An internal Amazon memo sent to software engineers in September and published this week has triggered anxiety about how the company will handle employees who find AI tools make their work worse, not better, according to a report by The Register. Written by Terk Terkowitz, a vice president and distinguished engineer at Amazon, the document addresses the role of developers in an AI-driven era and emphasizes customer outcomes and engineering judgment. But the memo's mixed signals have left engineers wondering whether raising concerns about AI ineffectiveness will be treated as legitimate feedback or a career liability.

The memo states that input metrics are "deeply flawed," yet advises managers to use them anyway, with caution. Within the span of three sentences, it warns that engineers who don't adopt AI are "falling behind" while also acknowledging that such metrics have serious limitations. Earlier this year, Amazon shut down KiroRank, an employee-created system that ranked workers by AI token usage, after staff began generating pointless tasks to boost their standings, a practice now called "tokenmaxxing." Meta launched a similar initiative called "Claudenomics" the same week in May, involving 85,000 employees. Amazon also announced another 250 job cuts this week, and engineers who read the memo believe these developments are connected.

The document advises that teams should examine the 90th percentile of AI adoption, looking at groups consuming the most tokens, and verify whether their output actually increased proportionally. According to the report, the memo includes a key phrase: "If your work does not benefit from current AI tooling, that is a finding, not a failing." It encourages engineers to document cases where AI hinders productivity, bring them to their managers, and work through the issue. The memo also identifies tasks that AI tools aren't currently improving, a recognition that not every objection to AI stems from fear of replacement.

The report raises concerns that the memo's contradictory positions give managers cover to interpret policy however they choose: citing "a finding, not a failing" when a strong engineer skips AI, or pointing to "falling behind" when it's time to put someone on a performance improvement plan. The question, the report notes, is whether promotion and performance decisions will reward long-term outcomes or simply optimize for what's easy to measure, like tool activity. The tension extends beyond internal operations: while Amazon's memo frames tokenmaxxing as contrary to its principle of frugality for internal costs, AWS bills its Bedrock and Kiro customers by the token, meaning that outside the company, token consumption drives revenue. Nothing currently ties AI spending to meaningful output like features shipped or bugs fixed. The industry faces a broader contradiction between measuring useful results and billing models based on consumption, leaving engineers to navigate policies that can be weaponized depending on management's intent. For companies adopting AI mandates, the real test isn't whether tools are available but whether employees who document AI making tasks worse are taken seriously or quietly managed out.