The honest math of AI ROI
Most AI business cases compare a subscription to a miracle. The arithmetic that survives audit: unit costs, time reclaimed, and the error bill.
The AI business cases that cross my desk share a structure: precise costs, miraculous benefits. The subscription is priced to the cent; the return is "20% productivity uplift" sourced from a vendor webinar and applied to the entire payroll. Finance signs it because the number is large, then remembers it forever, because the number was fiction. Here is the arithmetic that survives an audit.
Cost, fully loaded. Licenses and API spend, which is the easy column (the meters are public; per-token pricing even rewards well-engineered workloads with caching discounts). Then the honest lines: engineering time to integrate, the review time humans spend checking output, and the error bill: the reworks, the wrong answer that reached a customer. The review line is the one every optimistic case omits, and it is frequently the largest.
Benefit, in reclaimed hours you can name. Not payroll percentages. Specific workflows, measured against the baseline you took first: the report that took a day and takes an hour, the backlog that stopped growing. Hours are only money when the freed time has a destination, so the business case must say what the destination is. "Support handles 30% more volume without hiring" is a claim finance can check in two quarters. "Everyone is 20% faster" is a horoscope.
The portfolio view. Some deployments will return nothing; that is fine if the wins are measured and the losers are killed. What is not fine is the blended narrative where one good deployment's story subsidizes four unmeasured ones indefinitely.
One more line item the honest ledger needs: the cost of measuring. A pilot with a real baseline, a control group and a quarterly read costs perhaps a tenth of the pilot itself, and executives sometimes balk at it, which is how you learn what the enthusiasm was made of. A team unwilling to spend ten percent proving the value was never expecting the value to survive proof. Fund the measurement first and the rest of the portfolio starts disciplining itself, because everyone builds differently when they know the number is coming.
The uncomfortable summary: AI ROI is usually real and usually smaller than the deck said, and companies that measure honestly end up deploying more of it, because their second business case gets believed.