An AI business case is the paper that asks finance and the board to fund an AI initiative. It compares a measured baseline with at least two options, one of them doing nothing, prices one-off and running costs, says how each benefit will be measured, and shows ROI, net present value, payback and the assumptions that would break the case.
Key takeaways
- 01 Keep a do-nothing option in the comparison until the end. The UK Treasury's Green Book makes the business-as-usual option compulsory on every shortlist.
- 02 Hours saved are capacity, not cash. Name the budget line where freed time will show up, or expect finance to discount the benefit.
- 03 Price run costs for every year of the horizon: inference, monitoring, evaluation reruns, human review and support.
- 04 Ramp adoption year by year. Assuming full benefits from the first month is one of the easiest ways to inflate an AI case.
- 05 Publish switching values, the point at which each key assumption turns net present value to zero, so the pilot knows what it has to prove.
What an AI business case is and who owns it
An AI business case, sometimes called an AI investment case, is the decision document that asks for money to build, buy or scale an AI system. It answers one question for whoever holds the budget: is this the best use of the money, compared with the alternatives, including doing nothing?
It is neither a strategy nor a project plan. The one-page AI strategy says where AI should create value. The business case argues for a single investment inside that strategy. The plan schedules the work once the money is approved.
| Role | What they own in the case |
|---|---|
| Business sponsor | The case and its benefit commitments |
| Finance business partner | Discount rate, horizon, baseline cost checks, and which benefits count as cash |
| AI or IT lead | Build and run cost estimates, technical feasibility, platform choice |
| Risk, security and privacy | Risks and controls; confirms the data may be used |
| Benefit owners | One named manager per benefit line, accountable for delivering and reporting it |
Write one case per investment decision. Several use cases can share one platform cost, but each should carry its own benefit lines so a weak one can be cut without sinking the rest. To decide which candidates deserve a full case, score them on value, risk and effort first, and use the phase-by-phase implementation playbook to see which gate the case serves. Narrower angles are covered in the board-level argument for private AI and ROI arithmetic for workflow automation.
What finance and the board expect from an AI business case
McKinsey’s State of AI survey, published on 25 August 2026, found that 80% of respondents say AI has improved their own productivity, yet only 37% attribute any EBIT impact to it. That gap between personal gains and company results is what a finance reviewer probes first.
Gartner told CFOs on 4 November 2026 that hours saved, looked at in isolation, do not measure business value, and that the gain leaks away unless the organisation decides where freed capacity goes. Its survey of 1,303 senior functional leaders, run from January to April 2026, found that 23% reported headcount reductions because of AI while 17% reported increases. A case that turns minutes saved straight into salary savings will not survive that review.
Public-sector appraisal rules are a useful benchmark even for companies. The UK Treasury’s Green Book, in the 2026 edition last updated on 28 October 2026, requires the business-as-usual option to go through to the shortlist, tells appraisers to adjust explicitly for optimism bias, and asks for sensitivity analysis and switching values. Its business case guidance, revised on 30 June 2026, rests on the Five Case Model: strategic, economic, commercial, financial and management.
| What the reviewer asks | Template section that answers it | Five Case Model dimension |
|---|---|---|
| Why act now, and what does the problem cost today? | Problem and baseline | Strategic |
| What if we do nothing, and who would supply the alternatives? | Options | Economic and commercial |
| How will we know the benefit is real? | Benefits and measures | Economic |
| What will it cost to run, not only to build? | Costs | Financial |
| What could go wrong, and who controls it? | Risks and controls | Management |
| How sure are you of the numbers? | Sensitivity and assumptions log | Economic |
The AI business case template
Use the nine sections in this order; the examples come from the insurer case further down. The DOCX in the template pack repeats the structure with blank fields.
1. Problem and baseline
| Field | What to write | Example |
|---|---|---|
| Problem statement | The problem in one or two sentences, without mentioning AI | Intake takes 25 minutes per claim and 9% of claims need rework |
| Process in scope | Start and end points, yearly volumes, teams involved | First notice of loss to coverage check; 120,000 claims a year |
| Baseline measures | Today’s value for every metric the benefits will move | Handling minutes, rework rate, backlog, complaints |
| Measurement method | How the baseline was measured, sample size and dates | Time study of 400 claims across four weeks |
| Cost of the current process | Fully loaded cost per unit and per year | €48 per handler hour including on-costs and overhead |
| Strategic link | The strategy objective served | Absorb volume growth without new headcount |
2. Options, including doing nothing
| Option | What to write | Example |
|---|---|---|
| 0. Do nothing | What happens if current arrangements continue, including costs you will incur anyway | Hire nine handlers next year to absorb 8% volume growth |
| 1. Do minimum | The smallest change that meets the objective | New intake forms and checklists, no AI |
| 2. Preferred AI option | The proposed system, its scope and how it is sourced | An assistant drafts the claim summary and coverage check for the handler to approve |
| 3. Alternative AI option | A different scope, supplier or deployment model | An add-on module from the core claims system vendor |
Say why each rejected option lost. Comparing the AI option only with doing nothing invites the question of what a cheaper fix would achieve.
