Board Meetings
What Should Boards Ask AI Before Making Strategic Decisions?
The strategy team walks in with a recommendation, confidently phrased, backed by a scenario model, three supporting charts and a summary an AI system produced in minutes. Nobody on the board built it, nobody fully understands how it arrived at its conclusion, and the meeting is already running long. The temptation is to nod it through, and by the time anyone thinks better of it, the item has moved on to the next page of the pack.
Accepting a good-looking answer is not the same as asking a good question, and boards that have built the habit of continuous monitoring, the discipline covered in an earlier piece on continuous strategic adaptation, still need a separate skill: knowing what to ask before a recommendation becomes a decision. The quality of a board's AI-assisted decisions will depend less on the sophistication of the tools management uses and more on the sharpness of the questions directors put to whatever those tools produce, because the tools keep improving regardless of who is using them well.
This article sets out four questions worth asking before any AI-assisted recommendation reaches a vote, plus what it takes to embed that habit into how a board runs its meetings, and a short way to check whether the habit has taken hold.
In this article:
Why AI does not remove board accountability
What is the board's role in AI-assisted decision-making? Unchanged from before AI arrived: making the judgement call and living with the consequences, regardless of how much analysis, human or machine-generated, sat underneath it. AI can produce the recommendation. It cannot be the one accountable when the recommendation turns out to be wrong.
Regulators are making that distinction explicit rather than leaving it to assumption. Under Provision 29 of the UK Corporate Governance Code, effective for financial years beginning on or after January 1, 2026, boards must go beyond listing the controls they have in place and demonstrate a genuine understanding of why those controls might fail. Applied to AI-assisted decisions, that's a direct challenge to boards that treat a recommendation's polish as evidence of its soundness. A tidy summary and a working control are two different things, and a board that can't explain why it trusted a particular AI-assisted recommendation isn't meeting that standard.
Boards that have spent the last few years building comparable discipline around cyber risk and data governance already have a template for what this looks like in practice. AI oversight mostly applies the same rigour to a newer category of decision, not a different kind of scrutiny altogether.
Governing AI-assisted decisions well is a governance question that happens to involve AI, not a technology question at all: who is accountable, what was tested before the board relied on it, and whether the board's own scrutiny would survive being examined months later. The four questions below are what that scrutiny looks like in practice.
A practical questioning framework for directors
What questions should boards ask AI? Four, asked consistently, cover most of what matters: what assumptions it's making, what evidence is missing, what alternatives were never considered, and where it could simply be wrong.
What assumptions is AI making?
Every AI-generated recommendation rests on assumptions nobody stated out loud: that a market keeps behaving the way it has, that a supplier relationship holds, that a competitor doesn't move first. Surfacing them is the first job, because an assumption nobody named is an assumption nobody can challenge.
A concrete example: an AI-assisted growth forecast that silently assumes current customer acquisition costs stay flat. That assumption might be reasonable. It might also be the single most fragile part of the entire recommendation, and the board has no way of knowing which without asking. The question isn't whether the assumption is stated somewhere in the supporting material. It's whether a director can name it without digging for it, and whether they would even recognise it as an assumption, not simply as a fact.
What evidence is missing?
Completeness is easy to fake and hard to verify from the outside: a recommendation can look complete while leaving out, without anyone noticing, the one category of evidence that would have changed the conclusion.
One sharp version of this question is whether cyber and business risk were weighed alongside the commercial case, or whether they sat in a separate report nobody cross-referenced against the recommendation itself. A growth strategy that never asks what happens if the underlying data infrastructure is compromised has a gap, not a flaw in its logic, a gap in what it considered at all. Directors probing for this ask what wasn't included, not only whether what was included looks credible. A board that has never asked to see the risk register alongside the recommendation is not in a position to know whether that cross-check happened at all.
What alternative scenarios should be considered?
Felipe Csaszar, a strategy professor at the University of Michigan's Ross School of Business, made a sharp point in Harvard Business Review about how few real alternatives most strategy sessions generate: teams walk away with three or four options not because that's all that existed, but because “that's all your team had the time and mental bandwidth to develop.” AI removes that particular constraint. It doesn't remove the need to ask whether the alternatives on the table are the right ones, or just the ones easiest to generate.
