Amazon has told its software engineers that the AI usage figures it relies on to gauge adoption are "deeply flawed" - and, a few paragraphs later in the same memo, that those figures still track performance well enough to be used "with care and judgement." The memo, published on the company's news site on October 27, 2026, explains the thinking behind a revised role guideline for software development engineers (SDEs). It was posted by Terk Terkowitz, vice president and distinguished engineer for Worldwide Amazon Stores, who described it as a memo shared with software engineers across Amazon.

In Short

Amazon rewrote the job description for its software engineers and decided to judge them mainly on what they deliver for customers, not on how much AI they use. That matters because many companies, Amazon included, have counted things like AI tokens and lines of code to see who is using AI, and the memo admits those counts can be gamed and can raise costs. Engineers are still expected to learn AI quickly, but one who can show where AI does not help their work is now treated as making a useful finding rather than failing.

A guideline built on what does not change

Engineering leaders from across the company, including its Distinguished and Senior Principal Engineers, came together to "reaffirm what it means to be an SDE at Amazon," according to Terkowitz. The updated guideline drew on "key learnings and observations from the most successful SDEs at the company." Terkowitz's post on Amazon's news site does not say when that meeting took place or when the memo first reached staff. According to The Next Web, which cited CNBC, Amazon first shared the memo internally in September.

The memo is addressed to SDEs and to software development managers (SDMs), and it opens with the question its authors faced: "how much do we focus on AI?" Amazon's SDE guidelines have historically avoided naming technologies, according to the memo, because the company runs many different businesses and its engineers do many different kinds of work. That approach held through two decades of change the memo lists as "the cloud, microservices, serverless, etc." In the Senior Principal and Distinguished Engineer communities, the memo says, leaders "deeply debated whether this time is different."

They concluded that it is. "AI is changing the work of software engineering both more quickly and more fundamentally than these previous trends," the memo states. "What it means to be a software engineer, and what software engineering even is, is changing quickly and durably."

Even so, the authors chose not to write a definition of good AI practice into the guideline. Their reasoning was about shelf life. "Any definition we write now will likely look antiquated by this time next year," the memo says, warning that a fixed definition risks "freezing our engineering practices in amber while the world moves on without us." The memo sets its answer in bold: "To solve this problem, we have decided to focus on what is not changing as we evolved the SDE roles." Those constants are impact on customers, on the business and on teams. "The outcomes that matter aren't changing, even if the inputs, tools, and techniques are changing faster than ever."

To support the principle, the memo quotes Jeff Bezos from a 2012 interview with Werner Vogels: "'What's going to change in the next five to ten years?' But I very rarely get asked 'What's not going to change in the next five to ten years?' That second question is actually the more important of the two."

None of this makes AI optional. "We all need to embrace this change and take the time to learn, or we risk being left behind," the memo says. The authors are "certain that whatever future best practices are, AI will be at the center of software engineering practice," and engineers not yet using AI are told to start "as soon as possible."

The renaissance developer, again

Terkowitz framed the revision in language Amazon has used before. "We're now in the era of the Renaissance Developer, blending AI tools with human judgment and domain expertise to build entirely new things," he wrote, linking to Werner Vogels' predictions for 2026. Amazon's chief technology officer described renaissance developers in November 2025 as engineers who combine technical expertise with broader domain knowledge, and argued that generative AI would not replace developers.

Terkowitz's introduction also makes a claim about the guideline's audience beyond Amazon. "As a leader in agentic AI with thousands of SDEs, Amazon believes sharing our vision for the SDE role of the future can serve as a model for our customers," he wrote. The post gives no precise count of SDEs.

The metrics problem

The most candid passage concerns measurement. Amazon's stated preference is to "measure, and be measured by, long-term outcomes." Outcomes, though, can be "time delayed, difficult to measure, and hard to isolate from other trends," the memo concedes. So the company has used proxies. "In some cases, we have tackled those problems by using input metrics instead, like token usage or adoption of AI tools, or throughput metrics like code changes and reviews."

Then comes the admission: "We are all aware of how deeply flawed these metrics are, primarily because they are disconnected from what matters."

