Amazon accidentally spent $1.8 million using Claude for menial coding task, went 860% over budget — 'catastrophically expensive' coding blunders discovered in internal Amazon AI usage metrics

Amazon recently faced a significant budget overrun after an internal AI project spiraled out of control. The company spent $1.8 million on a task involving the Claude AI model that was meant to match author details with listings. This specific deployment went 860 percent over its initial budget. The issue remained undetected for five months, highlighting the hidden costs associated with scaling automated agents in a large corporate environment.

Internal reports reveal that this is not an isolated incident. Other projects saw similar financial strain, including a half-million dollar expense for a financial auditing tool and over $100,000 for a logistics system project. These figures represent a shift from the past where such technical mistakes were trivial to correct. Now, the high demand for tokens required by these automated agents makes even minor errors catastrophically expensive for the balance sheet.

Amazon characterizes these overruns as part of the normal learning curve that accompanies the adoption of new technology. The company argues that internal teams are actively finding ways to improve cost efficiency. Despite the high dollar amounts, Amazon notes these losses are small relative to its massive quarterly revenue. The tech giant is now refining how it grants permissions and monitors usage for AI agents to prevent further waste.

This incident adds to a broader conversation in the tech industry regarding the sustainability of current AI spending models. As companies pivot toward usage-based billing, many are finding that unchecked automated tasks drain annual budgets in a matter of weeks. The reliance on agentic AI has forced a rethink of how major corporations monitor and control their software expenditures.