Some teams within Amazon have incurred millions of dollars in unplanned costs as a result of the uncontrolled use of Anthropic’s Claude Sonnet AI model. According to the Financial Times, in one of the largest incidents, an error in an AI agent that checks data on book authors and service providers caused the company to spend $1.8 million (860%) more than budgeted on token usage. It took five months to detect the problem.
Another project, automating financial audits, incurred additional costs of $541,000, and a logistics optimization project incurred additional costs of $134,000. Thus, the total unexpected costs for the three projects reached $2.48 million.
Amazon said that these are only the experiences of individual teams and do not reflect the overall picture of the use of AI across the company. Nevertheless, the incident shows that the uncontrolled use of AI agents can pose serious financial risks for large companies.
