AI Economics Playbook

The Sentence That Costs the Most

A prudent-sounding AI request can authorize spending that nobody has priced. Give the yes a cost ceiling and a review point.

Kiran SusarlaOriginally published November 17, 2026 · 6 min read
A person lifts the edge of a wall panel, revealing a cobalt river that continues beyond the wall.
Conceptual illustration: a small opening reveals a continuing flow beyond the visible boundary. Illustrative, not measured financial data.

Executive summary

AI spending can be approved through an apparently reasonable conversation without anyone agreeing to its full cost. The article asks finance leaders to examine how costs change with usage, set a spending ceiling and agree when the decision must return for review.

The unpriced yes

A few months ago I had someone over to look at a patch of damp spreading across a bedroom wall. He pressed it, tapped along the skirting, frowned a little, and said exactly the sort of thing you want to hear. We would not really know how far the water had travelled until we opened the wall up, but he would only do what was needed, nothing more. I said yes. Of course I said yes. It was the responsible answer, the one that refused to overcommit, and it came from the only person in the room who actually understood damp.

What I did not notice, standing there nodding, was that I had just agreed to a cost I had no way of seeing. The number depended entirely on something neither of us had looked at yet, which was how far the water had already gone. (I would love to tell you I asked for a ceiling before he picked up the chisel. I did not.) The sentence that got my yes was not reckless. It was careful, and that is precisely why I agreed to it without a figure.

Somewhere in a conference room this quarter, an AI lead is going to say the budget-table version of that sentence, and a finance leader is going to nod for the very same reason. In the book I wrote on AI economics, I put the sentence as plainly as I could: "We need to scale the model. Demand is strong, but inference costs are climbing and we will need more compute to hold latency down." Read it slowly. Every word is familiar and the whole thing sounds prudent. They are managing demand, watching costs, protecting the experience. The meeting moves on, and a commitment has been made to something nobody can price, on a cost curve nobody in the room has been shown.

The approval happened in the nod. There was no invoice to argue with, no line item to send back, no moment that felt like signing a cheque. It felt like agreeing with a sensible person who had clearly thought about the problem, which is a very different sensation from making a financial decision, even though a financial decision is exactly what it was. This is why the most expensive sentence in an AI budget is rarely the one with a frightening number in it. The frightening number gets scrutiny. The dangerous sentence is the one with no number in it at all, delivered in the register of restraint.

When the bill arrives

If that sounds like an edge case, the most recent evidence says it is closer to the house style. A Capgemini Research Institute snapshot published on the twelfth of November, drawn from five hundred senior executives at large organisations across ten countries, found that half of them say AI spending gets approved without adequate visibility into its full cost. Not half of the careless ones. Half. Fifty-nine percent said their AI spending over the past year came in significantly above what they had expected or budgeted. A Menlo Ventures report a week earlier found that, on average, enterprises planned to grow their AI spend by about 1.9 times this year and actually grew it by 4.7 times, close to two and a half times their own plan.

None of that describes reckless companies. It describes careful people saying yes to sentences that sounded careful, and meeting the bill later. And the room for that kind of surprise is shrinking. In a Fortune report the same week, Mark Williams, a Boston University lecturer and former Federal Reserve bank examiner, speaking at a dinner for finance chiefs in New York, put it plainly: “I think you should assume the cost of capital is going up.” When money gets dearer, every unpriced yes has to compete harder for it, and the scrutiny you skipped at the nod tends to show up later, at a far less convenient moment.

A run rate, not a repair

What makes the AI version worse than my damp wall is the shape of what you agreed to. My wall, at least, had an end. Once it was fixed, the spending stopped. A yes to "scaling the model" is not a yes to a repair. It is a yes to a run rate, a cost that recurs every hour the system stays live and moves with usage you do not control. Double the users and the cost might double, or it might triple when shared infrastructure runs out and you cross into dedicated, costlier capacity, or it might fall if volume works in your favour. Intuition cannot tell you which. Only the curve can, and the curve is the one thing a prudent-sounding sentence leaves out.

So the useful move is not to say no, and it is certainly not to go away and learn how inference works. You no more need to understand the model than you need to understand a jet engine to run an airline without losing money. The move is to stop the nod from being the whole transaction, and to put an edge on the yes before you give it.

Give the yes an edge

Start by asking for the shape of the cost rather than a single figure. What does this run to at today's volume, at ten times that, at a hundred, and where does it stop being worth it? Nobody is being asked to predict the future perfectly, only to draw the curve and mark the cliff. A team that understands its own economics can produce that in an afternoon. A team that cannot has just told you something more useful than any number would have: it does not yet understand what it is asking you to fund, and that, far more than the size of the request, is what should give you pause.

Then, while you still have the leverage, attach a boundary. Approve a ceiling rather than an open tab, and name the point at which the request comes back to the table. Yes, up to here, and we look again when the monthly run rate reaches this. That one sentence turns an open-ended commitment into a bounded one with the decision already built in. It costs nothing to say in the calm of the approval meeting, and it is almost impossible to say later, because later is exactly when the leverage has gone and the spend has become something everyone quietly depends on.

There is nothing hostile in any of this. You are not doubting that the person in front of you is good at their job. You are giving your own yes an edge, so that if the curve turns out steeper than anyone hoped, stopping is a plan you wrote together in a quiet room rather than an argument you have to start in a loud one. Hmm... maybe that is the whole discipline, really. Not suspicion. Just refusing to let a reasonable tone stand in for a reasonable boundary.

I have started keeping a small test for myself, and it works on walls and budgets alike. If I cannot say out loud the most this could cost me, I have not actually decided anything. I have only been agreeable. Agreeable is a lovely thing to be at dinner! It is an expensive thing to be across a budget table, because the person asking is usually not trying to trick anyone. They are speaking their own language fluently, and a fluent, prudent sentence with no number in it will get a yes almost every time, unless someone in the room insists on putting the number back in before the nod.

The sentence that costs the most is the one you approve by nodding. So before you nod, make it priceable, and give the yes an edge.

I would really like to hear how others handle this. What is your cleanest way to bound an open-ended 'yes' before it turns into an open-ended bill? Always ready for a learning or two.

Original sources

Sources: Capgemini Research Institute, AI economics snapshot (November 12, 2026): https://www.capgemini.com/news/press-releases/most-organizations-plan-to-adopt-a-multi-model-ai-strategy-as-nearly-six-in-ten-faced-significantly-higher-ai-costs-than-expected/

Menlo Ventures, 2026: The State of Enterprise AI (November 3, 2026), figures as reported by SaaStr: https://www.saastr.com/ai-spend-hit-177-billion-this-year-only-22-of-it-came-out-of-software-budgets/

Fortune, "CFOs face a tough choice on AI spending as borrowing costs rise, expert warns" (November 12, 2026): https://fortune.com/2026/11/12/cfo-face-a-tough-choice-on-ai-spending-as-borrowing-costs-rise-expert-warns/

#CFO #FinOps #EnterpriseAI #AICostManagement #AIGovernance

Editorial note: The original article wording and sequence are preserved. The executive summary, section headings, takeaways and cover illustration are editorial additions.

Continue reading

Browse Insights