Amazon Internal Reveals AI Cost Overrun, Minor Error Leads to Million-Dollar Bill

According to Perceptive Beating monitoring, The Financial Times reported citing insiders that Amazon internally flagged multiple AI projects for severe cost overruns. In one particularly egregious instance, funds were spent on using Claude Sonnet to match author information on product pages. The project ended up as a failure but still incurred $1.8 million in expenses, exceeding the budget by 860%, and was only discovered five months later.

Another financial audit tool project unexpectedly overspent by approximately $541,000. A project aimed at optimizing logistics and delivery speed also misused $134,000, a discrepancy that was only noticed over two weeks later.

Amazon engineers attributed the issues to deployment errors and a lack of cost controls. Minor mistakes that would hardly cost anything in traditional systems could translate into "disastrously expensive" bills in AI projects. The company is working on implementing automation measures to prevent similar projects from spiraling out of control.

In response, Amazon stated that these cases were isolated to a few teams and do not reflect the overall picture of AI usage in the company's day-to-day operations. Previously, Amazon had also shut down an internal AI leaderboard because employees were gaming the system by having agents carry out meaningless tasks, resulting in a significant additional token consumption.