Amazon has learned the hard way that artificial intelligence can turn a small configuration mistake into a seven-figure loss. The company's own internal metrics show multiple AI projects running far beyond budget, including one that burned $1.8m before anyone noticed. Senior engineers reportedly described the results as "catastrophically expensive." The details were shared in an internal meeting, and they offer a warning for every organization now rushing to adopt generative AI.
The $1.8m Claude failure
The most serious case involved Amazon using Anthropic's Claude model to match author records against product listings in its store. The job went 860 percent over budget, ran for five months before the cost was spotted, and then failed to deploy successfully. In other words, Amazon received a massive bill and still did not get the working system it wanted.
Two other internal projects also overshot. A financial auditing tool exceeded its budget by about $541,000. A system intended to speed up deliveries added $134,000 in extra AI costs. These are not the kinds of expenses that would have appeared in traditional software development. A regular program with a bug would crash or throw an error, but an AI model keeps working, sending prompts and tokens, quietly generating charges every step of the way. The bigger the task, the bigger the bill, and the harder it is to notice when something is wrong.
Why AI mistakes do not stop themselves
The core issue is a mismatch between how ordinary software fails and how AI systems fail. In a conventional codebase, a badly written function will return an error or stop execution. In an AI pipeline, a bad configuration simply continues. A retry loop that sends the same prompt again and again, or a job pointed at an entire catalogue instead of a small test sample, does not crash. It just produces tokens, and tokens have prices.
The billing also arrives on a different cycle. Build logs show failed jobs immediately. AI costs arrive on a monthly invoice, so teams can go for weeks or months without noticing that a model has been running in the background. The shift to per-token pricing makes this worse. AI agents, which generate far more tokens than a normal chatbot, can multiply the problem. A single agent loop can invoke dozens of model calls in a minute, and each call adds to the total. When an agent is asked to complete a broad task, it may search, re-read, correct itself, and try again, all before the user realizes that no answer is coming.
The problem is not unique to Amazon. One developer reportedly received a $1.3m OpenAI bill from a runaway AI process. Enterprise AI budgets are rising even as the price per token falls, because the amount of tokens consumed increases faster than prices decline. Companies are discovering that model costs are not something that can be left unsupervised. Some are adding dedicated FinOps teams for AI. Others are building custom software to monitor token usage in real time. The technology is still young, and the best practices for managing its costs have not yet caught up with the speed of adoption.
Amazon sells the fix on AWS
What makes Amazon's example especially awkward is that its cloud division offers tools designed to prevent exactly this kind of overspending. AWS Bedrock includes batch inference at half the on-demand price, a discounted Flex tier for workloads that can wait, prompt caching that cuts input costs by up to 90%, and prompt routing that sends simple requests to cheaper models. These are straightforward levers that can reduce AI bills without sacrificing quality. Many teams ignore them because they are not obvious, or because no one has set up a policy that requires their use.
There was also a cheaper model available. Anthropic's Haiku costs about a third as much as the Sonnet model that Amazon used for the failed job. Choosing the frontier model by default, and leaving cost controls disabled, is the opposite of the guidance Amazon and Anthropic give to customers. In many cases, a large model is not even necessary. A small model can handle routine classification, extraction, or matching tasks at a fraction of the cost. Performance may be slightly lower, but the cost difference can be dramatic. For an internal tool that matches records, a compact model may be more than adequate.
The lesson from Amazon's overruns is not that AI is too expensive. It
Source: TNW | Amazon News