We've covered how AI expands the range of scenarios a board can test in more depth in an earlier piece on AI-assisted scenario modelling. The question that matters here is narrower and harder to dodge: did anyone deliberately try to break the recommendation with a scenario built to challenge it, or did every version confirm the same conclusion, dressed up in slightly different numbers?
Where could AI be wrong?
Directors identify AI blind spots by asking where a model's confidence and its accuracy might diverge, since the two aren't the same thing, and a fluent, well-formatted answer can be confidently wrong.
The practical version of this question: what would have to be true for this recommendation to fail, and how would the board know if it had? One that can't answer that, that has no stated condition under which it would be considered wrong, hasn't really been tested. It has just been produced. Asking where AI could be wrong isn't adversarial for its own sake, and boards that treat it as such usually stop asking after the first uncomfortable answer. It's the difference between a recommendation the board interrogated and one it merely received.
Embedding challenge into board discussions
None of the four questions above do much good if they only occur to someone after the meeting has already moved on. Embedding them into how a board runs takes a few specific habits, and none of them require slowing the board down as much as directors sometimes fear.
Frame agenda items as questions, not recommendations. “Should we approve this AI-assisted growth plan?” invites a nod. “Which of this recommendation's assumptions are we least confident in?” invites a real discussion.
Circulate AI-derived analysis before the meeting, not during it. Directors challenge more effectively when they've had time to sit with the material and form a view, instead of absorbing it live in the room.
Expect individual review before collective discussion. A director who has already identified which assumption worries them arrives ready to press the point, instead of waiting to see what others think first.
Chairs should actively ensure debate against alternative views. A recommendation that goes unchallenged for want of anyone assigned to challenge it is not a tested recommendation, whatever the vote count suggests. That's a role for the Chair to hold deliberately. As Simon Laffin puts it on the Boardroom Confidential podcast: “There is an argument that if everybody agrees, the chair should stop and say, ‘We’re not going to agree this until somebody disagrees.’
Confirm shared understanding before concluding. Before a decision closes, the Chair should be able to state, and have the room agree, what was decided and why, beyond the fact that a vote happened.
The test of good AI-assisted judgement
Three questions reveal whether this discipline is real or just theoretical. Could a director explain, without notes, the single assumption a recent AI-assisted recommendation depended on most? Has anyone on the board ever changed a recommendation by asking one of these four questions, or does asking them always confirm what was already decided? And would the board's reasoning survive being read back eighteen months later by someone who wasn't in the room? A board that answers no to any of these is receiving recommendations, not testing them.
How Sherpany supports better board deliberation
Better deliberation depends on the questions above leaving a trace, not evaporating once the meeting ends and living only in whoever remembers asking them.
Sherpany gives boards a secure, structured place to capture that record: which assumptions were challenged, what evidence a director flagged as missing, and why the board ultimately trusted, or didn't trust, a particular AI-assisted recommendation. That audit trail matters twice over. It makes challenge visible in the moment, since a question raised and logged carries more weight than one raised and forgotten. And it makes the board's reasoning defensible later, when a regulator, an auditor or a new director asks why a decision was made the way it was.
None of this decides anything on the board's behalf. It gives the board's own judgement, and the questions that shaped it, somewhere to live beyond the meeting where they were asked.
A Good Answer Is Not the Same as a Tested One
A good-looking recommendation and a tested one aren't interchangeable, and the gap between them is exactly what these four questions are built to close. AI can produce the first. Only the board, asking what it assumed, what it missed, what else it should have considered, and where it could be wrong, produces the second.
Asking these questions well tends to speed decisions up rather than slow them down, because a bad assumption gets caught in the meeting instead of eighteen months later, when it's expensive and public. The board that gets this right isn't the one with the most sophisticated AI tools. It's the one that never stopped asking what those tools might have missed, meeting after meeting, long after the novelty of having AI in the room has worn off.
If you would like to see how Sherpany supports better-documented, better-challenged board decisions, book a free demo and find out how Sherpany can help.