Amazon is wiring budget caps into its AI deployment path after internal post-mortems revealed what token billing can do to a project. Engineers presented cases on July 28 where model-driven work blew past cost estimates, with the Financial Times reporting that one job burned $1.8M on Claude Sonnet calls and delivered nothing.
The worst case matched author records against retail product listings. It exceeded its planned spend by 860%, took five months for anyone to notice, and the rollout still failed. Smaller overruns surfaced too, roughly $541,000 on financial auditing tools and $134,000 on delivery-speed improvements across the logistics network.
The pricing math explains the scale of the damage. Claude Sonnet costs $3 for a million input tokens and $15 for a million output tokens, so a loop that keeps resending the same context, or a batch run pointed at a full catalogue, never raises an exception. It just generates a monthly invoice instead of a build-log error.
Management responded by shifting enforcement earlier in the cycle, building automated guardrails that sit inside the deployment pipeline rather than in monthly reviews. The move follows the shutdown of an internal leaderboard for the Kiro developer platform, which staff had gamed by inflating token usage, a practice employees dubbed “tokenmaxxing.”
Amazon stresses the examples represent a handful of groups out of a workforce near 300,000, and says it is experimenting and improving its AI usage.
