Amazon staff have identified catastrophically expensive cost overruns caused by mistakes in AI deployments and a lack of spending controls. Senior engineers said switching tasks to AI models led to unplanned spending and are building automated guardrails to curb future waste.

One project spent $1.8mn on matching author details with Anthropic’s Claude Sonnet despite failing, an 860% budget overrun detected after five months. A second incident incurred $541,000 in unexpected costs for financial‑auditing tools, and a third logged $134,000 in accidental spending on logistics‑speed AI, discovered after two weeks.

Amazon says it is “experimenting, learning and improving” cost efficiencies and that the overspending is not isolated. The cases involve only a handful of teams among roughly 300,000 corporate employees and $180bn quarterly revenue, while the company plans $200bn in capex this year, mostly for AI and data‑centre infrastructure.

Across the industry, token‑based billing from providers like Anthropic and OpenAI is driving cost spikes, prompting firms to shift to cheaper mid‑tier or open‑weight models. Amazon recently shut an internal Kiro leaderboard after “tokenmaxxing” inflated AI token consumption.