My CEO wants ROI numbers on AI. How do I measure that?

Your CEO wants a number on what the AI spend is actually returning. And the four-hours-saved slide that used to carry the room just got a raised eyebrow. You have presented it twice before to nods and a "good, keep going". This time someone asked, quite reasonably, where those hours show up in the accounts, and you did not have the answer ready.

So here is the answer, before the diagnosis. To measure AI ROI in a way a board will accept, you stop reporting activity and start reporting the language the board already reads: payback period, net present value, internal rate of return, total cost of ownership, with revenue movement and margin as the headline. Productivity is an input. The board is asking about outputs. The whole reason the slide stopped working is that you were answering a financial question with an operational metric, and the gap between the two has become the entire conversation.

Why did "four hours a week saved" stop being a business case?

Because the market priced it in and the board has caught the same signal. The shift is documented. In the Futurum Group's 1H 2026 survey of 830 IT decision-makers, direct financial impact, revenue growth and profitability taken together, nearly doubled to 21.7% as the primary measure of AI return, while productivity gains fell from 23.8% to 18% as the leading metric. The plainest version: the era of productivity metrics as AI ROI has closed.

There is a second, quieter reason. The hours often do not survive contact with reality. A team saves four hours generating a draft, then spends three of them correcting errors, rewriting weak output, and verifying what the model produced. The time-saved figure is real on the slide and gone in the workflow. Gartner's own data points the same way: in a survey of 724 respondents, teams using generative AI reported marginally lower high-productivity gains, 34%, than teams using traditional AI at 37%. Productivity is hard to measure and does not automatically become organisational value. So when your CEO raises an eyebrow, they are not being difficult. They are reading the same reports the analysts read, and they want to know whether the hours turned into money.

What numbers does a board actually find credible?

The ones already on the page when they read any other capital decision. You are not inventing a special AI dashboard. You are translating the AI case into the four measures the board uses for a new plant, an acquisition, or a systems migration, and letting the operational metrics sit underneath as supporting evidence, not as the headline.

That TCO line is where most cases quietly leak, and the timing is where they get exposed. Deloitte's 2025 survey of 1,854 senior executives found that only 6% report an AI payback period under one year; for most organisations satisfactory returns arrive across two to four years, far slower than the seven-to-twelve months a board expects from ordinary technology. The 72% of CIOs who report breaking even or losing money are rarely overpaying for models. They are underpricing the change, then meeting the real bill later. Build the full picture and your number becomes defensible. Hide the hidden costs and your CEO finds them in a worse meeting than this one.

How do I measure AI ROI without measuring the wrong thing?

You point the measurement at where value is actually created, and the research is unusually clear about that location. Deloitte found that 66% of organisations report productivity and efficiency gains from AI, yet only 20% are already growing revenue, while 74% are still only hoping revenue will follow. The gain is real and the money has not landed, because the gain sits in activity and never gets routed into a revenue line or a cost line. So a dashboard that tracks model cost and tokens measures the input and ignores the change that produces the output. The metrics that predict returns track behaviour: how many people actually adopted the tool, which workflow changed shape because of it, and what that change did to a number you can point to in the accounts.

This is the part worth keeping in view. The MIT NANDA report found that about 5% of integrated AI pilots reach rapid revenue acceleration, and Deloitte's separate Q4 2025 CFO Signals survey of 200 finance chiefs at billion-dollar companies found that only 21% of active AI users report clear, measurable value. The leaders in that minority are not the ones with the best models. They are the ones who changed how people work and then measured that change in money. The bottleneck is no longer the technology. It is whether the organisation around the technology actually shifted, and whether you instrumented that shift so it shows up in the accounts.

Here is the sequence to carry into the room:

1. Name the financial outcome first. Pick one revenue line or one cost line the AI initiative is meant to move, and state it before you mention any tool.
2. Build the full TCO. Add the retraining, the integration, and the change management to the licence, so the denominator stays honest.
3. Translate into payback, NPV, IRR. Set the return against your hurdle rate so AI is judged on the same terms as any other capital decision.
4. Instrument adoption and workflow as the leading indicators. These tell you, months before the profit-and-loss does, whether the human and process layer is moving.
5. Report movement, then attribution. Show what changed in the accounts, then show your working on how much of it the AI earned.

Do that and you give your CEO the thing they were really asking for: not a defence of the spend, but a credible claim that the next dollar put in returns more than a dollar. That is a number a board can stand behind, and a position you can keep extending quarter after quarter.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.