The memo names two ways the proxies backfire when treated "as a target instead of a measure." Each is set against one of Amazon's Leadership Principles. On cost: "tokenmaxxing is at odds with frugality." On complexity: "measuring LOC is at odds with invent and simplify." LOC stands for lines of code. Tokenmaxxing describes inflating consumption of tokens - the units AI models read, generate and bill by - to look productive on a usage ranking. The distinction between a measure and a target is the core of what economists know as Goodhart's law, although the memo does not use the name. Its proposed safeguard is cultural rather than technical: the authors say they expect the engineering community's ownership and focus on delivering results to keep those outcomes in check.

Flawed, but correlated

The following paragraph turns. "More positively, we have found that these input metrics correlate well with the outputs that matter and expect they will continue to do so for some time," the memo says. It attributes the correlation to two effects: rising productivity and delivery on the teams that have embraced AI the most - "we call them frontier teams" - and the fact that "teams and individuals who are not adopting AI or increasing output metrics are falling behind."

From there, the authors conclude that "using these input metrics, with care and judgement, as you think about your career and your organization's performance is appropriate."

The tension is hard to miss. The same document calls the metrics disconnected from what matters and says they correlate well with what matters. Can both be true? Perhaps, if the correlation holds on average while individual numbers are easy to game. But the memo publishes nothing to settle the point. It does not say which outputs were measured, how strong the correlation is, how many teams count as frontier teams or how much faster they deliver. It does not say what share of Amazon's code is written with AI. The correlation is asserted, not shown.

Percentiles instead of targets

What the memo does prescribe is a way of reading the distribution of AI adoption. Leaders are told to pay "special attention" to two points in it. The first is the 10th percentile: "finding those teams who aren't embracing AI effectively and understanding why." The second is the 90th percentile: "finding those teams who are heavily using AI and understanding whether they're seeing delivery increases to match."

The second check is the more revealing one. A team at the 90th percentile of usage whose delivery has not improved is, in effect, the tokenmaxxing case the memo warns about. The first check looks for laggards. Neither sets a number that engineers must hit. "Working together, over time, we will find better metrics and set smarter goals," the memo says - an acknowledgement that the current set is provisional.

A second Bezos passage, from his 2016 letter to shareholders, backs the approach. Good inventors, Bezos wrote, "study and understand many anecdotes rather than only the averages you'll find on surveys." The memo adapts the line for managers: "Good leaders deeply understand their team's work. They spend tremendous energy developing that intuition. They study and understand many anecdotes, and carefully track outputs, rather than only surface-level metrics."

"A finding, not a failing"

One sentence addresses engineers whose work does not suit the current tools. "If your work does not benefit from current AI tooling, that is a finding, not a failing," the memo says. Such engineers are asked to document why, share the result with peers, raise it with their manager and work on solutions. "An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for."

The allowance is narrower than it may first appear. It covers work that the tooling does not help, not engineers who have not tried. The 10th percentile check exists to find exactly those teams "who aren't embracing AI effectively."

A hands-on test for managers

Much of the burden falls on leaders. Technical and organizational leaders must ensure their teams use AI "to the best effect to improve delivery for customers," according to the memo, which requires "working with them hands-on rather than relying on what you hear." The memo then defines how much understanding is enough:

"Sufficient understanding means you have used these tools on your production work, can describe a task where they helped and a task where they did not, and can explain that judgement to your team."

That is a three-part test: use on production work, one concrete example on each side, and the ability to explain the call. The memo describes AI deployment as a balance between exploring new and better methods and doubling down on existing ones. It adds: "As a leader, we expect you to be taking ownership of connecting the inputs and outputs in your team or organization."

The leaders having the most impact, according to the memo, are those "doubling down on learn and be curious" - another Leadership Principle - by making time to learn AI and by giving their staff time to do the same. Teams that followed their curiosity, balanced with "deliver results" and "insist on the highest standards," saw "the most positive impact from AI adoption," the memo says. The returns are described as "disproportionately large." No figures accompany the claim.

