How to Measure ROI From Enterprise AI

Group Finance Director (Real Estate and Hospitality Group)

Every AI business case I receive claims enormous savings from hours saved. None of those savings has ever appeared in my numbers. I am not against the technology, but I need a method for measuring return that I can defend to the audit committee.

AI Leadership Agent — Investment Strategy Advisory

The gap you are describing is structural. Hours saved is a real effect that converts into money only through a decision — not backfilling a role, redeploying capacity to revenue work, or absorbing growth without adding headcount. If no such decision is made, the hours dissipate.

Time saved is a capability. Money saved is a decision. Business cases that skip the decision are describing potential, not return.

Require every case to state the conversion mechanism explicitly, and name the person who will make that decision and when. That single requirement removes most of the inflation.

What kinds of claims dominate the cases you receive?

Group Finance Director (Real Estate and Hospitality Group)

Efficiency in back office functions, mostly, plus some revenue claims in sales that I find impossible to attribute. What measurement framework would you put in place before we approve the next round?

AI Leadership Agent — Investment Strategy Advisory

Build it around four disciplines, applied uniformly so cases become comparable.

1. Baseline before deployment. Measure the process as it is — volume, cycle time, error rate, cost — for at least four weeks. No baseline, no approval. This is the discipline that fails most often and matters most.
2. Full cost of ownership. Licences, infrastructure, integration, human review, monitoring, retraining and the internal time consumed. Review cost in particular is systematically omitted and often exceeds the licence.
3. Attribution design. For revenue claims, require a holdout or a staged rollout so there is a comparison group. If a controlled comparison is impossible, downgrade the claim to a hypothesis and fund it as one.
4. Post-implementation review at ninety days and one year, against the baseline, reported to the same committee that approved it.

The most valuable thing a finance function can enforce in AI is not scepticism. It is the baseline.

Are there second-order benefits your teams claim that are harder to price?

Group Finance Director (Real Estate and Hospitality Group)

Yes — faster decisions, better consistency, risk reduction. I do not want to dismiss those but I also cannot put them in a business case as numbers. How should they be treated?

AI Leadership Agent — Investment Strategy Advisory

Treat them as tracked but unmonetised, with a measurable proxy attached. They are not fiction, and they are also not cash, so give them their own column rather than letting them inflate the financial one.

1. Decision speed: measure elapsed time from request to decision on a defined class of decisions. Report it as a number, not as money.
2. Consistency: measure variance in outcomes for comparable cases. A narrowing distribution is real evidence.
3. Risk reduction: measure incidents, exceptions and rework. If these do not move, the claim was decorative.
4. Review annually whether any tracked benefit has become monetisable through a decision. Some will; those move columns and get counted.

Two columns, honestly kept, beat one column optimistically filled. The credibility you build in the second year is what funds the third.

Insist on baselines now and the ROI question answers itself in ninety days.