How to build an AI business case your CFO will actually approve

Build it around four numbers a finance team can audit: the measured baseline cost of the process today, the incremental benefit after the do-nothing case is removed, the 24-month whole-life cost, and the payback month. IBM’s 2025 survey of 2,000 CEOs found only 25% of AI initiatives delivered the ROI expected.

• Number 1 — Baseline. What the process costs today, from actuals over a stated window.
• Number 2 — Incremental benefit. Gross uplift minus what would have happened anyway, as gross margin, not revenue.
• Number 3 — Whole-life cost. Build, run, internal staff time and exit over 24 months — not the pilot invoice.
• Number 4 — Payback month and break-even input. When cumulative net cash turns positive, and the assumption value at which the case fails.

Everything else — strategic rationale, vendor shortlist, risk register — exists to make those four numbers believable.

What your CFO is actually testing when they read an AI business case

Not whether AI works. Whether the benefit lands on a budget line somebody owns, whether the cost shown is the whole cost, and whether the money returns inside the horizon they are accountable for. The technology question was settled before the meeting.

An AI case fails differently from a capex case: the costs are legible — licences, integration, a partner fee — while the benefits arrive as freed-up hours, which are not money. A finance function will approve a smaller cash-releasing benefit over a larger capacity-releasing one, every time. If your document does not distinguish the two, the CFO does it for you in the meeting.

The most fully documented public structure is the HM Treasury Five Case Model — strategic, economic, commercial, financial and management cases. It is the standard framework for UK and Welsh Government spending proposals — many local authorities in England and Wales require appraisal in line with it too — rather than a private-sector standard, but the five headings travel and the guidance is free. The four-number test is not a replacement for it; it is the part of the financial case that gets argued about.

The four-number test: an AI business case on one page

Fill these four rows first. An empty cell is the question you will be asked.

The four-number test is a completeness check, not a scoring model: a case missing any one of the four is not a weak case, it is an unfinished one.

Number 1: what the process costs you today, measured not estimated

Pick one process, not a department, and pull at least six months of actuals with the window stated on the page. The baseline that survives scrutiny is a unit cost with a denominator finance can re-derive: cost per qualified appointment, cost per opportunity, cost per enrolment. Allocate payroll by a time study — a finding of the form “46% of inside-sales hours go to first contact and follow-up” is auditable; “most of their time” is not.

The baseline also prices the do-nothing option. If the process is at capacity and demand is growing, the honest counterfactual is not “stay the same” — it is hiring, and that has a number. Our measurement methodology defines the denominators we report against, including the disclosure that our 7× average sales lift has a median closer to 4×.

Number 2: the incremental benefit, and which kind of benefit it is

Two adjustments turn a gross uplift into a number finance will book. Strip out what would have happened anyway — which needs a holdout or spend-down design, not an attribution report; our guide to running an incrementality test on lead gen spend covers both. Then classify the benefit, because finance credits the four types at very different rates.

Freed hours are not savings until someone’s budget falls or someone’s output rises. Say which you are claiming, on the page, before you are asked.

Number 3: the 24-month whole-life cost, not the pilot invoice

The most common cost error is presenting the pilot price as the cost of the programme. UK Treasury guidance names the concept you need — whole life cost — and states it “is not simply the upfront capital costs of construction or development”. The AI lines people leave out: internal FTE time to own the agent after go-live, data and telephony pass-through, change and training in the receiving team, re-work when the first qualification ruleset is wrong, privacy and security review, and exit — notice, data portability, transcript ownership.

Budget 24 months, not 12. A 12-month view flatters any deployment with a build phase: RAND’s interview study of AI project failure advises leaders to be prepared to commit a team to a specific problem for at least a year, and concludes that a project not worth that commitment is probably not worth starting.

Number 4: the payback month and the input that breaks it

Present payback as a month on a cumulative net cash line, not a ratio. “Payback in month 5” can be checked in month 5; “3.2× ROI” cannot be checked at all, which is why experienced CFOs discount it on sight.

Then run the model backwards. Take the weakest assumption — usually incrementality or close rate — and solve for the value at which the case stops paying for itself, stating whether you have solved the steady-state month or the whole 24 months, because the two thresholds differ. That number converts “trust me” into a testable threshold and hands the pilot a kill criterion. Set it before launch, alongside the measurement design in how to run an AI outbound pilot that can actually fail.

A worked example: the four numbers for an AI appointment-setting case

Every figure below is an illustrative placeholder, shown so the arithmetic is visible end to end — replace them with your own. None are LeadsNow prices or a forecast of any vendor’s results. Scenario: an inside-sales team of six takes 1,200 enquiries a month, contacts 62% within 24 hours, and sets 190 appointments a month at a 70% show rate.

