The board is asking for AI ROI. What do I tell them?

Tell them the truth in three parts, in this order: here is what we spent, here is what we can and cannot yet evidence, and here is the dated plan to close the gap. Do not open with the plan. Do not open with a definition of agentic AI. Open with the number you spent, because that is the number they already have in front of them, and the fastest way to lose a board is to make them feel you are managing them rather than reporting to them.

The second thing to tell them, once the room has absorbed the first part, is that this is the normal outcome rather than a local failure. PwC surveyed 4,454 CEOs across 95 countries and found 56% report seeing neither revenue gains nor cost reductions from their AI initiatives, with only 12% achieving both. That is not cover. It is context, and there is a difference. Cover is “everyone’s struggling, give me more time.” Context is “the distribution of outcomes in this market is known, here is where we sit in it, here is what separates the 12% from the 56%, and here is which of those things we have and have not done.”

Why the board is asking now, and what they are actually asking

Boards do not usually ask about AI ROI because they have developed a sudden interest in machine learning. They ask because someone on the board reads the same trade press you do, an audit committee flagged an unbudgeted line, or an investor asked them the question first and they had no answer.

BCG’s survey of CEOs and boards found the two groups are measuring different things. CEOs estimate 35% of their performance evaluation depends on achieving AI ROI, against boards’ estimate of 27%. An eight-point gap sounds academic until you are the one sitting in the chair. It means you are answering a harder question than the one being asked, and answering it defensively, which makes an ordinary reporting item feel like an inquest.

So clarify the question before you answer it. Ask: are you asking whether the spend was justified, or whether we know what we are doing? Those need different answers. The first is arithmetic. The second is governance, and the governance answer is the one that actually settles a board down.

Part one: what we spent

Bring the full number. Not the AI project budget. The full number.

That means the licences, including the ones bought on departmental cards. The consultancy days. The internal time, costed at a defensible rate rather than treated as free because it was salaried. The integration work that got booked to IT rather than to the AI line. If you have shadow AI in the business, and at 500 people or more you almost certainly do, estimate it and say you are estimating it.

A board can forgive a number that is higher than they expected. They struggle to forgive a number that grows between meetings because you only counted the obvious parts the first time. Give the total, then give the breakdown, then say plainly which parts of it you are confident in and which are estimates. The estimate you flag as an estimate costs you nothing. The estimate you present as fact and then revise costs you the room.

Part two: what we can and cannot yet evidence

Split your outcomes into three columns and put them on one slide.

Evidenced. Something changed, you measured it, and you can name the before and after. A process that took four hours takes ninety minutes, measured across sixty instances. Ticket deflection at a rate you can pull from the helpdesk. If you have any of these, lead with them, and be specific about the measurement window. One month of good numbers is an anecdote.

Plausible but unproven. People say it is faster. Usage is high. Nobody wants to give the tool back. This is real information and you should not pretend otherwise, but you must label it as what it is: self-reported, unbaselined, and therefore not something you would put in a set of accounts.

No return yet. The pilot that ran well and stopped. The licence nobody logs into. Say so. Name it.

The third column is the one that earns you credibility, and most leaders cut it before the meeting. Do not cut it. A board that sees you volunteer your own dead projects stops wondering what else you are not telling them, and that shift is worth more than any of the numbers in the first column.

The usual reason column one is thin is not that AI does not work. It is that nobody recorded a baseline before switching the tool on. You cannot evidence a 30% improvement in a process nobody timed. This is boring, it is not a technology problem, and it is fixable in weeks rather than quarters.

Part three: the dated plan to close the gap

This is where most leaders reach for a strategy. Do not. A board that has just heard you cannot prove a return will not be reassured by a document promising a bigger version of what did not work.

What settles a board is dates and names. Something like:

• By the end of next month, a named owner for every AI tool currently in use, and the list of tools with no owner, which we will cancel.
• By the end of the following month, baselines recorded for the three processes we intend to measure. Actual timings, actual error rates, actual cost per transaction.
• By the next board meeting, a scored readiness position across data, process, people, technology, strategy and governance, with the gaps ranked.
• Ninety days from today, either evidenced numbers on those three processes or a documented reason they could not be produced.

