AI led to 860% budget overrun — Amazon admits neural networks made errors 'catastrophically expensive'

Internal Amazon reports show that using AI led to significant cost overruns across projects. One project saw a $1.8 million overrun when the Claude Sonnet model failed to match customer data with product cards, exceeding its budget by 860%, discovered only five months later. Other cases include a $541,000 overrun on a financial audit tool and $134,000 on a logistics system. Amazon stated they are experimenting and learning, but the examples reflect isolated incidents. The overruns represent less than 0.1% of Amazon's monthly revenue. Previously, AWS had issues with an AI coding bot that was later restricted, and Amazon dropped an AI usage leaderboard due to rising costs.