Every AI budget request that dies in a finance review dies the same way. Someone stands up and says the tool "saves time," "boosts productivity," or "drives efficiency." The CFO nods politely, asks how that shows up in the numbers, gets a vague answer, and the request goes to the bottom of the pile.
The problem is not that AI lacks ROI. The problem is that most teams pitch it in a currency finance does not accept. "Saves time" is not a number. It is a feeling. And no CFO funds feelings twice.
Aggregate ROI is a story. Unit economics is a case.
The instinct is to build a big top-down number: "AI will save the company 3 million a year." Finance has seen a hundred slides like that, and they discount it to zero on sight, because it hides all the assumptions and none of them are testable.
Flip the direction. Do not model the company. Model one run of one workflow. What does a single quote, a single collections email, a single monthly report cost to produce today โ and what does it cost with an AI teammate doing the drafting and a human approving? A number built from the bottom up, per unit, is one the CFO can interrogate, verify, and โ crucially โ extrapolate. That is the difference between a story and a case.
The four terms of an honest AI unit-economics model
A credible per-workflow model has exactly four inputs, and most pitches conveniently omit the last two.
- Current cost per unit. Fully loaded labor time to produce the output today, times the loaded hourly cost. Three hours of a specialist's time to build a quote is a real, defensible number.
- AI cost per unit. The actual run cost โ model tokens, plus a fair share of platform cost. This is small but not zero, and pretending it is zero is how you lose credibility in the first meeting.
- Human-in-the-loop cost per unit. The minutes a person still spends reviewing and approving. If your model assumes zero human time, the CFO knows you are lying, because someone has to own the output that goes to a customer.
- Error/rework cost avoided. The mispriced quotes, the missed follow-ups, the compliance slips that used to happen and now do not. This is often the biggest term and the one everyone forgets.
Why the research points at augmentation, not replacement
There is a temptation to model AI ROI as headcount removed. Resist it โ it is both politically radioactive and empirically shaky. A field study of over five thousand customer-support agents found generative AI raised productivity by about 14 percent on average, with the largest gains going to less-experienced workers, and it worked by making people better and faster, not by deleting them.
That shapes the honest model. The value shows up as more output per person, faster turnaround, and fewer errors โ not as a smaller payroll. A CFO trusts an augmentation case precisely because it does not require them to bet on layoffs to hit the number.
Instrument the workflow so ROI is measured, not estimated
Here is the move that changes the conversation permanently: stop estimating ROI and start recording it. If every run logs its cost, its turnaround time, and whether the output was approved as-is or corrected, then after two weeks you are no longer arguing about a projection. You are showing an actual ledger.
This is why per-run cost tracking is not a nicety โ it is what converts the second budget request from a debate into a formality. The first ask is a hypothesis. The renewal is a spreadsheet. Give your CFO the spreadsheet, denominated per workflow, and AI stops being the thing finance questions and becomes the line item finance defends.
References
- Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper No. 31161, 2023.
Want to build a per-workflow ROI model on one of your real processes? Book a 30-minute call and we'll draft the numbers together.