Amazon sees AI projects go off the rails due to unexpected costs

Amazon’s internal AI projects ran far over budget. One failed Claude Sonnet effort reportedly reached about $1.8 million, while other projects added hundreds of thousands of dollars in costs.

The overruns were driven by token-based billing and autonomous AI agents, which made usage and spending harder to forecast than expected.

In response, Amazon is building automated cost controls and tightening permissions for AI agents.

The cases show why companies need guardrails for AI spending, not just performance targets.