Half a billion dollars. One month. One company.

Axios reported it in late May: an AI consultant watched one of their clients run up roughly $500 million in Claude charges in a single month. The company hasn't been named, and the number is big enough that it has to be one of the largest corporations on earth. But the cause wasn't exotic. They handed out AI licenses to employees and never set a spending limit on any of them.

That's it. No cap, no budget, no visibility. Thousands of people with an expensive tool and no reason to think about what it cost.

How a number gets that big

The headline makes it sound like a freak accident. It isn't. It's the predictable result of two things stacking up.

The first is that the new generation of AI tools costs wildly more than the old one. A quick question to a chatbot is cheap. An agentic tool — the kind that reads your whole codebase, runs tasks in a loop, spins up parallel sessions — can burn through a thousand times the tokens for a single job. One engineer running those workflows all day can rack up hundreds or thousands of dollars a month without doing anything wrong.

The second is that nobody felt the meter running. One CTO told Axios their employees were using a frontier AI model to check the weather. Not because they were reckless. The tool was right there, it was free to them, and nobody had ever suggested otherwise. Multiply that across a workforce and the math turns ugly fast.

And this isn't one company being dumb. Uber reportedly blew through its entire 2026 budget for Claude Code by April. Microsoft is moving some of its engineers off Claude Code and onto a cheaper option. The biggest, most technically sophisticated companies in the world are the ones getting surprised by these bills. These are smart companies. They just skipped the setup.

The same mistake, smaller numbers

You're not going to spend $500 million. You might spend $2,000 next month and have no idea why.

The shape of the mistake is identical at every scale. You sign up for a few AI subscriptions, or you put an API key into some tool a contractor set up, and then you stop watching. Three months later the card statement has a charge you can't explain, and you can't tell whether it bought you anything. Maybe an employee is running an automation in a loop that fires a thousand times a day. Maybe a tool defaults to the most expensive model when a cheaper one would do the same job. Maybe it's all working fine and you just have no way to know.

The enterprise version is a half-billion-dollar headline. The small-business version is quieter and a lot more common: paying for AI nobody is accountable for, with no way to see whether it's earning its keep.

It isn't just the giants

The half-billion number is the extreme, but the same thing is happening at every size, and it has been for a while.

A developer left a Google Gemini API key exposed and watched a routine bill balloon to more than $82,000 after someone found the key and ran it hard. Another reported on Hacker News that a leaked Anthropic key ran up $10,000 before they noticed. These aren't enterprise budgets. They're individuals who left one door open.

And it happens with no theft at all. When the AI coding tool Cursor changed how it billed last year, some users drained a month of usage in a single day, because newer models quietly spend far more tokens per task than the ones people had budgeted around. The tool worked exactly as designed. The cost just wasn't visible until the statement showed up.

Different tools, same root cause every time: usage nobody capped and nobody was watching.

The four things that keep it from happening

None of this is a reason to stay away from AI. It's a reason to set it up like you'd set up anything else that touches your bank account. Four things do most of the work.

Put a hard cap on every account. Most paid AI tools and API keys will let you set a spending limit. Use it. Pick a monthly ceiling that would make you raise an eyebrow if you hit it, and set the account to stop or alert before it gets there. Know the difference between a cap and an alert, too: OpenAI and Anthropic let you set a hard ceiling that actually stops at the number you pick, while some platforms (Google Cloud among them) only warn you as the charges keep climbing. The company in the story had a cap available the whole time. They just never turned it on.

Match the tool to the job. The most expensive model is not the right answer for most tasks. Drafting a review reply, sorting an inbox, pulling a name off a form: these run fine on cheap, fast models. You save the heavy, expensive ones for the work that actually needs them. A setup that reaches for the priciest option every time is the software equivalent of using a frontier model to check the weather.

Know what you're being billed for. You should be able to answer "what did AI cost me last month, and what did it do" in under a minute. That means usage you can actually see: a dashboard, a report, a number that maps to a result. If the spend is invisible, it's already out of control; you just haven't gotten the statement yet.

Give every workflow an owner. The $500 million happened because the cost belonged to everyone, which means it belonged to no one. Each AI tool running in your business should have one person who knows it exists, knows roughly what it costs, and would notice if that number doubled. That's most of the protection right there.

The actual lesson

It would be easy to read this story as "AI is a money pit, stay away." That's the wrong takeaway. The company didn't get burned by Claude. They got burned by deploying a powerful tool with zero guardrails and then not looking at it for thirty days.

A well-set-up AI workflow is affordable, and what you get back from it is obvious. The ones we build for service businesses around Phoenix run on a few dollars a month, because they're scoped to a specific job, pointed at the right-sized model, and capped so they can't run away. The cost is boring and predictable, and it's plainly worth more than it spends. Expensive AI is almost always uncontrolled AI.

The difference between the two isn't the technology. It's whether anyone set it up on purpose.

Want AI working in your business without the surprise bill? The AI Audit maps your most-repeated workflows and shows exactly where AI pays for itself, and where it doesn't. One week, $750 fixed fee, with cost estimates in plain numbers. If a workflow isn't worth the spend, the report says so.

Still figuring out which tool fits your operation? Here's our breakdown of Claude, ChatGPT, and Gemini for small service businesses →

Want to try a scoped, low-cost piece yourself first? The free Claude agents pack is three ready-to-use projects: inbox triage, follow-up nudger, and review replies, each pointed at a specific job instead of left to run wide open.

When you're ready for someone to build and run it for you — caps, monitoring, and all — that's the AI Concierge.