Amazon flags AI projects after one Claude Sonnet effort ran 860% over budget

Internal reports showed several AI initiatives generated steep unexpected costs, prompting Amazon to tighten controls with guardrails and spending caps as scrutiny grows over whether heavy AI use delivers value.

25d ago

Fact Check

The claim is corroborated by a Financial Times report whose headline directly confirms Amazon found AI causing runaway spending, and by multiple independent secondary outlets (Tom's Hardware, Unite.AI, AI Weekly, Daily.dev) that all cite the FT and consistently report the same key facts: a Claude Sonnet author-mapping project cost $1.8M — 860% over budget — and went undetected for five months, alongside smaller overruns of $541K and $134K. All sources agree Amazon is tightening controls with automated guardrails and spending caps and scrapped an internal usage leaderboard. The headline claim of '860% over budget' and 'guardrails and spending caps' is precisely matched. Only slight uncertainty remains because the underlying FT primary text is paywalled and full verification relies on consistent secondary reporting.

Summary

Amazon internally flagged several AI projects for major cost overruns, including a product-listing effort using Anthropic's Claude Sonnet that cost about $1.8 million and ran 860% over budget before management detected the problem nearly five months later. Internal reports also showed a financial audit tool generated about $541,000 in unexpected expenses and a logistics optimization project added roughly $134,000 in costs that went unnoticed for more than two weeks. Employees described some AI failures as "catastrophically expensive" because autonomous agents can keep consuming tokens when errors are not tightly constrained. Amazon had also shut down an internal AI-usage leaderboard after workers gamed it by boosting token consumption, and it scaled back permissions for its Kiro coding assistant after an incident that led to a 13-hour AWS outage. Amazon said it is experimenting and improving how it uses AI and that isolated cases do not reflect day-to-day use across the company. The episode highlights a broader industry debate over whether rising AI usage and token consumption are translating into real productivity gains, even if the dollar amounts remain manageable for a company of Amazon's size.

Terms & Concepts

Claude Sonnet: An AI model from Anthropic used by companies for generative AI tasks.
tokens: Units of text processing that help determine how much usage an AI model consumes and how much it costs.
guardrails: Controls that limit or monitor how AI systems operate to reduce risk and runaway spending.