Thought Leadership
Internal Amazon documents show that several AI projects at the company ran significantly over budget. The most expensive example involved a never-deployed project using Anthropic’s Claude Sonnet model.
According to the Financial Times, citing an internal presentation, the project incurred costs of 1.8 million US dollars, exceeding its original budget by 860 percent. The overrun went undetected for around five months, and the project was ultimately never rolled out. Engineers reportedly described the incident internally as “catastrophically expensive.” What used to cause comparatively low costs in the case of faulty programs can now become significantly more expensive through the use of AI models, especially since the spread of AI agents with substantially higher token consumption.
Besides the Claude project, the report names two further cases: a tool for financial auditing caused around 541,000 US dollars in additional costs, and a system designed to shorten delivery times in logistics caused another 134,000 US dollars. In total, the three documented cases amount to around 2.5 million US dollars in unplanned spending.
Amazon downplays the scale of the incidents
In an internal presentation, Amazon stated, according to the Financial Times:
“As with any new technology, we’re experimenting, learning and improving how we use it, including how we drive cost efficiencies. Cherry-picking small, isolated examples where teams are learning from one another and portraying them as business as usual doesn’t reflect how teams across Amazon are using AI.”
Amazon
Measured against the company’s quarterly revenue of more than 181 billion US dollars, the additional expenses amount to less than 0.1 percent of a single month’s revenue. This is not the first time AI-related problems have surfaced in Amazon’s day-to-day operations. Earlier this year, AWS had already reported several outages attributed to errors by AI-powered coding tools. In response, Amazon restricted the access rights of AI agents instead of leaving them with the same permissions as the senior engineers they are tied to. The company also discontinued an internal leaderboard that had previously shown which employees used AI tools the most, after sharply rising AI costs called the usefulness of that practice into question.
(red)
