Amazon has issued internal guidance to software engineers discouraging "tokenmaxxing"—the practice of using AI tools excessively with little regard for output quality—as the company grapples with rising AI costs and the challenge of measuring genuine productivity gains [1]. The memo, circulated internally in September and posted publicly on Tuesday, reflects a broader industry concern about how to reliably assess whether AI adoption delivers valuable returns on investment [1]. Amazon noted that common metrics like token usage (the units of data processed by AI models), adoption rates, and code changes are "deeply flawed" ways of measuring AI effectiveness, as they can drive up costs and unnecessary complexity [1].
Under CEO Andy Jassy, Amazon has rapidly integrated AI across its operations, with Jassy previously predicting that AI will "change the scope and speed" of work and potentially reduce corporate headcount through automation [1]. However, the company has discovered that employees sometimes game the system to appear productive—Amazon shut down an internal leaderboard in May after discovering employees were tokenmaxxing to climb rankings [1]. While Amazon has not instituted hard caps on token usage, the memo advises that token consumption and code generation metrics should be considered "with care and judgment" alongside other data rather than as primary performance indicators [1].
The company advocates for measuring AI's impact through "long-term outcomes" and metrics such as "feature velocity," according to Marc Brooker, a vice president and distinguished engineer at Amazon Web Services [1]. This approach reflects a maturation in how enterprises are thinking about AI adoption—moving beyond simple usage metrics to focus on tangible business value. The guidance underscores the tension many technology companies face as they scale AI tools: balancing the desire to encourage innovation with the need to control costs and ensure that AI investments translate into meaningful productivity improvements.
CONCLUSION
Amazon's anti-tokenmaxxing memo signals a shift in how major tech companies are approaching AI adoption, moving from usage-based metrics to outcome-focused measurement. This reflects industry-wide challenges in controlling AI costs while ensuring genuine productivity gains, suggesting that companies may need more sophisticated frameworks to evaluate AI ROI beyond simple adoption rates.