Amazon’s memo puts judgment ahead of AI token volume

Amazon circulated an AI-use memo to employees in September, posted it on its website on Tuesday, and discouraged software engineers from “tokenmaxxing” when AI tools do not improve their work.

The September document revises how Amazon defines the roles of software development engineers and managers as artificial intelligence takes on a larger role in software practice.

Amazon’s warning targets what the report calls tokenmaxxing: using AI as much as possible with little regard for whether the output is useful.

Amazon’s memo enters a market with financial and managerial stakes: companies adopting AI face rising model bills and limited evidence that heavy usage produces higher productivity or a worthwhile return on investment.

What does the memo change?

Amazon starts from a position of uncertainty. Its memo says the company is seeing only the beginning of AI’s effect on software engineering and that many best practices remain unsettled.

“We're just seeing the beginning of AI's effect on software engineering. While we have a vision for what the future looks like, there's a lot that's still emerging. We are certain that whatever future best practices are, AI will be at the center of software engineering practice.”

That statement sets a clear boundary for the revised role definitions. Amazon is not presenting AI as a temporary experiment or a tool to remove from the engineering workflow.

Instead, the memo makes judgment part of the job. Engineers and managers must decide where AI adds value, where it fails, and how to evaluate the resulting work over time.

Amazon singles out three common measures of organizational AI use and calls them deeply flawed indicators of team performance:

• Token usage.
• The rate of adoption of AI tools.
• Code changes and code reviews.

The company says these measures can drive costs higher and create excessive complexity. More tokens or more generated code do not, by themselves, establish that an engineering team is producing better software.

This distinction changes how adoption data should be read. A high rate documents tool use, while a low rate says little about output quality. Amazon treats both as incomplete rather than dispositive.

Amazon does not tell employees to discard the measures entirely. The memo says token usage and code generation should be considered with care and judgment alongside other data.

Its preferred standard is long-term outcomes. Amazon says the best approach is to measure, and be measured by, long-term outcomes, rather than the volume of AI activity consumed.

That distinction addresses a central measurement problem. A leaderboard rewards activity, while an outcome standard asks whether the activity changes the work in a useful and durable way.

Why Amazon is targeting tokenmaxxing

The report says employees have rushed to adopt AI tools, prompting companies to confront two linked problems: the cost of processing data and the difficulty of proving a return.

Tokens are the units of data processed and generated by AI models. Some companies have set limits on employee token use to slow rising AI bills, according to the report.

Other companies have moved selected tasks to cheaper, open-source models. Those choices represent cost control, but they do not settle the separate question of whether AI-generated output is useful.

Amazon’s internal position differs from those hard limits. An Amazon spokesperson said the company had not instituted caps on employees’ token usage.

Amazon did take one concrete step before publishing the memo. In May, it shut down an internal leaderboard after discovering that employees were tokenmaxxing to climb the rankings.

The shutdown removed an incentive to consume tokens for the leaderboard’s sake. The September memo addresses the same behavior without announcing another numerical ceiling.

The contrast matters because usage limits and outcome standards solve different problems. A cap limits expense; it cannot determine whether an engineer’s work improved.

Amazon’s memo also stops short of treating no AI usage as an achievement by itself. The company accepts that some tasks may benefit from current tools and others may not.

What does it mean for engineers and managers?

The revised role definition places judgment alongside technical execution. An engineer now has to assess current AI capability against a specific task, including cases where the tool offers no useful assistance.

Amazon makes that expectation explicit. It says recognizing that current AI tooling does not benefit a particular task should count as a finding rather than a failure.

The memo then gives engineers a specific standard to demonstrate:

“If your work does not benefit from current AI tooling, that is a finding, not a failing. An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for.”

That wording matters for managers evaluating staff. It makes an explanation of ineffective use a sign of professional judgment, rather than evidence that an employee has failed to embrace a company tool.

It also narrows the meaning of AI adoption at Amazon. Adoption rates can count how many people use the tools; the memo asks whether those tools fit the work and produce results that last.

Amazon still places AI at the center of software engineering practice. The memo says future best practices will do the same. Its target is indiscriminate use, not adoption itself: employees should not maximize token consumption or force AI into work that does not benefit from it.

How does this fit Amazon’s workforce strategy?

Under CEO Andy Jassy, Amazon has rapidly integrated AI across its sprawling business operations. Jassy has said the technology will change the scope and speed of day-to-day work.

Jassy also predicted that AI could shrink Amazon’s corporate workforce in the next few years through automation, while creating other jobs. The memo supplies no headcount target or timetable.

The memo arrives as Amazon considers both sides of that forecast. It supports broader AI use, but it also asks the company to measure the return from that use more carefully than raw usage data permits.

That approach also reflects the cost pressure described across the industry. The report says companies are weighing token limits and cheaper model choices while searching for reliable productivity evidence.

Amazon’s May leaderboard closure and September guidance show two different responses. First came removal of a usage-based ranking; next came a written standard for evaluating long-term work.

What should readers watch next?

Amazon’s memo makes managerial implementation the immediate test: engineers and managers must judge task fit and long-term outcomes, but the document does not define a single productivity measure that can settle either question.

It also does not announce a token cap, a replacement leaderboard, a numerical benchmark, or a new workforce schedule. An Amazon spokesperson confirmed only that no employee token caps are in place.

For other companies, the practical comparison will be between restrictions and incentives. Token ceilings can control bills, while usage rankings can encourage waste; neither tells managers whether software work improved.

For Amazon, managers must turn the stated judgment standard into consistent evaluation without recreating the conditions that produced tokenmaxxing. That is the unresolved operational issue the memo identifies.

Until the company and the wider industry establish better measures, Amazon’s own stated standard is explicit: judge AI by long-term engineering outcomes, and treat a task where current AI does not help as a legitimate finding.