Back to the customer

The memo's final instruction has little to do with tooling. Engineers and managers who are closer to their products and businesses "are able to take the most advantage of AI and agentic software development," it says. Then comes an admission about Amazon's own structure: "As Amazon has grown, we've seen some of our ICs and SDMs become less connected to their customers and business. We want to reverse this trend." Individual contributors (ICs) at all levels are asked to become "deeply connected" to customers and to product direction, with the scope of that ownership rising with seniority.

Amazon has pushed in this direction before. In September 2024, Andy Jassy set a goal of raising the ratio of individual contributors to managers by at least 15% by the end of the first quarter of 2025, alongside a return to full-time office work.

The emphasis on domain knowledge is also familiar from outside Amazon. In January 2026, Google principal engineer Jaana Dogan said Claude Code had produced in about an hour an architecture resembling what her team built over the previous year, while pointing to her years of distributed systems experience as the reason she could judge the output. She called the result a toy version, not production-grade software.

The memo closes by saying Amazon "will continue to evolve the SDE role over the coming years, as AI's effect on our business and engineering practices becomes clear." It predicts that what matters will stay the same: "obsession with our customers, and a long-term focus on delivering results."

The leaderboard Amazon already retired

The reference to tokenmaxxing did not come out of nowhere. In May 2026, Amazon shut down KiroRank, an employee-built leaderboard that ranked staff by their activity on Kiro, its AI development platform, according to The Next Web citing CNBC and HRD Canada citing the Financial Times. Amazon said the beta dashboard was not a formal or approved tool and had been deprecated, according to HRD Canada. Senior vice president Dave Treadwell told staff the tool had good intentions but raised costs because workers inflated their token use, the same report said. According to HRD Canada, Amazon has set a target for more than 80% of its developers to use AI each week, and now tracks a metric called normalized deployments, which counts how often engineers use AI to ship useful code rather than raw token totals.

Business Insider reported that Amazon tracked engineers' AI usage, token consumption and lines of code, and used leaderboards to rank some teams, according to The Next Web. The same article said CNBC reported that an Amazon spokesperson described token usage as uncapped.

Read against those reports, the memo appears to formalize a correction that was already under way. It also confirms, in Amazon's own words, which proxies the company has used: token usage, AI tool adoption, code changes and reviews.

Lines of code, in particular, have featured in Amazon's own messaging. In August 2026, Beth Galetti, the senior vice president who leads Amazon's human resources organization, wrote that she had produced more than 100,000 lines of code with AI tools after a 25-year break from programming. The engineering memo, published two months later, describes measuring LOC as at odds with invent and simplify.

A buyer and a seller of tokens

There is a further complication. Amazon consumes tokens internally and sells them to customers. Jassy's letter to shareholders, released April 9, 2026, said Bedrock processed more tokens in the first quarter of 2026 than in all prior years combined. The same letter described a six-person engineering team using Kiro to build Mantle, a new Bedrock inference engine, in 76 days. When Terkowitz writes that Amazon's approach "can serve as a model for our customers," those customers include businesses deciding how much to spend on the same kind of tools.

Thirty thousand roles and a $220bn budget

The guideline follows the largest round of corporate job cuts in Amazon's history. The company announced 14,000 corporate job cuts on October 28, 2025, with Galetti describing AI at the time as the most transformative technology since the internet. Amazon confirmed a further 16,000 cuts on January 28, 2026, bringing the total to roughly 30,000, about 10% of its corporate workforce. Its stated reasons that month centered on fewer layers and less bureaucracy. On October 6, 2026, Amazon eliminated fewer than 1,000 more roles, mostly in its Stores business, according to a person familiar with the matter. Amazon did not confirm that figure.

Spending has moved the other way. Amazon flagged about $200bn of capital expenditure for 2026 in February, then raised the forecast by $20bn to $220bn at its second-quarter results. The memo mentions neither the cuts nor the capital budget.

What outside research shows

The memo's claim that input metrics track outcomes is hard to test from outside, but published research offers some reference points - and few of them are simple.

A Stanford team reviewing 84 papers on agentic AI evaluation found that technical metrics appeared in 83% of evaluations, while human-centered assessments appeared in 30% and only 15% combined the two. The authors argued that benchmark success often fails to predict real-world value - the same gap between inputs and outcomes that the memo describes.

