Amazon warns against tokenmaxxing, telling engineers that using more AI doesn’t necessarily mean they’re doing better work

The e-commerce giant published a memo on Tuesday revising its guidelines for software engineers, saying Amazon has entered “the era of the Renaissance Developer” who blends AI and their own expertise and judgment “to build entirely new things.” This includes telling management when AI isn’t useful to their work and explaining why. 

“If your work does not benefit from current AI tooling, that is a finding, not a failing,” Terk Terkowitz, vice president and distinguished engineer at WorldWide Amazon Stores, wrote in the memo. “Document why, share it with your community of peers, bring it to your manager, and work to find solutions. An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for.”

The idea is that indiscriminate AI use could end up working against the company. Terkowitz wrote that counting tokens and measuring lines of code are “deeply flawed” ways to gauge productivity because they don’t necessarily reflect the quality of work. He wrote that turning those metrics into “targets” could drive up costs—”tokenmaxxing is at odds with frugality”—and add unnecessary complexity, which undermines Amazon’s principle of “invent and simplify.” 

“They are disconnected from what matters,” he wrote. 

Amazon has already grappled with the pitfalls of encouraging AI use for its own sake. Employees were reportedly running AI agents on meaningless tasks to inflate their token count and climb an unofficial internal AI leaderboard called Kirorank. Amazon nixed it in May after the practice raised concerns about unnecessary computing costs, which can add up quickly. An unnamed company accidentally spent $500 million on Claude in one month after failing to put limits on employees’ usage, Axios reported.  

The e-commerce giant isn’t alone in specifying AI best practices. Shopify’s CEO Toby Lütke, who previously told employees in a memo that embracing AI is “a baseline expectation,” said last month that some workers are producing “slop grenades”—AI-generated emails and code that create more work for their colleagues to review afterward. 

Still, Amazon is far from discouraging AI use, and even far from abandoning its AI adoption metrics altogether, since Terkowitz wrote they still “correlate well with the outputs that matter.” Terkowitz encouraged managers to compare the 10th percentile and 90th percentile of AI users on their teams—for those using it the least, to understand why, and for those using it the most, to determine whether their output is increasing accordingly.

“Using these input metrics, with care and judgement, as you think about your career and your organization’s performance is appropriate,” he wrote. “Working together, over time, we will find better metrics and set smarter goals.”