3. Quantified benefits and how each is measured
Express every benefit in one of four forms, and agree the measurement before you put a number on it.
| Benefit type | Unit | How to value it | How to measure it | Counts as cash when |
|---|---|---|---|---|
| Hours | Hours freed per year | Hours × loaded hourly rate × share finance agrees to bank | Same time-study method before and after, on the same task mix | The hours replace planned hires, overtime, contractors or outsourced work |
| Error rates | Errors or rework per 1,000 items | Errors avoided × cost per error | Quality sampling or system logs, before and after | Rework cost or write-offs fall in a named budget line |
| Revenue | Conversion, throughput or retention | Extra volume × contribution margin, never × list price | Pilot group compared with a control group | Finance agrees the attribution method in advance |
| Risk | Expected loss per year | Probability × impact, before and after controls | Risk register scores, incident data, audit findings | Rarely; show it as a separate line outside the cash NPV |
Give each benefit line an owner, a start date, a ramp and a reporting source. When two lines draw on the same hours, keep only one of them.
4. One-off and recurring costs
| Cost line | One-off or recurring | What to include |
|---|---|---|
| Integration | One-off, plus yearly maintenance | Connectors to core systems, identity, document stores |
| Platform | Recurring | Licences or subscription; check whether price scales with users or usage |
| Inference and compute | Recurring | Tokens or GPU hours per task × volume; hardware depreciation if on-premises |
| Data preparation | One-off, then periodic | Cleaning, access approvals, the evaluation test set |
| Evaluation and monitoring | Recurring | Test reruns after changes, drift checks, quality sampling |
| Human review | Recurring | Review minutes per task × volume × loaded rate |
| Security, privacy and legal | One-off, then yearly | Assessments, data protection impact assessment, contracts |
| Change and training | One-off, then refreshers | Training hours, communications, launch support |
| Programme team | One-off | Project manager, product owner, analysts |
For how the deployment model shifts run costs, see the guide to on-premises AI platform cost and TCO, and check the categories that pilot budgets tend to leave out.
5. Risks and controls
| Risk | Effect on the case | Control | Owner |
|---|---|---|---|
| Output quality below threshold | Benefits fall and review time rises | Evaluation gate before each rollout wave | AI lead |
| Low adoption | Benefits arrive late | Training, team targets, monthly usage reporting | Business sponsor |
| Data cannot be used as planned | Scope shrinks | Privacy review before any build work | Data protection officer |
| Run costs above estimate | NPV falls | Cost per task tracked monthly; cheaper model for simple tasks | Finance partner |
Carry the detail into your AI risk register so the case and the register do not drift apart.
6. ROI, NPV and payback
Pick one horizon, often three to five years for an AI investment, and apply the same adoption ramp to every metric.
- Net cash flow in year t: benefits(t) − costs(t)
- ROI: (total benefits − total costs) ÷ total costs, over the horizon
- NPV: the sum, for t = 0 to T, of net cash flow(t) ÷ (1 + r)^t, where r is the discount rate
- Benefit-cost ratio: present value of benefits ÷ present value of costs
- Payback period: the year in which cumulative net cash flow turns positive, interpolated within that year
Ask finance for r. Companies normally apply their hurdle rate or cost of capital. UK public bodies use the Treasury’s social time preference rate, which the 2026 Green Book sets at 3.00% in real terms for years 1 to 30. The Green Book keeps the economic case in real terms and the financial case in nominal terms; say which convention each of your tables uses.
7. Sensitivity analysis
Change one assumption at a time and report the NPV for each change. Then add a low and a high scenario that move several assumptions together. Finish with switching values. The Green Book defines a switching value as the value a key assumption would have to reach for an option to stop being value for money.
8. Assumptions log
| ID | Assumption | Value | Source | Confidence | Owner | Tested in pilot |
|---|---|---|---|---|---|---|
| A1 | Minutes saved per claim | 10 | Supplier demo on 30 sample claims | Low | Claims operations lead | Yes |
| A2 | Share of freed hours banked | 70% | Hiring plan agreed with finance | Medium | Finance partner | No |
| A3 | Inference cost per claim | €0.50 | Token counts from 30 sample claims | Medium | AI lead | Yes |
Rate each assumption high, medium or low confidence. The low-confidence ones with a narrow switching margin are what the pilot or a proof of concept must test.
