Half a billion dollars in one month and no one saw it coming

Axios reported it without much drama. a company ran up roughly $500 million in AI-related costs inside a single month. The number is large enough to feel abstract, but the mechanism behind it is not. No one had set a limit. No one had mapped which teams were running what models at what volume. No one had made AI spend anyone's explicit job.

That is the actual story. Not the number. The number is just what happens when the structure is missing.

This is the pattern we are watching across organisations right now. AI tooling gets adopted fast, often at the team or individual level, and the financial and operational consequences get understood slowly, if at all. By the time someone in finance asks a question, the exposure is already real.

How it starts

Most AI spend begins sensibly. A product team adds a coding assistant. A content team experiments with a generation tool. Someone in ops connects an API to automate a workflow that used to take three hours a week. Each decision is reasonable. Each cost is small.

The problem is structural, not individual. These decisions get made in parallel, across departments, without a shared view of accumulation. Usage-based pricing means the bill scales with behaviour, not with a fixed contract. If behaviour changes fast, and it does when people find something useful, the cost curve changes fast too.

Coding assistants are a useful case here. Tools like Claude Code, GitHub Copilot, and similar products are priced per token or per seat, depending on plan. Token-based usage in particular can spike sharply when developers work on large codebases or run automated agents in loops. A single misconfigured agent job, with no spending cap set, can run continuously and generate a bill that looks nothing like the estimate someone made in a kick-off meeting.

The governance gap is real, and it is wide

Most organisations have procurement processes for software. SaaS contracts go through legal. Infrastructure costs sit inside engineering budgets with someone accountable. AI tooling, particularly in its current form, often falls through the gap between those two categories.

It is not classic software, because the cost is not fixed. It is not classic infrastructure, because it is not always managed by engineering. It lives in a grey zone: adopted by teams who want to move fast, invoiced by vendors who have no incentive to slow things down, and reviewed by finance teams who do not yet have the vocabulary to interrogate the bill.

A Gartner forecast from 2023 estimated that by 2026, more than 80 percent of enterprises would have used generative AI APIs or models in production environments. The pace of adoption is not in question. What lags behind is the management layer.

For agencies and studios, the risk is slightly different but equally real. We are often the ones embedding AI into client workflows, building on top of model APIs, and making architecture decisions that have long-term cost implications. If we do not understand the spend profile of what we build, we are creating problems we will eventually have to explain.

What good ownership actually looks like

Owning the AI budget does not require a new department or a 40-slide framework. It requires a few decisions that most organisations have not made yet.

First, someone needs to be named. Not a committee. One person or one role that has visibility across AI tooling spend, with the authority to ask questions and set limits. In a smaller organisation that is probably the CTO or a senior product lead. In a larger one it may need to be a dedicated function, but it starts with a name on a responsibility.

Second, token and usage limits need to be set at the infrastructure level, not the honour system. Most major model providers offer hard spending caps. Most organisations have not enabled them. Enabling a cap is a ten-minute task. Not enabling it is a choice that compounds over time.

Third, the adoption map needs to exist. Which teams are using which tools, on what pricing model, with what expected usage? This does not need to be elaborate. A shared spreadsheet that gets reviewed monthly is more useful than an unreviewed dashboard. The point is that someone has looked at it recently.

Fourth, automated agent workflows need a category of their own. Human-in-the-loop usage is generally predictable. Agents running in loops are not. Any workflow where a model is calling other models, or where a process runs on a schedule without human review at each step, deserves a separate line of scrutiny. These are the scenarios where costs can compound quietly and quickly.

Why this matters for agencies specifically

Agencies are in an interesting position. We adopt fast because clients expect it and because good work increasingly depends on it. But we also carry a particular kind of accountability: we make recommendations that other organisations act on. If we build a system that has unpredictable cost behaviour and we do not flag it, we own part of that outcome.

There is also a commercial angle. Agencies building on top of model APIs often pass usage costs through to clients, sometimes on a markup, sometimes as a pass-through. If those costs are not monitored, the invoice to the client can surprise everyone, including the agency. That is not a good moment to discover the pricing model was not well understood.

The organisations we respect most in this space are not the ones moving fastest. They are the ones that have built a clear view of what they are spending, why, and what they are getting back. That clarity does not slow down good work. It is what makes good work sustainable.

The $500 million story will not be the last one. The numbers will vary, but the structure behind them will be the same: fast adoption, slow governance, and a bill that arrived before the question was asked. The question is worth asking now, while the answer is still manageable.

Where to start

If you are reading this and your organisation does not have clear answers to three questions, start there.

Who owns AI spend? Not who uses AI tools. Who is accountable for the total number?

What are your hard limits? Not policies or guidelines. Actual caps set in the provider console.

What is running without human review? Automated workflows, scheduled agents, anything that calls a model without a person checking it first.

Those three questions will not solve everything. But they will surface the gaps that matter, and they will do it before the bill does.