For most teams, the first AI bill was a licence. So many seats at so much a month, the same figure every month, however much or little anyone used it. A finance team can plan around that.
The second kind of AI bill works differently, and it is arriving now. You pay for what gets used. That changes the question a manager needs to ask before anyone switches it on.
Two ways to pay for the same tool
A seat licence is a fixed cost. If the team ignores the tool for a month, you pay the same. If someone uses it all day, you also pay the same.
Usage pricing is a variable cost. You are charged for each piece of work the AI does, often measured in units nobody outside IT has heard of. A quiet month costs little. A busy month costs more. And a month in which something keeps running when nobody is watching can cost a great deal.
That last case used to be rare, because a person had to be at the keyboard for the meter to turn. It is becoming less rare, because AI tools increasingly do things on their own: start a task, call another service, set up something new to finish the job.
The argument for a hard cap
Simon Willison, a developer whose blog on practical AI is widely followed, made the case plainly on 3 October. Services that charge by usage, he argued, should come with a hard monthly limit switched on by default. A limit that stops the service when it is reached, not one that sends a warning.
His reasoning is about timing. “Nobody wants to wake up to an email sent at midnight warning about a budget limit,” he wrote, and find that while they slept, the service had kept running and run up “several hundred (or several thousand) more dollars”.
The warning email is what most services offer today. It assumes someone reads it straight away and acts. That assumption made sense when people were the ones spending. It makes less sense when an AI tool can keep spending while the team is asleep, or on leave, or in a meeting that runs long.
Willison’s point is that the safe setting should be the default one. Anyone who wants to run without a limit should have to choose that deliberately.
The providers are starting to move
Willison noted that two of the largest cloud providers have recently added caps of this kind. By his account, Amazon Web Services now lets a project pause for the rest of the month once it hits its limit, although that is still reaching only some customers. Google Cloud added monthly spending caps on specific services in July.
You do not need to know what either of those products does to take the point. The companies selling usage-priced computing have started to accept that a warning is not enough. If they think so, it is reasonable for a buyer to ask the same of the AI services their team uses.
Why this is a management question
Nothing here requires technical knowledge. It is the same discipline a manager already applies to a company credit card, a phone contract or a taxi account.
Picture a team that starts paying for an AI service by usage to process incoming documents. In the pilot, the monthly cost is modest and predictable. Then someone connects it to a shared inbox, and a newsletter subscription starts feeding it hundreds of extra messages a day. Nothing breaks. The documents get processed. The bill simply grows, and nobody sees it until the invoice arrives at the end of the month.
No one in that example did anything wrong. The tool did what it was set up to do. What was missing was a decision made in advance about the most the team was willing to spend, and what should happen when it got there.
That decision belongs to whoever owns the budget, not to whoever set up the tool. In most teams those are different people, which is exactly why it gets missed.
The person who sets it up is thinking about whether it works. The person who owns the budget is thinking about next quarter. Neither is thinking about what the tool will do on a Saturday night when a connected inbox fills up.
Two questions before anyone switches it on
What is the most this can cost us in a month? If the answer is “it depends”, ask what it depends on and whether there is a setting that puts a ceiling on it. A good supplier can answer this in one email. An evasive answer is information too.
What happens when it gets there? There are three common answers: the service stops, the service slows down, or the service carries on and you get an email. Only the first one protects you overnight. Decide which you want, and write it down.
It is also worth asking whether anything in a licence you already pay for is billed separately by usage. Some seat-based plans include extra features or higher limits charged on top, and those are easy to switch on without noticing.
Then two smaller habits. Put a named person against each usage-priced service, someone who sees the running cost each week rather than at month end. And when a usage-priced tool is connected to something new, such as another inbox, another team or another system, treat that as a reason to look at the limit again.
None of this is a reason to avoid usage pricing. For many jobs it is cheaper and fairer than paying for seats nobody uses. It is simply a different kind of cost, and it needs a different question asked at the start.
A licence tells you what you will pay. Usage pricing tells you afterwards. The hard cap is how you get the first kind of certainty back.