9. Decision requested
State the decision in one sentence, then the options the approvers have:
- Approve the full investment
- Approve in stages, releasing each tranche on evidence
- Approve a pilot only, with a return date and exit criteria
- Reject or defer, with the reason recorded
List the conditions attached to approval, the gate dates, who decides at each gate and the metric that triggers a stop.
A one-page version for a single use case
For a small use case on an already funded platform, one page will do: baseline, preferred option against doing nothing, three benefits with measures, costs, NPV and payback, the two riskiest assumptions, and the decision.
What changes in an agentic AI business case
Agents take several steps per task and call tools, so both cost and risk vary per task. Price the case per completed task, including retries, tool calls and human approvals, and count the share of tasks that still end with a person. State what the agent may change in other systems and who approves those changes. In June 2025 Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, so give each of those three its own section.
AI business case example: a fictional insurer
Example Mutual is a fictional home and motor insurer that handles 120,000 claims a year. All figures are illustrative.
Problem and baseline. A time study of 400 claims across four weeks measured 25 minutes of handler time per claim for first-pass intake, excluding rework. Nine per cent of claims needed rework because information was missed at intake. A handler hour costs €48 fully loaded.
Options. Doing nothing means hiring nine handlers next year to absorb 8% volume growth; the do-minimum option of new forms and checklists would save about two minutes per claim.
Benefits at full adoption.
- Hours: 120,000 claims × 10 minutes saved ÷ 60 = 20,000 hours a year. Finance agreed to bank 70% against the planned hires: 14,000 hours × €48 = €672,000. The other 6,000 hours go to callbacks and quality checks and are reported as capacity, not cash.
- Rework: moving from 9% to 6% of claims means 3,600 fewer reworks × €35 each = €126,000. The baseline excluded rework time, so these hours are not counted twice.
- Risk and revenue: an expected-loss reduction from moving staff off public chatbots is reported on a separate line and kept out of the NPV. No revenue benefit is claimed.
Cash benefit at full adoption: €672,000 + €126,000 = €798,000 a year.
Costs. One-off in year 0: integration €180,000, platform setup and security review €60,000, test set and evaluation €50,000, training and change €60,000 and project management €30,000, a total of €380,000. Recurring from year 1: platform €150,000, inference €60,000 (120,000 claims × €0.50), AI operations at half a full-time role €55,000, quality sampling €38,400 (2% of claims × 20 minutes × €48) and evaluation reruns €20,000, a total of €323,400. Full run costs are budgeted from year 1 even though usage ramps up.
Cash flow at a 10% discount rate.
| Year 0 | Year 1 | Year 2 | Year 3 | Total | |
|---|---|---|---|---|---|
| Adoption (share of full benefit) | – | 40% | 80% | 90% | |
| Cash benefits (€) | 0 | 319,200 | 638,400 | 718,200 | 1,675,800 |
| Costs (€) | 380,000 | 323,400 | 323,400 | 323,400 | 1,350,200 |
| Net cash flow (€) | −380,000 | −4,200 | 315,000 | 394,800 | 325,600 |
| Discount factor | 1.000 | 0.909 | 0.826 | 0.751 | |
| Present value (€) | −380,000 | −3,818 | 260,331 | 296,619 | 173,131 |
| Cumulative net (€) | −380,000 | −384,200 | −69,200 | 325,600 |
- ROI over three years: €325,600 ÷ €1,350,200 = 24.1%.
- NPV: €173,131. Benefit-cost ratio: €1,357,379 ÷ €1,184,248 = 1.15.
- Payback: 2 + (69,200 ÷ 394,800) = 2.18 years after the go decision.
Sensitivity and switching values.
| Assumption | Base value | Switching value (NPV = 0) | Margin |
|---|---|---|---|
| Minutes saved per claim | 10 | 8.5 | −15% |
| Share of freed hours banked | 70% | 59% | −11 points |
| Adoption at each stage | 40%, 80%, 90% | 87% of plan | −13% |
| Recurring cost per year | €323,400 | €393,000 | +22% |
A low scenario, with 7 minutes saved, adoption 15% below plan and run costs 10% higher, gives an NPV of −€402,380. A high scenario with 12 minutes saved and everything else at base gives €401,743. Minutes saved has both the narrowest margin and the lowest confidence, so it is the assumption to test first.
Decision requested. Approve €380,000 one-off and €323,400 a year for three years, in two stages: €150,000 now for a pilot with 40 handlers, and the rest only if the pilot shows a median saving of at least 9 minutes per claim with rework at or below 7%.
How to review and update the business case after the pilot
When the pilot ends, replace estimates with measurements in six steps:
- Replace assumptions with measured values. Update every pilot-tested line in the assumptions log, with sample size and dates.
- Compare results with the switching values. Below a switching value, the case fails as written: change the scope or stop.
