A misconfigured AI model does not stop when it goes wrong. It keeps running, keeps generating tokens, and quietly converts a trivial mistake into a seven-figure invoice. Amazon’s own internal metrics, reviewed by engineers at a staff meeting on July 28, show the pattern in unusually concrete detail.
The largest example was a tool built on Anthropic’s Claude Sonnet to match author records against product listings on Amazon’s retail site. The project consumed $1.8 million, ran 860 percent over its original budget, went undetected for five months, and was never launched. The Financial Times first reported the case, which surfaced alongside two smaller overruns: a financial auditing tool that came in roughly $541,000 over plan and a logistics system meant to speed up deliveries that added about $134,000. Together, the three projects accounted for roughly $2.5 million in unplanned AI spending.
Why token billing turns small errors into large bills
The root cause is pricing mechanics. AI providers have largely moved away from flat subscription fees toward billing per token, the unit of text a model processes. A human engineer who writes buggy code gets an error and fixes it. An AI agent with a bad configuration gets no error at all: a retry loop that re-sends the same prompt, or a job pointed at an entire catalog instead of a sample, simply produces more compute activity and a larger invoice. Amazon’s engineers described the resulting overruns in internal materials as catastrophically expensive mistakes that used to cost almost nothing to fix.
Monthly billing cycles amplify the problem. A runaway job that would be obvious in a build log or a real-time dashboard becomes invisible on an invoice that arrives weeks later, which is how the $1.8 million project survived five months without scrutiny.
The awkward part: Amazon sells the fix
The company that overspent on AI also happens to run the cloud platform that sells the cost controls designed to prevent exactly this. AWS Bedrock offers batch inference at half price, a discounted Flex tier, prompt caching at a tenth of the standard input rate, and prompt routing that pushes simple tasks to cheaper models. Anthropic’s own Haiku model costs roughly a third of the Sonnet model the failed project used. The overrun came from defaulting to a frontier model with guardrails switched off.
Amazon’s public response framed the incidents as isolated. The cases involved a small number of teams within a corporate workforce of roughly 300,000, and the company described them as part of the normal process of experimenting with new technology. Internally, the reaction has been more concrete. Amazon has discontinued KiroRank, an informal leaderboard that ranked developers by AI token usage, after employees inflated their consumption to climb the standings, a practice workers called tokenmaxxing. Engineers are now building automated guardrails, peer review requirements, and spending caps so runaway projects are caught before the invoice lands.
The blunders follow earlier incidents involving Amazon’s own AI coding assistants. In December, the Kiro assistant was authorized to make changes to Cost Explorer, an internal tool AWS customers use to review their billing, and instead deleted and recreated the environment it was working in, causing a 13-hour outage affecting parts of mainland China. Amazon attributed that incident to user access controls rather than AI autonomy, but both disruptions shared a feature: the AI tools carried the same permissions as the engineers operating them.
A governance problem, not a money problem
For Amazon, the sums are rounding errors against quarterly revenue above $180 billion. The same failure mode is far more dangerous for smaller companies. Uber, which relies heavily on Claude Code, had already exhausted its entire 2026 budget for the tool by April, and its leadership acknowledged publicly that the company had not yet established a clear connection between rising token usage and features customers actually benefited from.
The lesson of the Amazon case is governance, not technology. AI cost overruns are silent by design, and the companies that contain them will be the ones that treat spending caps, model selection, and usage review as part of the engineering process rather than an afterthought.
Sources: Amazon finds cases of AI causing runaway spending on tech projects (Financial Times, July 30); Amazon accidentally spent $1.8 million using Claude for menial coding task, went 860% over budget (Tom’s Hardware, July 30); Amazon blew $1.8 million on a single Claude AI task and didn’t notice for five months (BetaNews, July 30); Amazon spent $1.8m on a Claude job that failed, and it sells the fix (The Next Web, July 30)