What should a chair ask the CEO about AI?

A chair has one instrument the rest of the board does not: the private conversation with the chief executive before anything reaches the room. What gets asked there determines what arrives in the board pack three months later. These are the questions that change what management does, rather than the ones that produce a better slide.

Why the published lists do not work#

There is no shortage of question lists. The Institute of Directors publishes one, the consultancies publish theirs, and most of them are good questions. The trouble is the format. A list of ten questions read out in a board meeting produces ten answers, a minute noting that AI was discussed, and no change in the organisation, because each question can be answered in the abstract and none of them costs anything to answer well.

The questions below are chosen on a different test: a chief executive cannot answer them without finding something out. Some of them will not be answerable at the first attempt, and the gap is the point of asking.

Five questions that require work to answer#

Which decisions do our machines already make without a person deciding? Not which systems exist, which decisions they determine: a credit limit, a refund, a shortlist, a price, a public reply. The first answer is usually a list of systems rather than a list of decisions, and the second attempt is where it gets interesting. A chief executive who comes back with a shorter list than the organisation actually has is not being evasive; nobody has ever compiled one.

Who would stop each of them, and when did they last do it? Ask for a name rather than a function, and ask for a date. The gap between a stop that exists in a document and one that has been rehearsed is the whole of the control. How to design a stop button people will use sets out why willingness, rather than mechanism, is the part that fails.

What could we no longer do by hand, and did we choose that? Every organisation is spending capability it has not counted. The question is not whether that is happening but whether it was decided. Capability debt.

How would I hear if one of them went wrong, and how long would it take? A chair should know the route and the elapsed time before there is anything to route. If the answer involves the phrase “it would depend”, there is no route. What counts as a serious AI incident.

Which assurance in the last board pack would you not want tested? This is the question that works. It is uncomfortable by design. It asks a chief executive to point at the weakest claim their own organisation has made, which is a thing a chair can only ask privately and a thing most chief executives can answer immediately.

Three questions to ask yourself rather than the chief executive#

Whether this board could tell a good answer from a fluent one. A board with no way of judging the reply is not exercising oversight, it is receiving comfort. What AI literacy means for leaders.

Whether AI arrives in this board's papers as a technology item or as a decision item. The agenda position tells you which conversation the board is having, and technology items get delegated to the person who understands them.

Whether the instinct to appoint a director with AI expertise is oversight or outsourcing. It is the most common reflex and frequently makes the rest of the board less engaged rather than more. Should we appoint a director with AI expertise.

What a chair does with the answers#

Three of the five questions produce a list, and a list is a thing a board can hold management to over quarters rather than discuss once. The practical move is to ask for the lists privately, let the chief executive produce them without an audience, and then put them in the pack as standing items so the board watches them move. What belongs in that pack is set out at what management should report to the board about AI.

The failure mode is a single excellent discussion followed by nothing, which is what most boards have already had.

What this does not show#

These are questions argued from what has worked in rooms and from what the oversight evidence suggests goes wrong. No regulator requires any of them, no research has compared boards that ask them against boards that do not, and a chair who asked a different five would not be wrong. The UK Corporate Governance Code does not mention AI and the FRC has published no AI guidance for boards, so nobody is in a position to tell a chair that a particular question is mandatory.