What does a non-executive director need to know about AI?

A non-executive director is in an odd position on AI: accountable for the oversight of systems they do not use, cannot see, and did not buy, on a few days a year, with information supplied by the people they are overseeing. The reflex is to try to understand the technology. The better move is to get good at a small number of questions and at reading the answers. This page is for non-executives, non-execs, independent directors, trustees, governors and board observers, whose formal duties differ and whose practical problem is identical.

The best reference for this is seven years old#

Search for what a non-executive director should know about AI and the most substantial result is a KPMG paper from 2019. It was written before large language models reached general use, before agents, before any of the current regime. It sits near the top because almost nobody has replaced it. A non-executive doing their own preparation is reading advice from a different technological era, which is worth knowing before you act on it.

Four things that are genuinely different about AI oversight#

The system acts, so there is no pause to oversee in. Every oversight model a board has learned assumes a gap between proposal and action in which a person can intervene. Agents remove the gap. Can a human approve an AI decision at machine speed.

The control everyone names is usually not one. Human in the loop appears in almost every AI policy and describes an arrangement rather than a control. Asking what the person in the loop is able to do, and whether they have ever done it, separates the two. Human in the loop is not a safeguard.

The loss is invisible until it is expensive. Capability the organisation is spending does not appear in any report, and the bill arrives when a decision turns up that the system cannot make and nobody in the room has been trained to make either. Capability debt.

Confidence is a property of the writing, not of the knowledge. A fluent wrong answer reads like a fluent right one, so a non-executive reading an AI-assisted board paper cannot use their usual instinct for when something is off. How do I know when AI is wrong.

What you do not need to know#

How the models work. Which vendor is ahead this quarter. The difference between the architectures. Whether the company should use one model or another. These are the things a non-executive most often tries to learn first, they take real effort, and none of them is a board matter. A director who has learned the technology and not the questions is in a worse position than one who has done the reverse, because they now feel qualified to accept an answer.

You also do not need to be the board's AI expert. The instinct to appoint one is common and frequently reduces the rest of the board's engagement rather than raising it. Should we appoint a director with AI expertise.

What to ask in your first meeting on this#

Three questions, and the state of the answer tells you more than its content. Which decisions do our machines make without a person deciding, and who approved that. Who can stop each one, and when was that last rehearsed. How would this board hear if one went wrong. If the answers exist and are written down, the organisation is further along than most. If they are produced verbally and differently by two executives, that is the finding.

The fuller set, and what a board should be shown each meeting, is at what a board should ask about AI and what management should report.

The question nobody asks out loud#

Whether directors should be putting board papers into an assistant. Many already are, using their own accounts, to summarise a two-hundred-page pack on a train. It is rarely raised because raising it implies an admission. It has an answer, and the answer is not a straightforward prohibition. Should directors put board papers into AI.

Trustees, governors and board observers#

The duties differ and the exposure differs. A charity trustee answers to the Charity Commission and to the charity's own governing document, an academy trust board answers to the Department for Education and its funding agreement, a school governor sits in a maintained school's instrument of government, and a board observer usually has no vote and no duty at all, which changes what they can do and not what they can see.

None of these instruments mentions AI, and neither does the UK Corporate Governance Code. What each of these roles shares with a listed-company non-executive is the practical problem: oversight of systems you do not operate, on limited time, using information prepared by the people being overseen. The four differences above and the three questions apply unchanged.

What this does not show#

Nothing here is a legal statement of directors' duties and it is not a substitute for advice on them. The claim that oversight arrangements commonly fail rests on evidence from aviation, medicine and automation research rather than from studies of boards, because studies of boards overseeing AI do not yet exist. The strongest version of the opposing case is that boards have absorbed unfamiliar technologies before without any of this apparatus and may do so again.