Questions to Ask Before Approving an AI Business Case

Questions about the problem

What decision or workflow needs improvement? Who experiences the problem? What is the current process, volume, cost, delay, error, and review burden? Which observations are measured, which are reported, and which are assumptions? What happens if nothing changes?

The UK government Green Book defines appraisal as assessing costs, benefits, and risks across options for achieving objectives (Green Book collection). NIST’s AI RMF Core recommends documenting purpose, context, alternatives, risk, metrics, and human oversight (AI RMF Core). These are decision disciplines, not proof that a proposed investment will work.

Sources: [1] [2]

Questions about options and feasibility

What non-AI process, narrower automation, policy change, training, or service redesign could address the problem? Which option is reversible? What data, integration, procurement, security, privacy, legal, accessibility, and operating requirements apply? Which product facts were verified for the current version, plan, and configuration?

Ask what must be true for each option to work. Record dependencies and owners. Avoid comparing an optimistic AI future with the worst version of the status quo.

Questions about value and cost

Which benefits are cashable, time-releasing, quality-related, risk-related, or experiential? What baseline, counterfactual, population, period, and attribution design support them? Which costs cover licenses, integration, data preparation, governance, review, support, change, monitoring, and exit?

Use ranges and sensitivity tests for uncertain assumptions. Do not report time saved as financial value unless the organization has a credible plan for how released time changes capacity or cost.

Worked question set

Imagine a synthetic proposal team considering an AI evidence-mapping assistant. The initial case claims 30% faster reviews. The question set reveals no measured baseline and no source for the percentage. It also identifies a manual spreadsheet improvement as a lower-cost option.

The revised case proposes a bounded pilot. It measures review time, missed requirements, corrections, and reviewer effort on synthetic tasks. It records implementation cost, data boundaries, stop conditions, and who decides whether evidence supports expansion. The 30% claim is removed.

Questions about risk and learning

What errors could propagate? Who remains accountable? Which controls need operating evidence? What conditions pause the use case? What will the pilot teach, and which result would reverse the recommendation? How will model, vendor, data, and workflow changes trigger reassessment?

Questions about the decision record

Who has authority to approve the next stage, and which specialists must review the evidence? What is being approved: discovery, a technical test, a bounded pilot, or operational expansion? Which conditions travel with the decision? Where will assumptions, dissent, data limitations, and rejected alternatives remain accessible after the presentation ends?

Require an evidence register that links each material business-case statement to a source, calculation, named assumption, or owner judgment. Record the source date and the person responsible for refreshing it. A financial model should expose formulas, ranges, and sensitivity to adoption, review effort, error handling, and exit cost. The decision log should state what new evidence would trigger reconsideration and when the case expires. This prevents provisional reasoning from becoming a permanent fact through repetition.

Approval checklist

These questions turn a business case into a testable decision. They do not establish return, productivity, or outcome improvement before evidence exists.

• The problem and baseline are evidenced.
• AI and non-AI options are compared fairly.
• Costs include implementation and operation.
• Benefits have owners, measures, and realization paths.
• Uncertainty and sensitivity are visible.
• Risks, controls, stop conditions, and exit are defined.
• Approval authority and next evidence gate are explicit.