Amazon spent $1.8 million on a failed AI project, and didn't notice the overrun for five months

Amazon’s internal AI projects sharply overspent, and some overruns went undetected for months.

The biggest case involved a Claude Sonnet-based tool that failed after running 860% over budget, with costs reaching $1.8 million.

Amazon says it is tightening monitoring and controls as usage-based AI pricing makes spending harder to predict.

The episode highlights how AI projects can introduce more volatile financial and operational risk than traditional software.