Amazon faced an unexpected expense item: AI agents led to millions in losses

Recent internal Amazon reports obtained by Financial Times reveal rather clumsy mistakes in the use of advanced artificial intelligence (AI) models, including overspending on the powerful Claude tool. The result was an unexpected budget “bloat,” with a colossal $1.8 million spent on routine programming tasks.

AI cost inflation: from minor errors to catastrophic consequences

Historically, errors in software development systems have generally not led to significant financial losses. However, with the emergence of newer, more powerful AI models, the cost of such failures has risen sharply. This became especially noticeable after the introduction of so-called AI agents, which greatly expanded capabilities and automation. This led to a sharp spike in spending on so-called “tokens” — units of information for which processing is charged.

In the case of Claude Sonnet, which was tasked with matching author data to advertisements on Amazon’s massive platform, a flawed implementation caused the assigned budget to be exceeded by 860%. The shocking fact is that the problem went undetected for five months, accumulating a huge debt in the meantime.

This is not an isolated case. According to the published data, the company also spent an additional $541,000 on a project aimed at developing financial auditing, as well as $134,000 on a system designed to optimize logistics chains and speed up product delivery. These incidents highlight the growing challenges associated with integrating advanced AI solutions and controlling their costs.

Amazon: “We are learning and experimenting”

In response to the published information, Amazon representatives said the company is actively experimenting with new technologies, constantly learning and improving the ways it uses them. “As with any new technology, we experiment, learn, and improve how we use it, including how we improve cost efficiency. Selective examples of teams learning from one another, and presenting them as standard practice, do not reflect how Amazon teams use AI,” the company’s internal presentation stated.

Although millions of dollars spent on failed AI projects may seem significant, it is worth noting that they represent only a tiny fraction of Amazon’s overall financial performance. The company’s latest quarterly revenue exceeded $181 billion, making these losses less than 0.1% of monthly income.

The trend of uncontrolled AI spending

These incidents are not Amazon’s first AI-related problems. Earlier this year, the company faced disruptions caused by errors in the code of AI bots. The problem was then solved by limiting the permissions of AI agents rather than giving them powers equal to senior engineers.

Amazon previously had a system for ranking employees’ use of AI, but it was discontinued. Rising AI costs forced the company to reconsider its policy. Many tech giants actively encourage employees to use AI as much as possible to boost productivity, which in turn leads to maximum token consumption.

As experts note, this policy does not always correlate with successful product delivery. Uber’s CTO, for example, points to the lack of a direct link between active AI use and project success. The situation is further complicated by the fact that many AI providers have moved away from subscription models to per-token pricing. This has led to situations where companies’ annual budgets can be exhausted in just a few weeks.

For tech giants such as Amazon or Microsoft, these costs, while unfortunate, may not be critical. However, for smaller companies that do not have such substantial financial resources, this trend becomes extremely dangerous and can threaten their very existence. This underscores the need to develop more effective mechanisms for controlling and optimizing spending in artificial intelligence to ensure its sustainable and safe development for the entire technology sector.