In brief: Amazon has incurred massive budget overruns in the implementation of Claude-based AI agents, highlighting the cost control risks of AI projects with unlimited token consumption.
An Amazon project using Anthropic’s Claude model to match author information with product listings overran its budget by 860 percent at a cost of 1.8 million US dollars and was never delivered. The case reveals cost control risks in enterprise-wide AI agent deployment.
According to internal documents obtained by the Financial Times, the project using Anthropic’s Claude Sonnet model incurred costs of 1.8 million US dollars instead of the original budget of approximately 200,000 US dollars. The budget overrun remained undetected for approximately five months. Developers internally referred to the project as “catastrophically expensive.” Beyond this case, Amazon has documented two additional problematic AI projects: a financial audit tool with 541,000 US dollars in additional costs and a logistics system with 134,000 US dollars in cost overruns. In total, these three cases amount to approximately 2.5 million US dollars in unplanned expenditures.
The central difference from previous software development lies in token consumption: while faulty or inefficient traditional programs previously incurred comparatively low additional costs, operating AI agents with continuous API usage leads to rapidly exponentially increasing costs. This economic dynamic requires more precise cost forecasting and ongoing expense control than traditional development processes.
Amazon downplays the extent of the incidents in an internal memorandum: the total expenditure of 2.5 million US dollars represents less than 0.1 percent of a single month’s revenue with quarterly volumes exceeding 181 billion US dollars. The company emphasizes that these cases are learning examples from an experimental phase, not the norm. Nevertheless, Amazon has already initiated concrete measures: following multiple failures of AI-powered programming tools this year, the company restricted access permissions for AI agents and ended an internal ranking that mapped the most intensive AI tool users among employees.
Source: www.it-daily.net · Published July 31, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.