🔍 Read the full analysis: Tokenmaxxing At Amazon: Why Outcomes Matter In The AI Metrics Debate on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Amazon published revised guidance for software engineers on October 27, warning that token counts and other AI-usage measures can become harmful targets when they are disconnected from results. The memo still describes those measures as correlated with useful outputs today, but says engineers and managers should judge AI by its impact rather than the volume of use.
Amazon published new guidance for its software engineers on October 27, warning that token counts and other measures of AI use can distort incentives when treated as performance targets. The memo says engineers should be judged by impact on customers, the business and their teams, not by how many tokens they consume or how often they use AI tools.
The memo was written by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business, and sets out a revised role guideline for software engineers. It says measures such as token usage, AI-tool adoption, code changes and reviews are “deeply flawed metrics” when used as targets rather than as indicators. According to the memo, pursuing those numbers can increase costs and add complexity without showing whether the work is useful.
The guidance does not say that AI adoption is unimportant. It says these input measures currently “correlate well” with outputs that matter, and describes teams that have adopted AI more extensively as delivering more. Its distinction is between using data to spot patterns and using it to score individuals or teams. The memo advises leaders to examine both the 10th percentile of adoption, to learn why some teams use the tools less, and the 90th percentile, to see whether heavy users are producing better results.
It also tells engineers that work which does not benefit from current AI tools is “a finding, not a failing”: they should explain why and share that information with peers and managers. Leaders are asked to use the tools on their own production work, including a task where the tools helped and one where they did not, and explain the difference. The guideline avoids fixed AI-use rules, saying definitions could quickly become outdated.
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 AI Activity to Measured Results
The guidance addresses a growing workplace question: whether companies can use activity data as a reliable stand-in for productivity. Token volume is easy to count; improvements in customer experience, software quality or delivery time can be slower and harder to attribute. Amazon’s memo argues that turning a convenient proxy into a goal risks rewarding the proxy rather than the result.
That distinction matters to workers because a usage target can encourage unnecessary activity, while to employers it can raise AI costs and complicate systems without proving that customers or the business benefit. Amazon’s direction is not to abandon measurement or AI adoption, but to treat usage statistics as diagnostic signals and test them against actual work outcomes.
The memo also leaves room for differences between jobs and tasks. If a tool does not help with a particular assignment, documenting that limitation may inform teams’ decisions about where AI is worthwhile. That is a different approach from assuming either that every task should use AI or that high usage alone proves productivity.
Products Worth Considering
DJI Osmo Pocket 3 Filter Bundle – Vlogging Camera with 1 CMOS Sensor and 4K/120fps
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
DJI Osmo Pocket 3 Creator Combo Vlogging Camera with Mic 2, 1 CMOS, 4K/120fps, 3-Axis Stabilization, 128GB Card
As an affiliate, we earn on qualifying purchases.
Earlier Leaderboards Prompted the Change
The memo follows reports of internal tracking practices at Amazon. In May, the Financial Times reported that some employees gave an internal agent tool called MeshClaw unnecessary tasks to raise their token counts. The report also described a target for more than 80% of developers to use AI weekly and internal token-use leaderboards. Those accounts were attributed to anonymous sources and are reported claims, not details established by the October memo.
Amazon’s response to one leaderboard is on the record: at the end of May, a company spokesperson told Business Insider that KiroRank had been deprecated. The spokesperson described it as an informal tracker created by a group of employees and said it was never intended to encourage 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 sets out a broader official standard for engineering roles.
The source material also compares the memo’s emphasis on outcomes with Amazon’s AI investments, but those comparisons are not the memo’s stated evidence for its policy. It reports that Business Insider said most flagship Nova models were moved to “keep the lights on” this summer, while AWS lifecycle information lists several Nova products as retired. It also cites Amazon’s reported investment in Anthropic and a June 30 filing valuation. These are separate business developments, not proof of why the engineering guideline was issued.
““Tokenmaxxing is at odds with frugality.””
— Amazon’s engineering memo, written by Terk Terkowitz and senior engineers
How Amazon Will Judge Outcomes
The memo sets a principle but does not specify a single companywide formula for measuring customer, business or team impact. It is not clear how individual performance reviews will apply the guidance, how teams will define useful outcomes across different jobs, or whether existing usage dashboards and targets will change. The source material also does not provide independent evidence that higher adoption itself caused teams to deliver more; the memo says the measures currently correlate with relevant outputs.
Reports about employees inflating token counts rely in part on anonymous accounts. Amazon’s confirmation that KiroRank was deprecated does not independently confirm every detail of those reports. Nor does the memo establish that its policy was prompted by a specific incident. The company’s upcoming financial results are a separate event and cannot, without further evidence, establish the success or failure of its AI strategy.
Guidance Meets Quarterly Results
Amazon’s third-quarter results are scheduled for October 29, one day after the memo’s publication, according to the source material. Those results may provide company-level financial information, but they will not by themselves answer how well AI tools are improving engineering work. The memo’s immediate next step is practical: managers and engineers are to examine where AI helps, where it does not, and whether usage corresponds to better outcomes.
Further evidence would include how Amazon applies the revised role guideline in reviews, whether it changes its adoption tracking, and whether the company publishes measures linking AI use to customer or engineering results. Until then, the memo establishes a stated standard—measure impact, not token volume—without setting out a detailed scoring system.
Key Questions
What does Amazon mean by tokenmaxxing?
Tokenmaxxing means maximizing the amount of text processed by an AI model, including by giving a tool unnecessary tasks to increase token counts. The memo says token use should not be treated as a performance goal.
Is Amazon telling engineers to stop using AI?
No. The guidance says AI-use measures currently correlate with outputs that matter and notes that teams adopting AI are delivering more. Its warning is against scoring people on usage volume rather than judging the results.
What should managers measure instead?
The memo points to impact on customers, the business and the team. It does not provide one universal formula, and asks leaders to examine why adoption is low or high and whether heavy use produces better work.
Did Amazon confirm reports of employees inflating token counts?
The source material says the Financial Times reported the practice based on anonymous accounts. Amazon separately confirmed that the KiroRank leaderboard was deprecated, describing it as an informal tracker; that confirmation does not establish every detail in the earlier report.
Source: ThorstenMeyerAI.com
