The Financial Times reports, citing sources, that Amazon has internally flagged several AI projects that went wildly over budget. The worst offender: a project using Claude Sonnet to match author info to product pages. It ultimately failed — but not before racking up a $1.8 million bill, 860% over budget, and going unnoticed for five months.
Another financial auditing tool inadvertently cost an extra $541,000. A logistics optimization project burned an extra $134,000, and it took over two weeks to catch.
Amazon engineers blame deployment errors and a lack of cost limits. In traditional systems, small mistakes cost next to nothing; in AI projects, they can become catastrophically expensive. The company is developing automated guardrails to prevent future overruns.
Amazon responded that these cases involve only a small number of teams and don't represent its overall AI usage. Earlier, the company also killed an internal AI usage leaderboard after employees started having agents perform pointless tasks just to climb the ranks, burning tokens.