Amazon spent $1.8 million on a Claude Sonnet tool built to match author records against product listings, which ran 860% over budget for five months before anyone caught it, and the product never actually shipped, per the Financial Times. Two other projects overshot in the same stretch, including a financial auditing tool by roughly $541,000 and a logistics system by $134,000, for about $2.5 million in all. Engineers laid the overruns out at a July 28 staff meeting, saying errors that were once cheap to fix became “catastrophically expensive” after AI agents began doing the coding. Amazon called the examples cherry-picked and said they misrepresent how its teams use AI.
