According to CoinMeta, multiple AI projects at Amazon were reported to have incurred serious overspending. The most extreme case involved using Claude Sonnet to match author information on product pages; the project ultimately failed, costing a staggering $1.8 million, which is 860% more than the budget, and it was not discovered until five months later. Another financial audit tool project unexpectedly cost about $541,000 more, and a project aimed at optimizing logistics and delivery speed also overspent by $134,000, with these issues only being identified more than two weeks after they occurred. Amazon engineers attributed the problems to deployment errors and a lack of cost constraints. What are minor mistakes that hardly cost anything in traditional systems can turn into catastrophically expensive bills in AI projects. The company is developing automated measures to prevent similar projects from getting out of control in the future. Amazon responded that these cases only involved a few teams and do not represent the overall situation of the company's daily use of AI.