Budgeting for AI that bills by the task

After rolling Claude Code out to roughly 5,000 engineers, Uber burned through its entire 2026 AI budget in four months, with some engineers running up bills of $500 to $2,000 a month. Its chief operating officer called it a head-exploding moment and began asking whether the spend was worth it. Most firms reading this will never use AI at that scale, but the mechanism that caught Uber out is heading for everyone. AI is moving off the fixed per-seat licence you can put in a budget and forget, onto pricing that rises with how much the tool actually does. If you are about to switch on an agent that runs whole tasks, the useful move is to find out how it charges, set a spending cap, and decide which work is worth a variable cost, before the bill arrives rather than after.

What is actually changing

For years, business software has been sold by the seat. You pay a fixed amount per person per month, you know the number a year ahead, and finance can plan around it. Most AI arrived the same way, as a per-user licence bolted onto tools you already pay for. That model is starting to break, for a simple reason. When AI stops helping a person with a task and starts running the whole task itself, it no longer maps to a headcount. An agent does not log in as one of your staff. It does work, and the amount of work is what costs money.

You can see the shift in Microsoft's own products. In June it made Copilot Cowork generally available, an agent inside Microsoft 365 that takes a goal and completes a multi-step job across Outlook, Teams and Excel rather than drafting a line for a person to finish. It is billed by usage, in credits, not only by the seat. So the cost of AI inside the tools your team already uses now moves with how much the agent does, and a busy month can cost more than a quiet one.

Why the bill is hard to forecast

What unsettles a budget is that usage does not rise in a straight line. An agent working through a task on its own can call a model many times, check its own output, and loop until the job is done, so an involved task can cost many times a simple one. Multiply that by a team that finds the tool genuinely useful, and the total climbs in a way a fixed licence never could. That is roughly what happened at Uber. Not waste exactly, but capable people using a capable tool more and more, with no ceiling on what it added up to.

It would be easy to read that and decide it is a big-company problem. The opposite is closer to the truth. A firm of five thousand engineers has the headroom to absorb a surprise and the finance team to catch it early. A smaller professional-services firm runs on tighter, more forecastable numbers, where an unplanned few thousand pounds in a quarter is the difference between a budget that holds and one that does not. The scale is smaller, but so is the cushion. A variable AI bill is harder to live with, not easier, when every line is already accounted for.

The trap to avoid

In our work with firms turning on their first agent, the pattern repeats. The tool gets switched on for everyone, people use it because it helps, and the bill is read at the end of the month. With per-seat software that was fine, because the number could not move. With usage-based AI it is exactly how costs run away. The spend does not announce itself. It accumulates across dozens of small jobs that each felt free in the moment, and the total only becomes visible weeks later, by which point the habit is set and the invoice is real.

What should a leader actually do?

None of this is an argument against agents. Pointed at the right work, they earn their cost easily. It is an argument for deciding the money question on purpose rather than by accident. Four concrete steps.

Find out how the tool charges before you switch it on. Per seat, per task, per credit, per result: ask the vendor exactly what triggers a charge and what a typical job costs, and do not accept a single per-seat figure for something that bills by usage. If you are weighing up a new tool at all, it is worth running it past the questions worth asking before any new AI tool first.

Set a spending cap, and set it low to start. Most serious AI tools let an administrator put a ceiling on spend per user, per team or per month. Turn that on before you roll the tool out, not after the first surprise. A cap turns an open-ended bill into a decision you make deliberately when you choose to raise it.

Match the spend to the value, one use case at a time. The question is not what AI costs in total, but whether a given task is worth what the agent charges to do it. A first-pass research note that saves someone an afternoon may be worth a lot of small charges. An agent running across every inbox because it is there is not. This is the same discipline as judging an AI use by the business value it returns rather than by how busy it looks.

Keep a person on the judgement-heavy work. The tasks worth handing fully to an agent are usually the repeatable, checkable ones. The advice, the client call and the final sign-off stay with someone who can answer for them, which also keeps the most expensive open-ended agent loops off your most sensitive work. It is the same boundary the agent-in-your-tools question turns on.

Working out which tasks justify a per-task cost, and which are better left to a person or a cheaper fixed tool, is the kind of decision the Good Transformer Lessons for Leaders sessions are built around. Not which tool to buy, but how to spend on AI so the money tracks the value. If that is worth an hour for your firm, book a discovery call.

The decision in front of you is small and specific. Before the next agent goes live in your firm, find out exactly how it charges, put a cap on it, and name the one or two tasks it is worth paying by the task to do. Do that, and usage-based pricing becomes a lever you control. Skip it, and you learn what your firm spends on AI the way Uber did, a month too late.

FAQ

Why is AI moving away from per-seat pricing?

Because agents do not map to headcount. Per-seat pricing assumes one licence per person using the software. An AI agent runs tasks rather than logging in as a user, and the cost depends on how much work it does, not how many people are signed up. As tools shift from helping a person to completing whole jobs, vendors increasingly charge by usage, by task or by result. Microsoft's Copilot Cowork, billed by usage in credits, is one example of the model reaching everyday office software.

How do you stop AI costs running away?

Set a spending cap before you roll a tool out, not after the first large bill. Most serious AI tools let an administrator limit spend per user, per team or per month. Then review usage against value for each use case, rather than looking only at the total, so you can see which tasks are worth what the agent charges and which are not. The aim is to make any rise in spend a decision you take on purpose.

Is usage-based AI pricing more expensive than per-seat?

Not inherently. For light or occasional use it can be cheaper, because you pay only for what you use rather than a flat fee per head. The difficulty is predictability: a usage-based bill rises and falls with activity, so a busy period costs more and the total is harder to forecast. Whether it works out cheaper depends on how much your team uses the tool, and whether that use returns enough value to justify it.

Should a small firm avoid agentic AI because the cost is unpredictable?

No. The unpredictability is a reason to put controls in place, not to stay out. Agents can do real work that justifies their cost when they are pointed at the right tasks. The sensible approach is to start narrow, on one or two repeatable jobs, with a spending cap set low, and to widen only once you can see the cost and the value clearly. That keeps the benefit while removing the risk of a surprise bill.