treadwell urging staff not to use AI merely for the sake of using it. The October memo turns that idea into formal engineering guidance.
The source places the memo alongside two company-level developments. Business Insider reported in July that most flagship Nova models had been moved to a “keep the lights on” status; AWS lifecycle information lists Nova Canvas, Nova Reel and the first Nova Sonic as retired. Separately, Amazon’s June 30 filing valued its Anthropic holdings at about $190 billion, compared with roughly $18 billion invested, according to the source. Those figures describe reported and filed positions; they do not by themselves show operating returns. Amazon was scheduled to report third-quarter results on October 29, 2026.
““Tokenmaxxing is at odds with frugality.””
— Amazon engineering memo
What the Memo Does Not Quantify
The memo does not set out a single replacement formula for evaluating AI-assisted work. It says long-term outcomes are slow and difficult to measure, and asks leaders to apply judgment, but the supplied material does not specify how teams should weigh customer impact, quality, speed, costs or maintainability against one another.
It also offers no quantified evidence that AI adoption caused the reported gains among high-adoption teams. The memo says usage measures correlate with outputs, not that every increase in usage improves performance. The reported MeshClaw behavior and May adoption target rely on media accounts, while Amazon’s confirmation about KiroRank is narrower: the company said the leaderboard was informal and had been discontinued.
The source does not provide Amazon’s October 29 quarterly results, which were still forthcoming when it was written. It also does not establish the final financial value of the Anthropic investment or explain the precise status of every Nova model beyond the cited reporting and AWS lifecycle listings.
Quarterly Results and Team Reviews
Amazon’s next scheduled public milestone in the source material is its third-quarter earnings report on October 29. Those results may offer financial context on the company’s broader business and AI-related investments, but they will not necessarily answer how engineering teams measure the outcomes described in the memo.
Within engineering, the guidance points to ongoing reviews of low- and high-adoption teams, with leaders testing tools on real production work and recording where they help. The memo does not announce a new leaderboard, numerical target or timetable for publishing results. How Amazon turns its outcome-focused guidance into consistent evaluations remains open.
Key Questions
What did Amazon announce?
Amazon published revised software-engineering guidance on October 27, 2026. It says token counts and other AI-use figures should not be treated as performance targets, and directs attention to customer, business and team outcomes.
Is Amazon telling employees to stop using AI?
No. The memo does not reject AI adoption. It says usage measures can help identify patterns, while warning that rewarding usage for its own sake can increase costs and complexity.
What does “tokenmaxxing” mean here?
It refers to using an AI tool for unnecessary tasks to raise token counts or improve a usage ranking. Amazon’s memo says that practice conflicts with its frugality principle.
What remains unclear about the new guidance?
The memo does not provide a numerical performance formula or quantified evidence that AI use caused higher output. It asks leaders to apply judgment and examine whether adoption is producing useful results.
When is Amazon’s next scheduled update?
The source says Amazon was scheduled to report third-quarter results on October 29, 2026, two days after the memo’s publication.
Source: ThorstenMeyerAI.com
