Vol.029 — Tokenmaxxing at Amazon: What Happens When “Use AI” Becomes the KPI
Date: 2026-05-18 | Newsletter
Key Summary
Amazon built an internal platform called MeshClaw to track AI token consumption per developer, set an 80%-weekly-usage target, created internal leaderboards — and employees responded rationally by building agents that do nothing useful except consume tokens. The Hacker News community named it “tokenmaxxing” and called it “Goodhart’s Law going corporate.” Zenta uses the episode to make a structural argument: we do not yet know how to measure AI productivity honestly, and when a matrix-driven culture meets an unmeasurable thing, the matrix wins — it measures what it can count, which is input, not outcome. The failure pattern transfers directly to ASEAN SMEs, Zenta warns, and faster: Amazon can absorb two years of fake-AI-usage; a 60-person company cannot absorb six months. An ASEAN-specific accelerator makes it worse: AI adoption grants from bodies like IMDA often carry reporting requirements on adoption metrics — number of employees trained, workflows touched — meaning the Goodhart trap is baked into the funding structure contractually. Functional AI Partners recommends three questions before setting any AI-related KPI: what specific business problem am I solving, what does success look like as an outcome not an activity, and am I willing to wait rather than set a KPI I know will be gamed? Do it simple. Define the problem before you mandate the activity.