First published on LinkedIn, 12 September 2026. Statistics updated November 2026.
In short
- Directors have adopted AI faster than their boards have governed it: 82 per cent have used it in their board work, and only 6 per cent sit on a board with a formal policy for it.
- AI proposals are capital allocation decisions being overseen with the habits of a technology decision.
- Three answers a board should have before approving: the business number and its owner, what stops, and who answers when the model is wrong.
The paper ran to fourteen pages. Twelve of them described the technology. Two described what it would do for customers, and both of those were written in the future tense.
Nobody in the room asked a bad question. The board asked about cost, timeline, vendor lock-in, and where the data would sit. All reasonable. All the sort of thing a competent board asks. What nobody asked was what the business would be able to do afterwards that it could not do before, stated plainly enough that someone could be held to it.
The paper was approved.
I have sat through enough of these to recognise the shape of it. This is not incompetence, and it is not negligence. It is a board applying the oversight habits of a technology decision to what is in fact a capital allocation decision, and the two call for different questions.
The gap sitting inside the numbers
Directors are not ignoring AI. In What Directors Think 2026, 38 per cent of directors named deploying AI across the business as a leading priority, and technology adoption is the top area for capital investment, cited by 42 per cent. Research by the Diligent Institute and Corporate Board Member, published in July 2026, found that 82 per cent of directors have used AI in their own board work, up from 66 per cent less than a year earlier.
The same research found that only 6 per cent have a formal board-specific policy on how directors use it, and 54 per cent have no policy at all.
Part of the reason is fluency. A board that isn’t confident in a domain asks about the parts of the paper it can assess, which is why AI discussions gravitate to cost, vendor, and data protection. Those are real questions, but they aren’t the governing ones. The governing question is what the business can do afterwards. Anyone can ask it, which is why it’s the board’s question rather than the technology function’s.
Read those figures next to each other. More than four in five directors have adopted the technology personally. Fewer than one in ten sit on a board that has decided how they may use it. Adoption ran ahead of governance in the one room whose job is governance.
The picture on results is no more comfortable. In the NACD’s 2026 Governance Outlook, 76 per cent of directors said AI will factor into their growth strategy, yet most organisations reported only slight or moderate success in realising operational or financial benefits from it so far. That is a fair description of where the technology is. It also describes a period in which a great deal of capital gets committed on the strength of direction rather than results.
Momentum is not a business case.
The question I have found most useful in these discussions is deliberately blunt. Is this serving the customer and the business, or is it serving the momentum?
Momentum is not always hype. More often it is inherited. It is the strategy the company has run long enough that nobody remembers deciding on it.
I worked with an AI business whose product was strong and whose market definition was not. The company had been founded to serve one narrow vertical, and every board discussion was a variation on how to win more of that vertical. The technology roadmap followed the market definition without testing whether it still held.
The question that moved the discussion wasn’t about the technology. What does this capability do that has nothing to do with the market we chose at the start?
The answer was cross-vertical automation and optimisation. Financial services was the first market to open, but it wasn’t the market the company had been built around. Repositioning around that capability moved the business from competing for share of one vertical to selling a capability into several. The addressable market didn’t enlarge; it was redefined. The product barely changed. The definition of what the company was for changed completely.
That is what the momentum question is for. It does not ask whether the technology works. It asks whether the thing the technology is being pointed at is still the right thing.
The seam where these decisions fail
I have been on the wrong side of the same gap myself, in a different context.
Leading a post-acquisition integration for a multinational consumer and industrial goods group, I had board approval, a plan and a clear mandate. What I didn’t have was dedicated resourcing, and I didn’t push hard enough when the board approved the outcome without deciding what we’d give up to get there. The integration stretched to two years. It should have taken about half that time, and the extra year included two field sales forces where one was needed, a duplicated management layer above them, and a second set of systems underneath. Reconstructed at today’s costs, that is somewhere around five and a half million pounds of avoidable spend. It is a reconstruction rather than a figure anyone produced at the time, and that is the part worth noticing. Since no one was asked to own the number, no one produced one.
I have run every integration since on the lesson: a board that approves an outcome without resourcing the route to it has not made a decision; it has expressed a preference.
AI proposals fail in that same seam far more often than they fail technically. The model works. The pilot works. Then the capacity it was supposed to release is never reallocated, the process it was supposed to replace runs alongside it for another eighteen months, and the business case quietly becomes a cost.
Three answers a board should have before it approves
Not three questions to ask the CTO. Three answers the board itself should be able to give, in its own words, before the paper is approved.
- What changes in the business if this works, stated as a number, and which executive owns that number? General efficiency is not an answer. A percentage against a named line, owned by a named person, is.
- What does the organisation stop doing? Capacity released and never reallocated is a saving on paper and a cost in practice. If nothing stops, nothing has been saved.
- What is the exposure when the model is wrong, and who answers for it? Not who maintains the system. Who answers for the decision it made.
If those three answers don’t exist, the decision isn’t ready, no matter what the technology is worth.
None of the three requires the board to understand the model. They require the board to insist that business consequences are considered with the same rigour as technology, and to notice when they aren’t.
A board that cannot say what an AI investment is supposed to change is not overseeing that investment; it is watching it.
And if most of your directors are already using AI in their board work while the board has never written down how they may use it, that is the first paper to commission, not the last.
Terry Irwin is a Non-Executive Director and Board Adviser in the UK and Europe, and helps boards test AI claims, spend and risk. More on his work: AI at board level.