Do more AI tokens mean more problems? Amazon is telling its engineers to stop chasing usage volume and instead focus on the actual value that AI can deliver to customers.
Amazon aims to move beyond the hype surrounding artificial intelligence, focusing instead on its practical application in engineering. The company has issued new guidelines for software development engineers, drawing a clear distinction between using AI to get actual work done and merely using it to project an image of high productivity.
According to an internal memo, engineers are encouraged to leverage AI to solve real customer problems, rather than simply inflating usage metrics. This guidance comes in response to reports that Amazon employees have faced pressure to increase their AI adoption rates.
As tech companies push for broader AI integration, Amazon is grappling with a complex challenge: what happens when employees are rewarded for the act of using AI, rather than for the outcomes they achieve?
Meet the “Renaissance Developer”
Amazon’s memo characterizes the next phase of software engineering as the era of the “Renaissance Developer.” This term refers to engineers who combine AI tools with technical expertise, human judgment, and a deep understanding of customer needs.
The company acknowledges that previous shifts—such as the adoption of cloud computing, microservices, and serverless architectures—have already reshaped software development. However, it argues that AI is transforming the process at a faster pace and a more fundamental level.
Yet, this does not mean engineers should cede their judgment to AI. Their core value remains the ability to build reliable systems and deliver meaningful results, rather than simply demonstrating that they are using the latest tools.
Amazon Warns Against “Token Maximization”
Perhaps the most pointed aspect of the memo is its warning against “token maximization”—a term used to describe developers assigning unnecessary or low-value tasks to AI agents solely to drive up token consumption.
Tokens are the units of text processed by AI models. Consuming more tokens increases computational costs without necessarily improving productivity. This practice has already become a point of concern within Amazon.
Previous reporting by YourStory highlighted internal pressure at Amazon to boost AI usage, with metrics established to track token consumption and the frequency with which employees utilize AI tools. It has been reported that some staff members have even resorted to using “MeshClaw”—an internal Amazon AI agent—specifically to inflate their usage figures. …to handle various trivial tasks.
This creates a challenge when it comes to measuring the adoption of artificial intelligence. If employees feel pressured to demonstrate that they are using AI, they might automate tasks that could actually be completed faster manually, or generate unnecessary prompts simply to boost their performance metrics.
According to reports, Amazon’s latest guidelines caution against using metrics like token consumption or the volume of AI-assisted code modifications as indicators of productivity. Such metrics might reward mere activity without reflecting whether the work actually provides value to customers. Furthermore, they could run counter to the company’s long-standing core value of frugality.
What does Amazon actually want to measure?
Amazon wants engineering leaders to focus on outcomes—specifically “feature velocity,” or how quickly useful updates reach customers.
The memo asks leaders to examine both ends of the spectrum. AI adoption varies widely; teams that use AI sparingly may need assistance in identifying meaningful applications, while teams with high AI usage must demonstrate that this increased engagement actually improves delivery efficiency.
Managers are also encouraged to use these AI tools themselves—to understand their limitations and identify where they truly add value. If a tool fails to improve a specific task, documenting that finding is a form of useful feedback, not a failure.
The guidelines also suggest that usage data should serve as a tool for leaders to understand how their teams operate, rather than becoming a target for employees to chase.
AI adoption is not a productivity race.
Amazon’s guidelines prioritize customer needs over raw AI usage statistics. Engineers who deeply understand the problems they are solving are better equipped to judge when AI can be helpful—and when it merely adds unnecessary complexity.
The company is not asking developers to avoid AI.
Rather, it wants them to demonstrate that the technology is worth using. For engineering teams, this is a crucial distinction: while AI adoption rates may be easy to track, the creation of superior software is something that cannot be faked.