Anthropic researchers tested 52 developers in a study released January 29, 2026. Those using AI assistance scored 50% on a comprehension quiz against 67% for the manual group, while finishing about two minutes faster, a difference the study did not find statistically significant. A study from UC San Diego and Cornell University found that experienced professionals did not practice vibe coding: all 13 observed developers kept control of software design, and survey respondents rated agents as unsuitable for business logic by 15 to 2.

At the heavy end of usage, the distribution is steep. Top-decile enterprises produce about 8x the output tokens of a typical enterprise, according to a16z data from August 2026, which is the kind of spread the memo's 90th percentile check is designed to interrogate. Experience appears to matter, though less than headline numbers suggest: Anthropic's March 2026 Economic Index report found that users with six months or more on Claude succeeded on 73.1% of tasks against 66.7% for newer users, a gap that shrank to about 3 to 4 percentage points after controls.

Trust remains low among developers in general. In Stack Overflow's 2025 developer survey, 84% of developers said they use AI tools, while 46% said they distrust the accuracy of AI output.

Why it matters for marketing and advertising

Amazon's memo is a statement about engineers, but the questions it raises apply to any team under pressure to show AI adoption.

Several companies have gone further in the other direction. Shopify chief executive Tobi Lütke made AI usage part of performance reviews in April 2025, and teams there must show why a task cannot be done with AI before requesting more headcount. At Opendoor, chief executive Carrie Wheeler made "Default to AI" the first line of every employee's job expectations in September 2025, with managers assessing how often staff reach for AI first. Amazon's memo keeps adoption tracking, but it says usage alone is a weak guide and that engineers who can show where AI does not help are exercising judgement.

Marketers face a version of the same measurement gap. In MiQ's survey of 3,169 marketers across 16 countries, 44% cited an inability to track results against goals as a barrier to AI use, while 72% planned to apply AI in more ways over the following year. Counting prompts or tokens is easy. Tying them to revenue or efficiency is not.

The memo also matters because of who reads it inside Amazon. Its guideline covers software engineers across a company whose advertising business generated $19.8bn in the second quarter of 2026, up 26% year over year. Amazon Ads has spent the past year building agentic AI products for advertisers, from Ads Agent, announced at unBoxed on November 11, 2025, to a server built on the Model Context Protocol that connects AI agents to Amazon Ads API functions and entered open beta on February 2, 2026. How the engineers behind such products are evaluated - by tokens consumed or by delivery - shapes what gets built.

Then there is cost. Agencies and in-house teams pay for tokens by volume. The memo's line that "tokenmaxxing is at odds with frugality" is, for them, a budget question as much as a cultural one.

Timeline

Summary

Who: Amazon's engineering leadership, including its Distinguished and Senior Principal Engineers, in a memo to software development engineers and managers, posted publicly by Terk Terkowitz, vice president and distinguished engineer for Worldwide Amazon Stores.

What: A revised role guideline for software development engineers that defines the job by impact on customers, businesses and teams rather than by AI tools. The memo calls input metrics such as token usage, AI tool adoption, code changes and reviews "deeply flawed," warns against tokenmaxxing and lines-of-code targets, but says the metrics correlate with outcomes and can be used with care. Leaders are told to study teams at the 10th and 90th percentiles of AI adoption, to use AI on their own production work, and to bring engineers closer to customers. Engineers whose work does not benefit from AI are told that this is "a finding, not a failing."

When: Published on Amazon's news site on October 27, 2026. According to The Next Web, citing CNBC, the memo was first shared with staff in September 2026, months after Amazon shut down the KiroRank leaderboard in May 2026.

Where: Amazon's software engineering organization worldwide, with the memo shared on aboutamazon.com as a model Amazon says its customers could follow.

Why: According to the memo, AI is changing software engineering faster and more fundamentally than the cloud, microservices or serverless did, so any written definition of good AI practice would quickly go out of date. The authors chose to anchor the role in outcomes that do not change, while acknowledging that the metrics Amazon has used to track AI adoption can drive up costs and complexity when treated as targets.