How to Build an AI Business Case Your CFO Won't Reject

The fastest way to lose a CFO is to open with hours saved. They have seen that number before, from the workflow tool, the RPA project and the last portal, and they have watched headcount stay flat every time.

This isn't cynicism on their part. It's pattern recognition. An AI business case that counts forty hours a week of saved effort as $200,000 of annual value is making a claim about a payroll line that nobody intends to change. The CFO knows that. So the case gets marked as soft and the project competes badly against something with a hard number.

There's a better way to write these, and it fits on one page.

The one-page template

Six sections. If a candidate can't fill all six, it isn't ready to be funded, and that's a useful signal in itself.

• The baseline — three counted numbers about the process as it runs today
• The change — what specifically will be different, stated so it can be measured
• Build cost — engineering weeks, integration, and the data work hiding underneath
• Run cost — inference, monitoring, evaluation and continued human supervision, per month
• The result — first-year net and payback period
• The sensitivity — the two assumptions that move the answer most, and what happens if they're wrong

A page, and no more. Anything longer is usually compensating for a weak baseline.

The baseline is the part nobody does

Three numbers per process: volume, handling time, and error rate. Counted, not estimated in a room.

Volume is the easiest and people still get it wrong, because they count the happy path. If four thousand orders a month arrive but nine hundred of them need a phone call, those are two different processes with two different economics, and lumping them together makes the case unfalsifiable.

Handling time means sitting with the person and timing it. Not asking them. When I do this, the measured time is routinely double what the manager estimated and half what the person doing it believed. Both are wrong in predictable directions.

Error rate is where the strong business cases come from. What fraction of these go wrong, and what does a wrong one cost? A process with a two percent error rate and a $3,000 cost per error is worth far more to fix than the same process with a two percent rate and a $30 cost, and the hours are identical in both.

Counting value without lying

There are four honest ways to state value, and they carry very different weight with a finance team.

Cost avoided, with a headcount decision attached

The strongest version. Volume is growing thirty percent, the plan was to add three coordinators, and now it isn't. That's a real number because it changes a real budget line. Note the shape: growth absorbed rather than headcount removed. Most AI value in mid-market companies is this, and it's both more accurate and easier to defend than a reduction story.

Error cost reduced

Errors have invoices attached: the reship, the credit, the detention charge, the write-off. Take your current rate, apply the expected improvement, be conservative about the improvement, and you have a number your CFO can audit against last year's ledger.

Cycle time converted to revenue

Only claim this where you can trace it. "Quotes go out in four hours instead of three days" becomes value when you can show your win rate on quotes returned within a day, which many companies can from their own CRM. Where you can't show it, say so rather than assuming a percentage.

Hours released, stated as hours

Sometimes the honest answer is that people get eight hours a week back and will spend them on other work. That's genuinely valuable and it isn't cash. Say it plainly and let leadership weigh it. A case that labels this correctly earns credibility for the three numbers above it.

Run cost: the section that gets skipped

Most AI business cases model the build and stop, which is roughly like buying a truck and budgeting only the purchase price. Four components belong in the monthly number.

• Inference — tokens or API calls at your real volume, including retries and the longer prompts that always arrive later
• Monitoring and evaluation — running the eval set on a cadence, plus the tooling and the person's time to look at it
• Human supervision — the review that remains, which for most workflows never reaches zero and shouldn't
• Maintenance — model and vendor changes you don't control, integration drift, and the fixes those force

For a mid-market document or routing workflow, this typically lands between $2,000 and $20,000 a month depending on volume. Over three years that regularly exceeds the build cost. A case that omits it isn't conservative, it's wrong, and a CFO who has been through one of these before will find the omission.

Calculating ROI so it survives scrutiny

The AI ROI calculation is net annual value divided by total annual cost, with run cost in the denominator. Then two disciplines that separate a real case from a hopeful one.

First, be conservative on the improvement rate and say so explicitly. If you think the system handles eighty percent of cases autonomously, model sixty. When it lands at seventy-five, you've built trust for the next case. When you model ninety and it lands at seventy-five, you've spent it.

Second, publish the sensitivity. Two assumptions usually move the answer: the automation rate and the volume. Show the result at both ends. A case that holds up at the pessimistic end of both is fundable without argument, and a case that only works at the optimistic end of both is a bet dressed as an analysis.

What makes a case fundable

A CFO rarely rejects a business case because the number is small. They reject it because the number cannot be attacked. A fundable case is one where every line survives being challenged on its own: the baseline came from counting rather than asking, the value is stated in the currency the improvement actually arrives in, the run cost is present, and the improvement rate you modeled is lower than the one you privately expect.

That last discipline is the one people skip, and it is the cheapest credibility you will ever buy. You will be asked to defend these numbers again in a year against what actually happened. How that conversation goes decides whether your second business case gets read at all.

What to do next

Pick your top candidate and try to fill the six sections. Most people discover they can't complete the baseline, which is the actual finding: you're not ready to fund it, you're ready to spend a week measuring it. That week is cheap and it changes which project you pick about a third of the time.

If you have five candidates and no way to rank them, that's the opportunity assessment: every candidate priced and scored, with a business case for the top few and an explicit list of what not to build. And before any of it, the readiness checklist will tell you whether the data supports the case at all.

Send me a business case you're unsure about and I'll tell you which number a CFO will go after first.