The CFO-Ready Business Case for AI

The business case for AI fails in the same place every time: it's a story about potential, and the CFO deals in numbers.

I've sat through the AI pitch more times than I can count. The slides are beautiful. The use cases are compelling. The vision is real. And then someone asks the question that ends the meeting: what does this actually cost, and what does it actually return? The answer is usually a story about transformation, and the story doesn't survive contact with a finance review.

The fix isn't a better story. It's a better unit. The business case for AI that survives finance is built on unit economics — the cost per resolved task, the payback period, and three numbers that hold up under questioning.

The unit that changes everything

The most important number in any AI business case is the cost per resolved task. Not the cost of the model call. The cost of the task being done — correctly, end to end, with the human review and the exception handling included.

Here's how you get it. Take the task the AI is doing — resolving a support ticket, processing an order, generating a report. Count everything the AI version costs: the model calls, the infrastructure, the review labor, the exception handling, the maintenance, divided by the number of tasks completed. That's the AI unit cost.

Then get the same number for the way it's done today — the labor, the rework, the error rate, the time. Divide by the same number of tasks. That's the baseline unit cost.

The difference between the two numbers is the unit economics of your AI investment. It's not a story. It's arithmetic. And it's the number that survives the CFO's review, because it's the number the CFO can check.

The business case for AI is one number: the cost per resolved task, AI version versus current version. Everything else is context.

The three numbers that survive finance

Beyond the unit cost, three numbers will get you through any finance review.

Payback period. How long until the savings from the unit-cost difference pay back the investment — the build cost, the integration, the training, the change management? The honest answer for most AI projects is six to eighteen months, and the honest answer is the one that gets the approval. A payback story that's too fast gets challenged. A payback story that's real gets funded.

The headcount reality. What does the labor actually do? Not the marketing version — "this will let us do more with less" — but the real version: the hours the AI removes from the task, the hours that return to the team, and the hours the team needs to spend reviewing AI output. Most business cases only count the hours removed. The CFO's finance team will find the hours added. Count both.

The failure cost. What does it cost when the AI is wrong? Not the average case — the cases that matter. The wrong answer that reaches a customer. The error that breaks a compliance process. The silent degradation nobody notices for a month. The business case needs a number for the risk it's taking on, and a plan for keeping that number small. The review layer is a cost. It's also the insurance premium that makes the business case honest.

Why the story-based case dies

The story case dies because it makes claims the numbers can't back. "Transform the customer experience." "Unlock new insights." "Give us a competitive edge." None of these are false. None of them are numbers. And in a finance review, a claim that isn't a number is a claim that doesn't exist.

The finance team isn't being hostile. They're being rigorous, and rigor is exactly what the AI investment needs. The story case treats the CFO's questions as obstacles. The unit case treats them as the spec. The CFO who asks "what's the cost per resolved task" is asking the right question — the question that will make you measure the thing, which is the thing that makes the project succeed.

How to build the case in a week

You don't need a consultant for this. You need a task, a week, and a spreadsheet.

Pick the highest-volume, most measurable task. Time it for a week — the current cost, the error rate, the review hours. Run the AI version on a sample, ideally in parallel, and measure the same things. Compute the two unit costs. Compute the payback. Write down the failure costs and the review plan. That's the business case. It's a week of honest measurement instead of a quarter of slides, and it's the version that gets approved.

The winning AI business cases all have the same shape. A specific task. A measured baseline. A unit cost that's lower. A payback that's real. And a risk number that's managed. The vision slides are fine — put them at the back, after the numbers have done the work.

The CFO doesn't need to believe in AI. They need to believe in a number. Give them the cost per resolved task, the payback period, and the honest risk, and the number will do what no transformation story ever could. It'll get approved.

Think this argument fits your event? Tell me about the room — the calendar is selective.