The AI ROI number a board can actually challenge
August 27, 2026
The board does not need a heroic AI ROI percentage. It needs a number with a clear baseline, a credible counterfactual, a quality floor, and an owner. The most useful unit is contribution margin per successful outcome: realized revenue or avoided cost minus the fully loaded cost of producing the outcome.
Start with the decision
State what decision the ROI analysis supports: fund expansion, renew a provider, automate a workflow, or stop an experiment. The decision determines the period, cohort, and counterfactual. A generic ‘AI saved money’ claim is not a capital allocation case.
Show the bridge
Bridge the baseline to actual through adoption, price, model mix, productivity, and quality. Separate realized cash savings from avoided hiring, faster throughput, and option value. Do not put hypothetical future benefits in the realized number.
Protect against false ROI
Pair the financial result with successful-task rate, customer or employee acceptance, error rate, and human-repair time. A cheaper workflow that creates downstream review is not cheaper. Make the quality floor explicit and report the p50 and p95 cost per successful outcome.
Make assumptions auditable
Show provider invoices, usage volume, rates, labor assumptions, revenue attribution, and the dates on which each assumption changed. Use conservative, base, and upside cases. The board can accept uncertainty when it can see where uncertainty enters the model.
A defensible AI ROI number is narrower than a marketing claim, but it is far more useful: it tells the board what to fund, what to measure, and what would change the decision.
Related
- Proving AI ROI
- AI spend variance analysis
- Reporting AI spend to the board