The AI bill is running five times plan. Now what?

Here is the scenario I keep hearing from leadership teams, and living through myself. You implemented a platform last year. The AI features were included, or priced per seat at a number finance could model. Then the vendor moved toward usage-based pricing, or your people actually started using the tools, and utilization went through the roof. The bill is now running at a multiple of plan. Some operators I talk to are seeing model spend up four to five times year over year. And the question the executive team is quietly asking is the bluntest one there is. Do we stop?

If you have asked that question and wondered whether everyone else already figured this out, they have not. The data says almost nobody has.

You are the norm, not the outlier.

The FinOps Foundation surveyed 1,192 practitioners managing 83 billion dollars in spend for its State of FinOps 2026 report. 73 percent of organizations exceeded their AI budget. Only 20 percent could predict AI spend within ten percent. And 80 to 90 percent of the money goes to inference, the running of AI, not the building of it.

Gartner told CIOs that organizations can miscalculate generative AI costs by 500 to 1,000 percent when they do not understand how those costs scale. So if your model was off by five times, you did not fail at math. You used the same math everyone else did.

Now look at what companies are actually doing about it. Mavvrik's 2026 cost governance research found unexpected AI costs changed business decisions at 62 percent of enterprises, forced board-level escalations at 40 percent, and pushed a quarter of them to delay or cancel AI initiatives. Forrester predicts enterprises will defer 25 percent of planned AI spend into 2027. And the sharpest example: Uber blew through its entire annual AI budget in four months, then capped every employee at 1,500 dollars per month per AI tool. A company with world-class engineering got surprised by its own adoption. That is the state of the art.

The three ways your stack bills you for AI.

Part of why the number is so hard to model is that your vendors do not charge for AI one way. They charge three ways, and each one fails differently.

• Consumption. Salesforce prices Agentforce at two dollars per conversation, or ten cents per action on credits. Metered like a utility. This is the model that blows up mid-year, because agentic workflows multiply the actions behind every business outcome.
• The seat tax. Microsoft 365 Copilot at 30 dollars per user per month, ServiceNow Now Assist at 15 to 25. Predictable, and completely decoupled from usage. In July 2026 Microsoft raised Microsoft 365 prices by up to 43 percent while bundling Copilot into paid tiers. The tax became unavoidable, even for the people who never use it.
• "Included, for now." Oracle announced NetSuite Next with AI included in existing licenses at no extra cost. Included does not mean free. It means the inference cost is invisible, unattributable, and waiting for you at renewal, on a vendor famous for renewal math.

Three models, three failure modes. The consumption model surprises you in month seven. The seat tax pays for shelfware. The included model cannot be assigned to a client, a project, or a P&L line at all.

"Do we stop the AI?" Wrong question.

Here is the part that matters. Utilization went through the roof because your people find the tools useful. That is not a cost problem. That is the adoption you were praying for a year ago, arriving with a bill attached. Switching it off to save the budget concedes the capability while your competitors keep compounding theirs. Uber did not stop. Uber governed.

The real decision is a choice between three moves, and most companies need some of each.

• Throttle. Caps per person or per tool, cheaper models routed to cheaper work, guardrails that stop wasteful calls before they burn money. This is what Uber's 1,500 dollar cap is. Not a retreat. A governor on an engine that is finally running.
• Re-cost. Attribute every AI dollar to the client, product, or workflow that consumed it. Until you can see cost per client, you do not know which accounts are quietly unprofitable, and you cannot make a single defensible pricing decision.
• Reprice. Once you can see the cost, your contracts have to carry it. Escalators, usage tiers, or pricing the expected drift into next year's agreements instead of eating it in this year's margin.

The order matters, and it starts with separating what you know from what you feel. The invoice is fact. The story that "AI is making us more productive so it is fine" is interpretation. So is the panic that says shut it down. This is the Evidence discipline: get the facts on one side of the table, the interpretations on the other, and decide from the facts.

The invoice is fact. The productivity story is interpretation. Govern the fact, test the story, and never let either one make the decision alone. Chris, on every AI budget conversation this year

The three ledgers of AI cost.

To re-cost your client book, you need three ledgers, because that is how the money actually leaves the building.

• Ledger one. Direct model spend. Your own API calls to model providers. Measurable per call, attributable per customer, if you instrument it.
• Ledger two. Consumption-billed vendor AI. Agent conversations, message packs, compute units. The vendor meters it. Mapping it to your clients is your job, and nobody does it for you.
• Ledger three. Seat taxes and included AI. Attributable only by allocation policy, and the included kind arrives silently at renewal.

Then the operator's choice is honest and simple. Re-cost the client book against all three ledgers now, or knowingly price the drift into forward contracts with an escalator. Doing neither is how the margin surprise happens, and the board meeting that follows it.

What I would do this quarter.

Pull the last six months of AI spend across all three ledgers and put it against your top twenty clients. The pattern will not be subtle. A handful of clients or workflows will be consuming most of the money, and at least one will be costing more to serve than it pays you. That one account is worth more than any dashboard, because it forces the pricing conversation your team has been avoiding.

Then set the throttle policy before the next invoice, not after. Caps with an exception process, the way Uber did it, keep the capability and the goodwill. Cuts announced after a blown quarter keep neither.

You are not the only one who did not see this coming. You would be one of the few who did something structured about it. If you want a second set of hands on it, that is the work.