How do I prove AI ROI to my board?

You do not prove AI ROI to your board in the way you prove ROI on a new warehouse. You present a position, not a projection. The structure that survives contact with a sceptical CFO has four parts: what you spent and on what, what changed that you can evidence, what you learned that changes the next decision, and what you are asking for next and on what condition it gets stopped. That order matters. It puts cost first, which buys you credibility, and it ends with a kill criterion, which is the thing boards almost never get offered and almost always want.

The reason the normal approach fails is that the normal metrics do not fit. EY found that 71% of CFOs believe common ROI metrics are ill suited for measuring gains from AI and other emerging technologies (CFO Dive). So when you walk in with a payback period and an IRR on a generative AI pilot, roughly seven in ten of the people you are talking to already suspect the number is theatre. They are right. Build the presentation on that assumption instead of fighting it.

Start with the number that makes the room honest

PwC’s 2026 CEO Survey found that 56% of CEOs report neither increased revenue nor decreased costs from AI investments in the last 12 months, and only 12% report achieving both (PwC 2026 CEO Survey).

Put that on the first slide. Not as an excuse. As a baseline.

It does two things. It tells the board that if you are in the 32% who achieved one of the two, you are ahead of the room, and if you are in the 56% you are in the majority and the question is what you do about it. It also pre-empts the comparison that is otherwise coming, which is some competitor’s press release about an AI deployment that has never been audited by anybody. Boards benchmark against announcements because announcements are what they can see. A named, sourced, recent survey of chief executives is a better benchmark than a rival’s LinkedIn post, and it is one they cannot argue with in the meeting.

The second number worth having to hand: only 35% of finance leaders believe their organisations effectively measure ROI from AI initiatives, despite 77% already using AI in operations (Founder Operator). Adoption has outrun measurement almost everywhere. Your board is not asking you a question anyone else has answered well.

Part one: what you spent, itemised, including the parts nobody budgeted

Lead with cost. It is counterintuitive and it works, because the first instinct of a board hearing an ROI case is to discount the benefit side and go looking for hidden spend. Give them the hidden spend yourself.

Itemise licences, implementation, and the two lines most AI business cases quietly omit: internal time and data remediation. Internal time is the hours your people spent prompting, checking, correcting and re-running. Data remediation is what you had to clean before the tool could do anything useful, and in most mid-sized businesses it is the largest single number on the page. If you have not spent anything on data remediation, say so, because it usually means the pilot was run on a narrow, already-tidy dataset and will not generalise.

One honest line beats three optimistic ones. If a licence is being paid for by a department that is no longer using it, put that on the slide. Shadow AI spend in a 500-person company is rarely visible in the finance system because it goes through expenses as a £20 monthly subscription, and a board that discovers it later will discount everything else you said.

Part two: what changed, evidenced, in three tiers

Separate your claims by how hard they are. Do this explicitly on the slide, because the mistake that kills credibility is presenting a soft claim with the same confidence as a hard one.

Tier one, directly measured. Things you counted before and counted after. Tickets resolved without escalation. Days to close. Error rate on a specific process. These are the only numbers you should defend under challenge.

Tier two, reasonably attributed. Outcomes that moved, where AI is a plausible contributor among several. Say it that way. “Conversion improved 4%, over a period in which we also changed the pricing page” is a stronger sentence than “AI improved conversion 4%” because the board will find the pricing change anyway.

Tier three, observed but not yet measurable. Capability you now have and did not before. People who can do something they could not do in March. This belongs on the slide labelled as what it is. Do not convert it into a pound figure. The moment you put a fabricated number next to a real one, both become suspect.

Most AI value in a mid-sized business in year one sits in tiers two and three. That is not a failure. It is what the early phase of any capability build looks like. The failure is pretending it is tier one.

Part three: what you learned, and why it changes the next decision

This is the part boards actually reward and almost nobody includes.

Infosys found that 72% of senior executives at large US firms say fewer than 25% of their AI pilots reached full production while delivering intended ROI (Infosys via O-Mega). The board has read a version of that statistic. They are braced for you to explain away a stalled pilot.

Do not explain it away. Most pilots that “fail” performed exactly as designed. What failed was the decision to run them on foundations that could not carry production volume. The four failure modes behind pilots that never scale are almost always upstream of the tool: data that cannot be assembled reliably, a process that three people run three different ways, no named owner with authority across the functions involved, and no governance anyone has formally assessed.

So say what you found. “We learned that we cannot produce a clean customer list on demand, which means the personalisation use case is blocked until that is fixed, and here is what fixing it costs.” That is a finding with a price attached. It converts a stalled pilot from a write-off into paid-for diagnostic work, which is a far better position than the one most executives walk into the room holding.

On governance, the gap is wide enough to be worth naming. Deloitte found that close to three-quarters of organisations plan to deploy AI agents within two years, while only 21% report a mature governance model for agentic AI (Deloitte 2026 State of AI). If your board asks whether you could pass an audit of how your AI systems reach decisions, and the honest answer is no, it is better to be the person who said so first. Compliance and governance are not the same thing, and confusing the two is how a business ends up compliant with regulations it has never applied to a system nobody is formally accountable for.

Part four: the ask, with a stopping condition

End with a specific request and the condition under which you will come back and recommend stopping.

“£60,000 over two quarters. If the finance close is not three days shorter by the end of Q2, we stop and I will tell you so.”

Boards approve money for people who will tell them when to stop spending it. The stopping condition is the single most persuasive thing in the deck, and it costs you nothing except the discipline to honour it. It also disarms the argument that AI spend is open-ended. It is open-ended only when nobody has defined the end.

Name the owner in the same breath. Not a committee, a person, with authority over data, process and tooling for the area in question. Whether that is a chief AI officer or an existing operator with the remit extended is a separate question, and for most businesses of this size the answer is not a new C-suite title. But the board should leave the room knowing a name.

Why readiness is the honest answer when return is not available yet

There is a version of this meeting where you have no tier one evidence at all. Nine months in, several tools bought, nothing you can defend with a number.

In that case, do not invent a return. Present a position.

Boards conflate readiness and maturity. Readiness is whether you can start something and have it survive. Maturity is how far you have got. Reporting maturity when you have not established readiness is how a business ends up with a maturity score of three on a scale of five and a customer database nobody trusts. If you cannot show return yet, show the board where the foundations are weak, what that specifically blocks, and what it costs to clear. That is a board paper. “We are exploring AI opportunities” is not.

The six things to have a position on before you walk in

The free AI readiness audit is 30 questions across six pillars, five questions each, and takes 7 minutes. Data, Process, People, Technology, Strategy, Governance. You get a readiness score out of 120, a band for each pillar, and a short synopsis. It does not give you an action plan, and it does not inspect your systems. Every answer is yours.

The value for a board meeting is that it forces a position on six things in the order a board will ask about them. Two of the questions are worth answering before you read any further.

If you needed a clean, accurate list of your best 500 customers right now, how confident are you in what you would get? And if you asked three different people in your business to complete the same task, how similar would their approach be?

If those two answers are uncomfortable, you have found what is actually limiting your return, and you have found it in under a minute. Measuring readiness before you spend anything is cheaper than measuring return after you have.

A score out of 120 with a weak band on Data is a defensible board slide. It says here is where we are, here is what it blocks, here is what it costs to fix. Boards can work with that. They cannot work with a payback calculation that seven in ten CFOs already think is the wrong instrument.