Key facts at a glance

• Amazon's internal metrics revealed several AI projects running far over budget.
• The worst case: an Anthropic Claude model used to match author records against product listings burned $1.8m before being noticed — 860% over budget and ultimately failed.
• A financial auditing tool overshot by about $541,000, and a delivery optimization system added $134,000.
• Bad configurations in AI systems don't crash; they keep running and generating token charges, often invisible until the monthly bill arrives.
• AWS, Amazon's cloud arm, sells tools like batch inference, prompt caching, and prompt routing that could have prevented the overspend.
• Amazon downplayed the incidents but acknowledged it is building automated guardrails and has scrapped an internal AI usage leaderboard that employees gamed.

Amazon has discovered that artificial intelligence can turn a trivial mistake into a seven-figure one. The company's own internal metrics show projects running wildly over budget, including one that burned $1.8m before anyone noticed. Senior engineers called the results “catastrophically expensive.” The cases were laid out at an internal meeting and first reported by the Financial Times.

The worst offender used Anthropic's Claude to match author records against product listings on Amazon's store. It ran 860 per cent over budget, took five months to spot, and the deployment failed anyway. Two smaller blunders sat alongside it, as noted by Tom's Hardware. A financial auditing tool overshot by about $541,000. A system meant to speed up deliveries added $134,000. Each was the kind of error that used to cost almost nothing.

How a cheap mistake becomes a huge bill

The reason sits on the price list. When a person writes buggy code, it throws an error and crashes. When a model does the work, a bad configuration just keeps running and quietly bills you. A retry loop that re-sends the same prompt, or a job pointed at the whole catalogue instead of a sample, produces no crash. It produces an invoice.

That invoice arrives on a monthly cycle, not in a build log. The gap is how a misconfigured job runs for five months before anyone connects it to a number. The shift to per-token pricing makes it worse. AI agents, which fire off far more tokens than a chatbot, pour fuel on it.

None of this is unique to Amazon. It is the same trap that produced a $1.3m OpenAI bill for one developer, and the reason enterprise AI bills keep climbing even as the price per token falls. The problem is not the price of a single token — it's the compounding effect of tiny mistakes amplified by scale.

Why AI errors don't behave like software errors

Traditional software fails loudly. A null pointer exception, a divide-by-zero error, or a syntax mistake stops the process immediately. The developer sees the failure, fixes the bug, and the cost is limited to the time spent debugging. AI systems, by contrast, are probabilistic and often invoked through APIs that bill per token. When a model is given the wrong instruction or the wrong input size, it doesn't throw a fit — it simply continues processing, often in a loop, generating billable tokens each time.

In Amazon's case, the Claude-powered matching job was likely configured to scan a massive dataset instead of a small sample. Or it may have been stuck in a retry loop, re-sending prompts because the model's output wasn't accepted. Without proper guardrails, such a job can run for months, silently consuming compute and tokens. By the time a human notices the anomaly on an invoice, the cost has already multiplied.

This is a fundamental difference between deterministic software and generative AI. The error handling is different, the observability is different, and the cost model is different. Most organizations are still applying traditional software governance to AI systems, which is why runaway costs are becoming a recurring theme across the industry.

The company that sells the fix

Here is the awkward part. Amazon overspent on AI while its own cloud sells the exact tools to stop it. AWS's Bedrock offers batch inference at half price, a discounted “Flex” tier, prompt caching at a tenth of the input rate, and prompt routing that pushes simple jobs to cheaper models.

Cheaper models were an option too. Anthropic's Haiku costs a third of the Sonnet model Amazon reached for. The overspend came from picking the frontier model by default and leaving the guardrails off, exactly the mistake Amazon and Anthropic tell other companies to avoid.

Amazon's own cloud customers use these features to keep costs under control. Batch inference is designed for workloads that don't need real-time responses, making it ideal for back-office data matching. Prompt caching reduces the cost of repeated context, which is common in agentic workflows. Prompt routing can send simple classification tasks to small, cheap models while reserving frontier models for complex reasoning. Any one of these could have shaved a significant portion off the $1.8m bill.

Amazon's answer

Amazon played it down. “As with any new technology, we're experimenting, learning and improving how we use it,” it said, adding that a few isolated examples do not reflect how its teams work. The cases involved a small number of groups inside a corporate workforce of roughly 300,000.

The maths backs the shrug. Amazon's quarterly revenue sits above $180bn, so the blown budgets do not add up to 0.1 per cent of a single month's sales. The AWS cost expert Corey Quinn was unsympathetic, posting that “getting a bill with two commas that you weren't expecting must be so hard for you.”

Behind the scenes, the response is firmer. Amazon has already scrapped an internal leaderboard that ranked staff by AI usage, after employees gamed it by inflating their token consumption. Its engineers are now building automated guardrails to cap what a project can spend before the invoice lands.

The deeper governance lesson

The internal leaderboard incident is especially telling. Ranking employees by AI usage sounds like a good way to encourage adoption, but it created a perverse incentive. Employees discovered that generating more tokens — whether through unnecessary prompts, verbose outputs, or intentional loops — would push them up the rankings. This not only wasted money but rewarded behavior that was directly opposed to cost efficiency.

By removing the leaderboard and shifting to automated caps, Amazon is moving from a culture of superficial AI enthusiasm to one of disciplined cost management. That is the real lesson, and it is a governance one, not a financial one. For Amazon, $1.8m is a rounding error. For the thousands of firms now leaning on AI to cut costs, a runaway job that hides for five months is the difference between a budget and a crisis.

Enterprises need to adopt a different mindset when dealing with AI systems. The old rules of software budgeting — design, test, deploy, monitor — still apply, but they need to be adapted to the unique characteristics of generative AI. Per-token costs require real-time observability. Model selection should be part of the design process, not an afterthought. And guardrails should be built into the platform, not bolted on after an invoice shock.

AWS's existing tools are a good starting point, but they only help those who choose to use them. Amazon's own misstep shows that even the builder of the platform can fall into the trap of assuming a frontier model is the right default. The company is now learning what it has long told customers: AI is powerful, but it must be managed with the same rigor as any other business expense.

Source:TNW | Amazon News