When an AI bill runs over, most large companies pause the usage that overran, and most need several business days to several weeks to do it. Nearly three in four CFOs at North American companies with $1B or more in revenue told Deloitte exactly that. A shutdown that slow is what a company falls back on when it never built a way to see which work caused the bill.
If you run a mid-market company, this is coming to your desk. The way AI is priced is changing, and the bill now moves with how your people work.
Usage pricing means the bill moves every day
More AI providers now charge per token instead of per seat or at a flat rate, Deloitte's third-quarter CFO Signals release notes, and a bill can spike quickly when a tool reads long documents or runs long back-and-forth loops. A token is a small piece of text the model reads or writes, and you pay for each one.
A seat license costs the same whether people use it or not. A usage bill rewards a good workflow and punishes a sloppy one, often within the same week. That makes it an operating cost to manage the way you manage freight or overtime, line by line and month by month.
Finance has a say in AI and still cannot see it run
Nearly 80% of the CFOs say finance has a formal role in AI decisions, yet only 20% use advanced controls, such as tying cost to a specific workload or routing requests through an AI gateway. Most of the rest lean on usage guardrails, team budgets or the IT budgeting they already had.
So finance weighs in up front, and the visibility does not follow the spend. When the overrun comes, just 19% deal with it the same business day. When nobody can say which workflow drove the bill, the only lever left is to pause the usage, and finding which usage to pause takes days.
Shutting it off teaches you nothing
All-in and all-out on any tool are the same mistake, because neither one is learning anything. A company that switches a tool on wide open and then pauses it when the invoice surprises it is back where it started, minus a few weeks of work, and it still does not know what went wrong.
My rule for AI output in finance applies here too. Trust it as directional, and have finance verify the specific numbers before anything goes out. If that full check has to happen every time a report runs, you have a problem. Cost works the same way. If understanding an overrun takes a shutdown, the problem is the missing meter, not the tool.
Give every AI workflow its own meter and its own owner
Before the next rollout, decide how the bill gets read. Four decisions cover most of it.
- Tag AI spend to the workflow that used it: the monthly close, the customer-service queue, the proposal drafts. A vendor invoice by itself tells you what you spent, not where.
- Give each workflow one owner who sees its cost every week, along with the result it is supposed to produce.
- Agree in advance what happens on an overrun: cap the usage, move the work to a cheaper model, or fix the loop that ran long. Decide it the same week, for that one workflow.
- Read your contracts. Among the CFOs surveyed, 27% say contract terms such as pricing, commitments or data egress limit their flexibility, and that is easier to fix before you sign than after.
Build this while the bill is still small enough for one person to read line by line.
If your AI is a handful of seat licenses, keep it simple
Some companies pay a flat monthly price per user for a few tools and run nothing usage-priced. Their bill does not spike, and a budget line and a quarterly review of who uses what are enough. The meter matters once a workflow is priced by use, and especially once AI runs inside a process without a person starting each request.
Start with last quarter's invoices
Pull last quarter's AI invoices and try to say which workflow each dollar went to and what it produced. If you cannot, that is the first thing to fix, before the next tool goes live.
I have written about finance teams that spend more time checking AI than using it and about fixing a finance process before you hand it to AI. Building cost and accountability into how a company runs AI is part of how I work with mid-market leaders. If your AI bill has surprised you, I am glad to talk it through.
Source: Deloitte CFO Signals Survey, third quarter 2026, 200 CFOs at North American organizations with at least $1B in annual revenue, released Oct 21, 2026.
John Wyatt · John Wyatt's page · All posts · RSS