- Re-price run costs from real usage. Use measured tokens or GPU hours per task and the review minutes actually spent.
- Re-plan adoption. Pilot users are often volunteers, so plan the wider rollout at a lower rate unless usage data says otherwise.
- Leave the pilot spend out. It is sunk, and the Green Book says sunk costs must not be included in appraisal; the scale decision rests on future costs and benefits only.
- Return to the same approvers with the new NPV, payback, margins and a recommendation.
After go-live the case turns into a benefits register: each benefit line, its owner, the agreed measurement, and a quarterly actual against plan. The guide to measuring returns once AI is live covers that tracking, and the ROI calculator is a quick way to rerun one use case with new inputs.
Common AI business case mistakes
- Counting benefits twice. The same hours appear as time saved and again as lower rework, or two cases claim the same team. Trace each benefit line to one measurement and one budget line.
- No baseline. Without a measured before, no after can be proven. A supplier’s demo figure is not a baseline.
- Ignoring the adoption ramp. Users join in waves and habits change slowly. In IBM’s 2026 CEO Study, published on 4 May 2026, the surveyed CEOs said only 25% of their workforce uses AI regularly in their job.
- Ignoring run costs. Inference, monitoring, evaluation reruns, human review and support continue every year. About one in five respondents to McKinsey’s 2026 survey said AI operating costs, token costs included, had constrained their use of AI.
- Treating hours as cash. Spell out how freed time becomes money: avoided hires, less overtime, insourced work or extra revenue.
- A single point estimate. Without a sensitivity table, optimism bias goes unchallenged.
- A do-nothing option that costs nothing. Business as usual carries costs of its own, such as planned hires or a growing backlog.
How VDF AI fits
VDF AI does not write the business case, set the baseline or decide which hours finance banks. What a governed platform adds is evidence. Audit logs and usage visibility show who used each assistant or agent and how often, the raw material for the adoption ramp. Role-based access control on every plan, and a deployment in VDF’s managed cloud or in your own data centre, private cloud or air-gapped network, decide which data a use case may touch, and so whether some benefits are reachable at all.
VDF AI Router sends requests to approved cloud models or to registered on-premises model endpoints, with an air-gap mode that keeps traffic on local models, so a pilot can try more than one approved model on the same task before the run-cost line is fixed. For VDF’s own licence costs, use the seat-sizing calculator and the published cloud plan prices.
Sources
Each source below was opened and read on 4 November 2026 before its figures went into this template.
Frequently asked questions
What should an AI business case include?
Nine parts: the problem and a measured baseline, the options including doing nothing, quantified benefits with a measurement method for each, one-off and recurring costs, risks and controls, ROI, NPV and payback, a sensitivity analysis, an assumptions log, and the decision you are asking for. Put a one-page summary on top. Expect finance to test the baseline, the run costs and the sensitivity table hardest, so give those sections the most care.
How do you calculate ROI, NPV and payback for an AI project?
ROI is total benefits minus total costs, divided by total costs, over a stated horizon. NPV divides each year's net cash flow by (1 + r) raised to the year number and adds the results, where r is the discount rate set by your finance team. Payback is the point at which cumulative net cash flow turns positive. Use the same horizon and the same adoption ramp for all three.
What does a good AI business case example look like?
It shows its arithmetic. The worked example on this page, for a fictional insurer, starts from a time study of 400 claims, banks only 70% of the freed hours as cash, ramps adoption from 40% to 90%, and budgets full run costs from year one. The result is a three-year NPV of about €173,000 at a 10% discount rate, and the case breaks even if the time saved per claim falls from 10 to 8.5 minutes.
How is an agentic AI business case different?
An agent takes several steps and calls tools for each task, so cost and risk vary from task to task. Price the case per completed task, including retries, tool calls and human approvals, and count the tasks that still end with a person. Gartner forecasts that over 40% of agentic AI projects will be scrapped by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, so an agentic case needs a clear answer on each of those three points.
What are the most common mistakes in AI business cases?
Counting the same benefit twice, having no measured baseline, assuming full adoption from the start and leaving out run costs. Close behind are treating saved hours as cash without saying how they convert, presenting one point estimate with no sensitivity analysis, and including money already spent on the pilot, which is a sunk cost.
Can I download the AI business case template as a Word or PDF file?
Yes. The template is part of the free AI Implementation Template Pack 2026, a DOCX file you can edit in Word or Google Docs and save as a PDF for the board pack. It contains the nine sections, the cash-flow table, the sensitivity table and the assumptions log, without the guidance text from this page.
Collect the evidence your business case depends on
VDF AI keeps audit logs of assistant and agent use and applies role-based access on every plan, so a pilot leaves a usage record you can set against each assumption.