What is the single most important question a board can ask about AI?

What evidence would make us reverse this? A deployment with no stated reversal condition is not a decision, it is a commitment. The question forces the organisation to state in advance what would count as failure, which is the difference between design and drift: not whether you adopted AI, but whether you ever decided anything you could later be held to. Asking it before rollout is cheap; asking it afterwards is a crisis meeting.

Why is zero overrides a warning sign?

Because it is evidence of an untested right rather than a good system. If nobody has disregarded the system this quarter, either the right to refuse is not practically exercisable, or the workload makes real scrutiny impossible, or nobody in the chain is competent to detect an error. Article 14 of the EU AI Act, in force since 2 August 2026, requires that overseers of high-risk systems can decide not to use them or disregard their output, which is a capacity rather than a permission.

What should a board ask about AI?

Q: What should a board ask about AI?

Twelve questions that test capability rather than activity. What can we no longer do without these systems? Who could tell if this system were wrong? How are we keeping people good at the work we have automated? Does this arrangement beat the better of human alone or system alone? Are we measuring usage or capability? Whose evidence are we relying on, and what does it not show? Who is accountable by name for each stage? How many times has anyone overridden the system this quarter? Which decisions are we now approving rather than making? Where are our next senior people coming from? What is our AI literacy programme actually producing? And what evidence would make us reverse this?

Q: How should these questions be asked?

Not as an agenda item. Ask two or three in the ordinary course of reviewing something else, and listen for hesitation rather than for content. Questions asked as a set invite a prepared answer; questions asked in passing get you the truth. Where the answer is that nobody knows, that is a legitimate answer and it belongs in the minutes rather than being resolved on the spot.

Most board questions about AI produce a presentation. These twelve test whether the organisation is quietly losing the ability to operate without the systems it is adopting.

Boards are asking the wrong questions about AI, and getting reassuring answers to them. "What is our AI strategy", "how are we using it", "what is the productivity gain" all produce a presentation. None of them tests whether the organisation is quietly losing the ability to operate without the systems it is adopting.

These are twelve questions that do test it. Each is paired with the answer that should worry you, because the value of a board question lies entirely in knowing what a bad answer sounds like. Most of these can be answered honestly in a sentence. If they cannot, that is itself the finding.

How to use this

Not as an agenda item. Ask two or three, in the ordinary course of reviewing something else, and listen for hesitation rather than for content. Questions asked as a set invite a prepared answer; questions asked in passing get you the truth. Where an answer is "we do not know", that is a legitimate answer, and it should be written into the minutes rather than resolved on the spot.

Capability

1 · What can we no longer do without these systems?

Worrying answer: "Nothing, we could always go back to how we did it before." Nobody has checked. Capability does not disappear on a schedule anyone tracks, which is why capability debt is only visible when something fails or a novel question arrives.

2 · Who could tell if this system were wrong?

Worrying answer: a function rather than a person, or a person who could not have produced the work themselves. Fluency carries no signal: a plausible wrong answer looks exactly like a correct one to anyone who cannot independently evaluate the content. See who owns verification.

3 · How are we keeping people good at the work we have automated?

Worrying answer: "That is the point of automating it." Reasonable, until you need someone to supervise it, override it, or do it when it fails. Nobody budgets for maintaining capability in automated work, because it looks like paying people to do something a machine does faster.

Evidence

4 · Does this arrangement beat the better of the human alone or the system alone?

Worrying answer: "We have not measured that." A meta-analysis of 370 effect sizes across 106 experiments found human-AI combinations performing worse on average than the stronger party alone, with the losses concentrated in exactly the arrangement most organisations have installed. A pairing that has never been tested against that baseline may be subtracting.

5 · Are we measuring usage or capability?

Worrying answer: seat counts, licences, prompt volumes, adoption percentages. Those measure activity. Nothing in them says anyone got better at anything. See usage theatre.

6 · Whose evidence are we relying on, and what does it not show?

Worrying answer: a vendor case study, or a consultancy survey from a firm selling the remedy. Ask what the study does not support. Ask whether an average is concealing opposite effects on different people: in the radiology evidence, the effect of AI assistance ran from strongly positive to strongly negative between individual readers and was not predicted by experience.

Accountability

7 · Who is accountable, by name, for each stage of this work?

Worrying answer: the team, the function, or a name that appears only after something goes wrong. Accountability assigned retrospectively is attribution, and it behaves very differently under pressure. The Delegation Boundary Map makes the gaps visible in about ninety minutes.

8 · How many times has anyone overridden the system this quarter?

Worrying answer: zero. That is not evidence of a good system. It is evidence of an untested right, and possibly of a workload that makes real scrutiny impossible. Article 14 of the EU AI Act now requires that overseers of high-risk systems can decide not to use them or disregard their output, which is a capacity, not a permission.

9 · Which decisions are we now approving rather than making?

Worrying answer: a defensive one. This is the personal version of the question and the most uncomfortable, because it applies to the board as much as to anyone below it. Executives reading machine summaries instead of source material are making a different kind of decision than they think they are.

The pipeline

10 · Where are our next senior people coming from?

Worrying answer: "We will hire them." Everyone is planning to hire them, from a pool that is being drained by the same automation. The work that built senior judgement is the work most easily automated. See the missing rungs and synthetic seniority.

11 · What is our AI literacy programme actually producing?

Worrying answer: a completion rate. Completion is not capability. Article 4 of the EU AI Act has required a sufficient level of AI literacy since February 2025, at every risk tier, and enforcement began in August 2026. It asks about understanding relative to context and to the people affected, not about tool familiarity. See what AI literacy means for leaders.

The one that matters most

12 · What evidence would make us reverse this?

Worrying answer: silence, or "we would look at it if something went wrong." A deployment with no stated reversal condition is not a decision, it is a commitment. Asking this before rollout is cheap. Asking it afterwards is a crisis meeting.

This question does more work than the other eleven combined, because it forces the organisation to state in advance what would count as failure. That is the difference between design and drift: not whether you adopted AI, but whether you ever decided anything you could later be held to.

Where this is uncertain

These twelve are drawn from practitioner research and from the published evidence cited above, not from a study of board effectiveness. No research establishes that boards asking these questions get better outcomes than boards that do not. They are offered as a decision-forcing device, and the honest claim is that they surface things other questions do not, which is a weaker claim than it may sound.

The regulatory references describe requirements that are newly in force, with no guidance or case law yet. This is not legal advice.

Related SuperSkills research

On oversight duties, meaningful human oversight. On the practical framework, the Delegation Boundary Map. On the leadership response, how should leaders respond to AI and AI workforce strategy. On what the evidence does and does not establish, what we actually know.

Key sources

Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful. Nature Human Behaviour, 8.
Yu, F. et al. (2024). Heterogeneity and predictors of the effects of AI assistance on radiologists. Nature Medicine, 30(3).
Article 14, Human Oversight and Article 4, AI literacy, Regulation (EU) 2024/1689.
Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. Lancet Gastroenterology and Hepatology.

About this framework

Rahim Hirji is the author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Free to use and adapt with attribution. Findings are attributed to the studies that produced them and kept separate from the interpretation. Reviewed quarterly.

Cite this

Hirji, R. (2026). What should a board ask about AI? The SuperSkills Intelligence Company. Last reviewed 26 August 2026. thesuperskills.com/research/what-should-a-board-ask-about-ai