Four items. Each has a date, an owner and a deliverable that either exists or does not. No item says “explore”, “align” or “build capability”. A board can hold you to a baseline. It cannot hold you to an alignment.

Notice that none of those four items requires buying anything. That is deliberate, and it is the single most persuasive thing in the plan. You are telling the board the next ninety days of AI work costs approximately nothing, because the gaps that stop AI paying off are almost never the model.

The numbers that show you are not uniquely behind

Use these sparingly. Two well-placed figures beat six, and a board that senses you are hiding behind benchmarks will stop listening.

On pilots: a March 2026 survey of 650 enterprise technology leaders found 78% of enterprises have AI agent pilots but under 15% reach production. If your successful pilot did not scale, you are in the overwhelming majority, and the reasons are structural rather than unlucky. The four failure modes are well documented and a pilot passes precisely by avoiding all of them, running on hand-picked data with hand-picked people and no governance overhead.

On the overall distribution: RAND’s breakdown puts 80.3% of enterprise AI projects failing to deliver promised business value, split as 33.8% abandoned before production, 28.4% reaching production but underdelivering, and 18.1% running but never recovering the investment. The useful part for a board conversation is the split. Which of those three are we? They are different problems with different fixes, and saying which one you are is more impressive than any recovery narrative.

On governance, if the board’s real question is about control rather than return: Grant Thornton’s 2026 AI Impact Survey found 78% of business executives lack strong confidence they could pass an independent AI governance audit within 90 days. If a non-executive asks whether you could pass one, the honest answer is probably no, and the answer that follows it should be the date by which you could.

Do not make these four mistakes in the room

Do not promise a return by a date you cannot control. “We’ll see returns by Q2” is a hostage. “We’ll have measured baselines by Q2” is a commitment you can keep.

Do not blame the vendor. You bought it. Schellman’s 2026 State of AI Governance Report found 42% of organisations place AI purchasing and adoption authority with the CIO or head of IT, and only 10% with the CEO, and in most of those organisations the same executive who approves the tool is the one blamed when it creates a compliance failure. If that describes your business, say it out loud as a structural finding rather than letting the board discover it during an incident. Ownership of AI outcomes is a board-level design decision, not an IT staffing one.

Do not ask for more budget in the same meeting. You have just reported that you cannot evidence the last spend. Asking for the next one in the same breath tells the board you have not changed your thinking. Ask at the meeting where you bring evidence.

Do not bring a maturity model. Five-stage frameworks with names like Exploring, Emerging and Optimising will be read, correctly, as a consultancy deck. Bring a score and the gaps behind it.

What to bring to the next meeting instead

The most useful thing you can put in front of a board at the following meeting is a scored position they can compare against, rather than a narrative they have to take on trust. A number that moves is an argument. A paragraph that says “we have made good progress” is not.

That is what the free AI readiness audit produces. Thirty questions, 7 minutes, and a score out of 120 with a band for each of the six pillars: Data, Process, People, Technology, Strategy and Governance, plus a short synopsis. The questions are the unglamorous ones, and that is the point. Could you produce a clean, accurate list of your best 500 customers right now? Can your key systems share data without someone exporting spreadsheets? Who is responsible for the quality of your data? Every answer is self-reported, so it is a structured account of your own position rather than a systems scan, which is exactly what a board paper is anyway.

Run it yourself, then have your CFO, your operations lead and your IT director each run it separately without comparing notes. Where their scores diverge by more than a band, you have found the gap that is actually costing you, and it is usually not the one you were about to fund. Bring the four scores to the board and let them read the disagreement. It will tell them more about your readiness than any slide you could write about it.

One closing thought on the conversation itself. The board is not testing whether you have AI returns. Most of your peers do not. They are testing whether you know, precisely and without prompting, why you do not and what you are doing about it. That is a question you can answer honestly today, at no cost, and it is a better answer than most of them are getting.