Who pays the bill for your AI agents?

Business software used to be sold by the seat. An AI agent is billed for what it does. That shift looks technical, and it leaves a variable expense without an owner in most organizations.

For twenty years, business software was bought the same way: a price per user, per month. You counted the seats, you signed, and the year's spending was known from day one. An AI agent is not billed that way. It is billed for what it does — for the action taken, the message handled, the conversation held, sometimes the result obtained. The difference sounds technical. What it really does is shift a responsibility that, in most companies, nobody has picked up yet.

The invoice line has changed in nature

A seat can be counted. A task cannot.

The price sheets published since 2025 all tell the same story. Some vendors now bill by the action: the agent looks up a file, drafts a reply, updates a record, and every move consumes a credit bought in advance. Others bill by the message, at a cost that can double depending on how complex the request is. Others bill by the result: nothing if the conversation goes nowhere, a flat amount if the customer's problem is solved.

None of these formulas is dishonest. They are simply variable, while a small company's software budget is built for fixed costs.

What this means for an owner

A fixed cost is negotiated once a year. A variable cost is steered continuously — or discovered at the end of the month.

DoiT's AI Spending Survey, conducted among 500 finance executives at companies with more than a thousand employees and published in June 2026, measures the scale of it: 79% went over their AI budget in twelve months. The more surprising figure is the next one: overruns are even more common at companies with a mature cloud cost management practice. They do not spend less. They see it better.

KPMG's Global AI Pulse survey, published in June 2026 among 2,145 executives across twenty markets, lights up the other half of the problem: about a quarter of organizations report full, real-time visibility into their AI costs. The rest learn the amount after the fact.

Where the money actually goes

There is one lever almost nobody watches, and it is enormous.

At a single model provider, output pricing today ranges from roughly $0.20 to $50 per million tokens — the unit of text these providers bill for — depending on the model chosen. Two hundred and fifty times the spread, for a task that may be identical. That choice is usually made in the code, by someone on the technical team, on a Tuesday afternoon. It appears nowhere in a budget.

Uber's case, reported in June 2026, shows what comes next. The company burned through its annual AI budget in four months, then capped spending at $1,500 per employee, per tool, per month. Its chief operating officer summed up the difficulty in a sentence that holds for an organization of any size: the link between what is spent and the value delivered has not been established yet.

Worth keeping in mind: the unit price of artificial intelligence has been falling for three years, and falling sharply. The invoice is going up. That is not a contradiction — it is what happens when the billed unit changes and the volume explodes.

The new problem: nobody owns this budget

This is the most uncomfortable conclusion in the available data, and it holds from one study to the next.

In DoiT's survey, technology claims ownership of the AI budget 55% of the time, finance 53% — a total above one hundred, which means both functions believe they are in charge. A report published by Harness in July 2026, among 700 engineering leaders, is blunter still: 52% say no cost owner is clearly designated, 72% have been hit by an unexpected billing spike, and more than half plan their AI spending by guesswork.

Gartner forecast in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, naming escalating costs as the leading cause. That is not a prediction about technology. It is a prediction about financial governance.

The Belarel perspective

A company that hires does not just hand over an access badge and a computer. It hands over a mandate: here is your role, here is what you may commit, here is who you ask when you go past it.

An AI agent today gets the access, and rarely the mandate.

An agent without a budget is an employee without a mandate. It works, it produces, it spends — and the question of how far it was allowed to go comes up only once the invoice has arrived. Cost control is not an accounting refinement to be added later: it is the same question as permissions, asked in dollars instead of access rights. Who may spend how much, on what, and who is warned when the threshold gets close.

This is an operations question, not a finance question. It belongs in the same place as the company's other rules.

Three questions to ask this week

An owner does not need to understand model pricing to take back control. They need three answers.

Who, by name, is responsible for AI spending? If two functions answer "we are," the answer is "nobody."

What is the ceiling, and what happens when it is reached? A ceiling with no defined consequence is not a ceiling, it is a wish.

Do we know what one task costs? Not the monthly total: the cost of one customer follow-up, one email handled, one file reviewed. Until that number exists, there is no way to say whether an agent pays for itself — and no way to decide whether to hand it more.

The day those three answers exist, AI stops being a cost you absorb and becomes a cost you choose. That is a change of posture more than a change of tool.