The use of AI is intended to accelerate software development, but within Amazon, it sometimes turns out to be an unexpected expense. Internal examples show that projects significantly exceeded their budgets because developers lacked sufficient insight into the costs of AI models and AI agents. The company is therefore working on stricter control measures.
According to the Financial Times, senior engineers discussed several incidents this week during an internal meeting in which AI projects turned out to be much more expensive than anticipated. Amazon is now developing automated controls designed to prevent projects from spiraling out of control unnoticed.
One employee succinctly summarized the problem: “It’s hard to determine how much something related to AI actually costs.”
Failed project costs $1.8 million
One example involved a project in which Anthropic’s Claude Sonnet model was supposed to link author data to product listings in Amazon’s online store. Although the implementation failed, the bill totaled approximately $1.8 million, an 860 percent budget overrun. Notably, according to the internal presentation, it took five months before the additional costs were discovered.
Other projects also spiraled out of control financially. The development of software for financial audits resulted in unexpected additional costs of over $541,000. An AI project aimed at improving the delivery performance of the logistics network led to another $134,000 in unintended expenses. That discrepancy was not discovered until more than two weeks later.
According to Tom’s Hardware, costs rose primarily because AI agents performed tasks autonomously and consumed far more tokens in the process than traditional AI applications. According to the engineers, errors that cost virtually nothing in traditional software development can have major financial consequences in AI workloads.
During the meeting, it was noted that programming errors that used to “cost virtually nothing” can turn out to be “catastrophically expensive” when using AI models. Furthermore, the incidents discussed are reportedly not isolated cases.
Amazon responds that it continues to improve its use of AI. “As with any new technology, we are constantly experimenting, learning, and improving how we use it, including ways to manage costs more efficiently.”
At the same time, the company emphasizes that the examples come from only a small number of teams and do not provide a representative picture of AI usage within Amazon. “Highlighting a few small, isolated examples in which teams learn from one another and presenting them as if this were the norm does not accurately reflect how teams within Amazon use AI.”
Although the amounts seem substantial, they are relatively modest for Amazon. The company generated revenue of over $181 billion last quarter.
AI pricing is changing rapidly
These incidents illustrate a broader trend. More and more organizations are struggling to keep AI costs under control, partly because suppliers are switching to pricing models based on per-token billing. As a result, programming errors or inefficient AI agents can more quickly lead to unexpectedly high bills.
Earlier this year, Amazon discontinued an internal ranking system based on the use of its Kiro development platform. According to the Financial Times, this led to so-called “tokenmaxxing,” in which employees increased their AI usage to rank higher on the list. This, in turn, drove up usage costs. Amazon stated that the initiative had been launched with good intentions but ultimately backfired.
Moreover, this is not the first time Amazon has encountered unexpected consequences of AI development. Earlier this year, AI coding agents caused multiple outages within AWS. The company subsequently restricted the permissions of these AI agents to mitigate both security risks and costly errors.
Also read: Atlassian gives employees “AI wallets” to curb tokenmaxxing