Amazon’s internal AI usage metrics revealed that a single Claude Sonnet deployment consumed $1.8 million in token spend, according to the Financial Times. The project, designed to automate commercial loan agreement reviews, exceeded its allocated budget by 860% and ran for five months before the overspend was detected.
The Three Incidents
The loan agreement task was the most expensive failure, but not the only one. Amazon’s internal cost tracking surfaced two additional overruns, as Tom’s Hardware reported: a $541,000 overage on a project building a financial auditing tool, and $134,000 in excess costs for a system designed to reduce delivery times in the company’s logistics network.
All three shared the same structural problem. Teams allocated budgets based on per-seat or fixed-cost assumptions from the subscription era. Token-per-request billing made those assumptions irrelevant. When an AI agent runs autonomously on a task, each iteration burns tokens with no natural ceiling.
Amazon’s Response
“As with any new technology, we’re experimenting, learning and improving how we use it, including how we drive cost efficiencies,” Amazon said in an internal presentation, according to the Financial Times. The company characterized the overruns as isolated learning experiences.
In dollar terms, the combined overspend is a rounding error for a company reporting more than $181 billion in quarterly revenue. But the detection gap matters. Five months of unchecked token burn on a single task suggests Amazon’s existing cost controls had no mechanism to flag agentic workloads running beyond their budgets.
Amazon has already taken corrective steps on AI agent access. Earlier in 2026, AWS limited the permissions of AI agents after coding bot blunders caused outages, restricting agents to lower access levels rather than granting them the same permissions as senior engineers. The company also scrapped an internal leaderboard that ranked employees by AI usage volume after the practice drove unnecessary token consumption.
The Pricing Structure Problem
The overruns illustrate a cost structure collision that extends well beyond Amazon. Enterprise procurement systems were built for fixed annual licenses and per-seat subscriptions. Token-per-request billing introduces variable costs that scale with agent autonomy: the more tasks an agent handles, and the more iterations each task requires, the higher the bill.
OpenAI’s decision on July 30 to cut GPT-5.6 Luna prices by 80% suggests the vendor side is adjusting. But lower per-token prices do not solve the governance gap. A cheaper token still burns budget when an agent runs without cost ceilings, approval gates, or real-time spend monitoring.
For enterprises deploying autonomous agents across business operations, the Amazon incidents point to a specific gap: cost controls that track cumulative token spend per task, trigger alerts at budget thresholds, and halt agent execution when limits are reached. Without those controls, every agentic deployment carries the same risk of silent budget overrun that Amazon discovered five months too late.