Amazon Reports AI Cost Overruns, With Some Projects Overspending by 860%

According to Beating Monitor, citing sources familiar with the matter, Amazon has reported multiple AI projects significantly exceeding budgets. The most extreme case involved using Claude Sonnet to match author information with product pages; the project ultimately failed but cost $1.8 million—860% over budget—and was not discovered until five months later. Another project, a financial auditing tool, unexpectedly incurred an additional $541,000. A project aimed at optimizing logistics delivery speed also overspent by $134,000, with the error going unnoticed for over two weeks. Amazon engineers attributed these issues to deployment mistakes and the absence of cost controls; minor errors that would cost virtually nothing in traditional systems can become “catastrophically expensive” in AI projects. The company is developing automated safeguards to prevent similar projects from spiraling out of control. Amazon responded that these cases involve only a few teams and do not reflect the company’s overall use of AI. Previously, Amazon shut down its internal AI usage leaderboard after employees manipulated it by instructing agents to perform meaningless tasks, unnecessarily consuming large amounts of tokens.