How to measure AI ROI

The AI ROI formula, and why it is the easy part

Every AI ROI framework lands on the same arithmetic: add up the value a bet creates, added revenue, saved cost, and avoided cost, subtract the total cost of ownership (build, run, licences, and change management), and divide by that cost. Hard metrics like hours saved and error reduction sit alongside softer ones like adoption and risk.

The formula is not where teams get it wrong. They get it wrong on the inputs. Value and adoption are estimated optimistically in a business case, then the bet is funded, and the real numbers only arrive after the money is spent. A precise formula on guessed inputs still gives you a confident, wrong answer.

Why measuring AI ROI after you build is too late

The stakes are well documented. RAND finds more than 80% of AI projects fail, roughly twice the rate of non-AI IT projects, and McKinsey’s 2025 survey shows that while 88% of organisations use AI, only 39% report any enterprise EBIT impact. When ROI is measured only after launch, the failures are discovered after the capital is gone.

MIT’s NANDA initiative measured the same pattern from the other side: 95% of enterprise GenAI pilots never move the P&L. The problem is not that teams cannot do the ROI maths. It is that they run the numbers on the wrong side of the spend.

How to measure AI ROI before you build: pretotyping

Pretotyping, created at Google by Alberto Savoia and taught at Stanford, is how you get trustworthy ROI inputs before you commit. Instead of estimating demand and adoption, you measure them with fast behavioural tests, often with no production AI at all. A few proven patterns do most of the work:

• Wizard-of-Oz: a human quietly performs the AI’s job so you can measure whether the output is actually valued and used.
• Concierge: you deliver the outcome manually, end to end, to learn what good looks like before automating it.
• Fake door: you advertise the AI capability and measure real demand before any of it exists.
• Pinocchio: a non-functional stand-in that tests whether people will adopt the workflow at all.

From one ROI number to a governed portfolio

Measuring one bet is a start; the real job is keeping every bet honest. That is why pretotyping is the engine inside AI governance: it makes evidence cheap enough that the board can demand proof of ROI before capital without grinding delivery to a halt.

In a two-week AI Bets Audit I map and rank your portfolio, then pretotype the highest-stakes bets to replace their biggest assumptions with behavioural evidence, so each fund, fix, or kill decision rests on a defensible ROI number, not a debate. It is run by operators who have executed 4,000+ experiments, and Rapidly, my software, keeps the ROI picture live as new evidence arrives, not frozen in a consultant’s deck.