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

Tell them the truth, which is that you cannot evidence a return yet, and that this puts you with the majority rather than behind it. BCG found only 45% of surveyed finance executives could quantify the return from their AI investments, and among those who could, the median reported return was 10% against a 20% target. So the honest answer to the board is two sentences long: we do not have an attributable return figure, and neither do most finance functions asking the same question this quarter.

Then change what you are being asked. The board question “what did AI return” is unanswerable at month nine of a pilot, because the thing being measured has not run long enough, through enough of the business, with enough clean data behind it, to produce a number anyone would defend. The question you can answer is “which preconditions for return have we met, which have we not, and who owns closing the gap”. That is a position, not an excuse, and it survives follow-up questions. The rest of this piece is how to build it before the pack is due.

Why the number you want does not exist yet

There are three separate reasons an AI return figure is hard to produce, and it helps to know which one is biting you, because the board will ask.

The first is attribution. If a support team’s first-response time improved 18% in a quarter during which you deployed an assistant, hired two people and changed the triage rota, the assistant owns some fraction of that. Nobody knows which fraction. Finance will not sign a number built on a guess, and they are right not to.

The second is timing. Most of the cost lands up front and most of the benefit, if it arrives, arrives later and spread thin. The MIT figure everyone quotes, that 95% of generative AI pilots fail, used a roughly six-month success window and was preliminary rather than peer-reviewed. Six months is a short window to judge anything that touches process. If you measure a pilot at month six and call it a failure, you may just be measuring a pilot at month six.

The third is that the benefit often shows up as capacity rather than cash. An hour saved is only an hour returned if something valuable fills it. Most of the time nothing does, because nobody planned what the freed hour was for. That is not an AI problem. It is a management problem that AI made visible.

I have watched this pattern for a long time and written up the five failure patterns that keep repeating. None of them are technical.

What you can actually put in the pack

You can evidence four things today without inventing anything. None of them is a return figure. All of them are harder to argue with.

Spend to date, broken out. Licences, integration, internal time, consultancy. Internal time is the one people leave out because it is uncomfortable. Put it in. A board that sees an honest cost line trusts the rest of the page.

Usage, not seats. How many people hold a licence against how many used it in the last 30 days. If you have bought 400 seats and 90 people logged in last month, that is your headline, and it is far more useful than a return estimate. It tells the board exactly where the money is going and exactly what to fix.

One narrow claim you can defend. Not “AI saved us £400k”. Something like: the contracts team processed 1,100 renewals last quarter against 820 the quarter before, with the same headcount, and the only change was the clause-extraction tool. You may only have one of these. One defended claim beats five estimated ones.

What you have not met. This is the part most people skip and the part that earns the room. Say which of your core processes are documented well enough for a new starter to follow unaided. Say how many manual spreadsheet exports sit between your two most important systems. Say who is responsible for the accuracy of the customer data you would be feeding into any of this. If the answer to that last one is nobody, that is a finding, not a confession.

Reframe the ask: preconditions, not returns

Here is the sentence to use. “We cannot evidence a return yet. What we can evidence is our position against the conditions that make a return possible, and here is where we fall short.”

That reframe works because it is checkable. A return estimate invites a sceptical CFO to pull the thread. A readiness position invites them to look at the same six things you looked at. The six that matter are data, process, people, technology, strategy and governance, and I have set out what a readiness audit actually measures under each one.

Use it to do something uncomfortable but useful: tell the board what you would need to be true before you could promise a return, and then say whether it is true. Most organisations find that three or four of the six are weak, and that the weak ones are cheap to fix relative to the AI spend already committed. Dirty customer data costs less to clean than a year of licences nobody uses, and dirty data is the cheapest mistake to fix first.

The ownership question will come up, so get there first

At some point someone will ask who is accountable. Answer before they ask.

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. Nearly four in five. If your board hears that number and then hears that you have a named owner with a defined remit, you are in a materially better position than the organisation next door with a glossier slide and nobody’s name on it.

The named owner does not need to be a chief AI officer. In a 500-person business it usually should not be. It needs to be a person who, when something goes wrong, is the one who gets called. Write the name down. If you cannot write a name down, that is the single most important line in your board pack, and it is worth more than any ROI estimate you could construct.

Expect the board to be less informed than it sounds

This is the part nobody says out loud, so I will.

Dataiku’s poll of 900 CEOs found 62% report their boards are actively applying pressure to deliver measurable AI-driven outcomes. Separately, 61% of CEOs say their boards are rushing AI transformation, and Deloitte research puts 66% of boards at limited-to-no knowledge of AI.

Read those together. The pressure in the room is frequently not an informed judgement that you are behind. It is a non-executive who read something on a flight and does not want to be the board that missed it. That changes how you answer. You are not defending a programme against expert scrutiny. You are giving a group of intelligent people who do not know the subject a frame they can hold on to.

Which means: do not go in with a number you half-believe. A soft number in a board pack becomes a hard number in the minutes, and in six months you will be asked why you missed it.

The failure rate statistic, and how to use it properly

Someone will quote the 95% figure at you. Here is the better one.

RAND found 80.3% of enterprise AI projects fail to deliver promised business value, and the useful part is the breakdown: 33.8% abandoned before reaching production, 28.4% reaching production but failing to deliver the expected value, 18.1% running but never recovering the investment. Three different failures with three different causes. The first is usually scope or sponsorship. The second is usually process, because the tool was dropped on top of a workflow nobody had documented. The third is usually that the saved capacity went nowhere.

Knowing which of the three you are at risk of is a far more valuable board conversation than arguing about whether the number is 80 or 95. I have broken the failure modes down with the numbers attached if you want the longer version.

And there is a counterweight worth having in the pack. The same Grant Thornton survey found organisations with fully integrated AI are nearly four times more likely to report revenue growth than those still piloting, 58% against 15%. The gap between pilot and integrated is where the return lives. That is an argument for finishing something, not for starting three more things.

A note on the strategy document

One more number, because it will save you from a particular trap. A 2026 enterprise adoption survey found 75% of executives admit their company’s AI strategy is “more for show” than actual internal guidance, with 39% having no formal plan to drive revenue from AI tools.

If the board pressure results in you commissioning a strategy document, you have a three in four chance of producing one of those. The test is simple. Six weeks after it is signed off, does anyone open it to decide something? If not, you have written a board answer, not a strategy. A one-page readiness position with a named owner and three gaps beats forty pages nobody reads.

What to do in the week before the meeting

Four things, in order.

Count your spreadsheets. Ask how many manual exports and imports sit between your two most important systems. Every one of them is a place an AI tool will inherit bad data.

Give the same routine task to three people and compare how they approached it. If the three approaches differ materially, you do not have a process, you have three habits, and automating it will automate the inconsistency.

Name the owner. One person, written down, for AI outcomes and for AI risk. Who should own AI in a mid-size company goes into what the remit should cover.

Then get a scored position across all six pillars so you walk in with a baseline rather than an argument. The free AI readiness audit is 30 questions, 7 minutes, and gives you a score out of 120 with a band for each of data, process, people, technology, strategy and governance. It is self-reported, so it is your honest view rather than an external verdict, which is exactly what a board needs from you. Bring the score. Bring the two weakest bands. Bring the name.

That is a defensible position in a room where most people are bringing a guess.