The cumulative net cash line behind that month-5 payback, so you can re-derive it: −$42,300 at month 1, −$46,600 at month 2, −$32,805 at month 3, −$915 at month 4, +$30,975 at month 5. Note what the break-even row is and is not: 32.3% is the threshold at which a steady-state month stops covering its own run cost. The full 24-month case also has to repay the $38,000 one-off and the two months of build, so the incrementality threshold over 24 months is higher than the monthly one. Quote whichever you mean, and say which. One more reconciliation, because the two cost lines are built differently: the whole-life row prices all 24 months at steady state, which is deliberately conservative. Costed on the ramp above — two build months at $4,300 and a half-volume month 3 — the same 24 months come to $637,850, or $57,750 less.

Now the uncomfortable part, still on the placeholder figures. Blended cost per held appointment after deployment is ($57,000 + $27,400) ÷ 196 = $431 — two dollars worse than the $429 baseline. Pitched as an efficiency case, this is rejected on the first slide. Pitched correctly it is a capacity case: 63 extra held appointments a month the team structurally could not reach, at the unit economics they already run, no headcount added. Most AI sales cases are volume cases wearing a cost-saving costume, and mislabelling them is why they die.

Two structural notes a CFO will raise. The fee is charged on attributed revenue while only the incremental share is bookable as benefit — that gap belongs in the model, not a footnote. And outcome pricing changes the risk shape, not the price: paying on outcomes — booked qualified appointments, or a share of the sales generated — rather than retainers or seats means a month with no appointments produces almost no fee, which is a cash-flow property, not a discount. Delivery detail sits on our AI appointment setting and enterprise lead generation pages. Show rates vary by offer and reminder cadence and reach up to 93% on our best-performing accounts, so model your own measured rate.

“My CFO rejected it” — where AI business cases actually fail

Rarely on the technology. In order of frequency:

1. The benefit has no owner. A saving nobody has agreed to give up is a forecast. Get the budget holder’s name on it before submission.
2. The same saving is booked twice. The hours this frees were already claimed by an automation programme approved last quarter. Finance has that ledger; you do not.
3. Run cost is missing. No internal FTE, no re-work, no exit. Number 3 exists for this.
4. No kill criterion. An unbounded commitment is priced as a higher risk class; a break-even threshold and a stop date lower it.
5. Benefits denominated in hours. Convert them to dollars off a named line, or move them to the strategic case and stop calling them savings.

The modelling errors specific to outbound, plus a seven-section document template and procurement-facing risk register, are in our companion page on building the internal business case for AI outbound.

Frequently asked questions

What is a business case in project management, and how is an AI business case different?

A business case justifies spending against the alternatives, including doing nothing. The most detailed published structure is the Five Case Model in the UK Treasury’s Guidance on Developing Business Cases (updated 30 June 2026): strategic, economic, commercial, financial and management cases. That is the standard framework for UK and Welsh Government spending proposals rather than a private-sector standard, but the five headings travel. An AI case differs in one way that matters — its benefits are far more often capacity-releasing than cash-releasing, so the classification must be explicit.

What should an AI business case framework contain as a minimum?

A one-paragraph problem statement with a number in it, the four numbers above, a classification per benefit, a risk register, decision criteria with a named metric and a date, and an exit plan covering data ownership and notice. Without the four numbers, no supporting evidence rescues it.

How long should the business case cover — 12 months or three years?

Twenty-four months: long enough for a build phase and a full run year, short enough that the assumptions hold. RAND’s report The Root Causes of Failure for Artificial Intelligence Projects, based on interviews with 65 experienced AI practitioners, recommends leaders be prepared to commit a team to a specific problem for at least a year, and notes a project not worth that commitment is probably not worth starting.

Do I need an ROI number before the pilot or after it?

Before — but as a break-even threshold, not a forecast. State what the pilot must achieve for the case to hold (“at least 32.3% incrementality in a steady-state month”), then measure against it. A pre-committed threshold is the difference between a pilot that produces a decision and one that produces a debate.

Why do AI business cases get rejected even when the technology works?

Because the benefit never reaches a budget line. The IBM Institute for Business Value’s 2025 CEO study — 2,000 CEOs across 33 countries, surveyed with Oxford Economics between February and April 2025 — reports only 25% of AI initiatives have delivered the ROI expected and only 16% have scaled enterprise-wide. That is CEO self-report, not audited financials, but it is the gap your document has to close.

Can I count staff time saved as a benefit?

Only if you say what happens to the time. Put it in the capacity-releasing row, name the redeployment, and attach the output it produces. Finance treats unredeployed hours as zero.