Amazon staff flag AI projects that ran far over budget

In short: Amazon employees say some internal AI projects have led to large, hard-to-spot cost overruns, and the company is adding controls to prevent repeat cases.

What's going on

Amazon staff have identified several cases where using AI systems caused spending to run well past what teams expected, according to a report by the Financial Times. Senior engineers told colleagues that moving work from normal software code to AI models sometimes created “unplanned” spending, and it could take a long time to notice.

One example shared internally involved matching author details with product listings on Amazon’s shopping site using Anthropic’s Claude Sonnet AI model. Employees said the effort failed, but still cost about $1.8 million, which they described as an 860 percent overrun, and it took five months to detect.

Other examples included roughly $541,000 in unexpected costs while building financial auditing tools, and $134,000 in accidental spending on a project to improve delivery speeds. Engineers said mistakes that would be cheap in older systems can become very expensive with AI, partly because some AI services charge based on usage. This is often tracked by “tokens”, which are small chunks of text or data processed, like paying per unit on an electric meter instead of a flat monthly fee.

Amazon said it is “experimenting, learning and improving” how it uses AI, including cost controls. It also said these were small, isolated examples and not typical of how teams use AI across the company.

What to watch

More companies are shifting from flat fees to usage-based AI pricing, which makes it easier for bills to spike if something is set up wrong. Watch for Amazon and others to add stronger spending limits and automatic alerts, like a credit card warning that triggers when a purchase looks unusual.

Source: Financial Times