The 10-minute test for whether an AI use case has a business case

An AI business case is not a forecast of how much work a model can perform. It explains how that performance will change a business outcome—and what evidence would prove it.

A team arrives with persuasive arithmetic. Its workflow handles 12,000 requests a year. AI could save three minutes on each request. Multiply the hours by a loaded labor rate, subtract the model bill, and the project appears to pay for itself. The spreadsheet may be correct while the investment case remains unproven. Nobody has said whether the freed minutes will reduce overtime, clear a revenue-generating backlog, prevent a loss or simply disappear into the working day.

Ten minutes cannot establish ROI. They can reveal whether a use case contains the ingredients of a business case at all.

The missing link is not a better formula

An AI output must be accepted as usable, change a workflow, alter a business outcome and meet an operating decision that captures the benefit. Each link is a hypothesis, not an automatic consequence. ⁣DIRECTIVE:0⁣

The UK's May 2026 guidance on evaluating AI interventions distinguishes technical performance from the impact of an intervention. It calls for a theory of change, a baseline and a comparison with what would have happened otherwise.¹ That is public-sector guidance, not a corporate ROI formula, but the distinction travels well.

Minutes 0–3: identify the unit that changes

Begin with a unit the business recognizes: a resolved claim, qualified lead, accepted design change, completed onboarding or prevented exception. Specify how AI changes that unit after review, corrections and exceptions. “Five minutes saved per employee” is not yet business value. “Twenty percent more claims closed within the same service window without increasing error rates” is a testable operating claim.

Without a unit, baseline or acceptance condition, the project may still merit exploration. It is too early to claim a return.

Minutes 3–6: show who can capture the change

In the 12,000-request example, three minutes saved per request represents 600 hours of gross capacity. If only 75% of requests are eligible and accepted, that falls to 450 hours. Deduct 120 hours of oversight and corrections: net capacity is 330 hours. At an illustrative loaded rate of $60 per hour, that is $19,800 of capacity equivalent—not $19,800 of cash savings.

If staffing, throughput and outcomes stay unchanged, the financial line has not moved. A manager could eliminate contractor expense, serve profitable demand without hiring, reduce penalties or redeploy capacity to measurable work. Each is a different case.

Name the person authorized to change staffing, procurement, backlog priorities or service commitments. If the claimed benefit depends on a decision nobody owns, the case has a missing link.

A randomized field experiment across 66 firms and 7,137 knowledge workers found that generative AI access reduced email time among users by roughly two hours per week, but did not detect corresponding changes in the quantity or composition of tasks from individual access alone.² Saved time is not worthless; it should not automatically be booked as organizational transformation.

Minutes 6–9: identify the counterfactual and full cost

A benefit is the difference between the intervention and what would plausibly have happened without it: the counterfactual. It may be estimated using a control group, staged rollout, matched comparison or documented baseline with explicit limitations. Do not invent the comparison after results arrive.

Include integration, data preparation, evaluation, monitoring, supervision, exceptions, security review, vendor dependency and recovery costs. Count work displaced into other teams. An accepted output must proceed through the business process, not merely pass a model benchmark. ⁣DIRECTIVE:1⁣

Minute 9–10: decide what happens next

Proceed to a full business case when the value unit, capture decision, owner, baseline and costs are credible enough to model. Run a bounded experiment when one decisive assumption remains uncertain, with a comparison and stop condition. Stop or redesign when the path ends at activity or nobody can make the change needed to capture value.

This differs from The biggest AI mistake companies make is starting with AI, which asks whether AI beats alternatives, and Your AI program can be efficient and still be strategically irrelevant, which asks whether the project deserves portfolio resources. The ten-minute test asks: is there a believable route from technical performance to captured value?

McKinsey's 2026 global survey reports that 80% of respondents say AI improved their individual productivity, while 37% report a positive contribution to organizational EBIT.³ These are self-reported survey measures, not causal estimates of a particular project.

A business case becomes credible when its chain is explicit enough to fail: a baseline someone can inspect, an output someone can reject, a cost someone can challenge and an operating decision someone has authority to make. That is what the ten minutes are for.

Sources and evidence notes

Footnotes

1. HM Treasury / UK Evaluation Task Force, Guidance on the Impact Evaluation of AI Interventions, updated 15 May 2026. https://www.gov.uk/government/publications/the-magenta-book/guidance-on-the-impact-evaluation-of-ai-interventions-html ↩

2. Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI, NBER Working Paper 33795, May 2025, revised November 2025. https://www.nber.org/papers/w33795 ↩

3. McKinsey, The State of AI in 2026: On the road to ROI, August 2026; self-reported survey, n=1,719. https://www.mckinsey.com/industries/industrials/our-insights/cutting-the-coordination-tax-how-agentic-ai-can-reshape-workflows ↩