When the evidence is thin

Some AI risks can be tested with evidence. Is the model accurate? Is it biased? Is the data secure? A board can ask for the tests and challenge the results.

Others cannot, at least not yet. What would a supplier’s failure do to us? What might an agent do that nobody told it to? Will the bets we are making on AI pay off? These are questions no report can settle. They are judgements, made around the board table, under uncertainty. John Adams calls them virtual risks: knowledge is thin, and the experts disagree.

For risks like these, the quality of the board’s conversation is the control. And the person with most influence over that conversation is the chair.

A recent article by Rusty O’Kelley and Laura Mantoura of Russell Reynolds Associates, How Chairs Drive Board Performance, sets out what effective chairs do, drawing on the firm’s 2025 survey of board culture. None of it is about AI. All of it applies with more force when the subject is AI. Four of its points matter more on AI.

Expertise must not become a monopoly. In the survey, 89% said directors often or always apply their personal expertise to board issues, but only 71% said chairs draw out the expertise most relevant to the issue. On AI, boards tend to defer to the one director who “gets tech”, or to the chief technology officer. That expertise is invaluable on what the technology can do. It is no better placed than anyone else to judge what the firm should bet on, or how much uncertainty it can bear. The authors’ remedy is to keep expertise “at the right altitude”, with three questions: what assumption are we testing, what trade-offs does management need to consider, and what risk should the board understand? On AI, those are exactly the right questions.

Hold back your own view. The authors warn that “the chair’s authority can unintentionally anchor the discussion”, so chairs should frame the issue and ask questions before offering a conclusion. Under uncertainty this matters more. When nobody has the facts, the first confident view anchors the rest, and the chair’s anchors hardest. If the chair says early that the board is comfortable with its AI suppliers, few directors will be the one to say they are not. Only 68% of respondents said their chairs actively seek different points of view.

Put the bet on the agenda, not the progress report. The authors ask chairs to decide where a topic “would benefit from directors’ discussion rather than a management update”. AI usually arrives as an update on adoption: pilots launched, hours saved, use cases approved. That crowds out the question only the board can answer: what are we betting on, and what would we do if it didn’t pay?

Give feedback, and ask for it. Only 43% of respondents said their chairs often or always give constructive feedback. The authors want it given soon after the behaviour, not saved for the annual board evaluation, and they want chairs to hold themselves to the same standard. On AI this matters because the behaviours that weaken a board’s judgement are quiet ones: the director who defers on every AI item, the expert who closes the discussion, the board that waves through a progress report. Left until the annual review, a year of AI decisions has gone by. And someone has to be able to tell the chair when it is the chair who is anchoring the discussion.

The article stops there. I would add one more thing. Boards need to be able to hold on to uncertainty for longer than executive teams might like. Management has to act, and the pressure to reach a conclusion is understandable. But a chair who lets the board settle too quickly can turn a plausible assumption into an agreed fact.

That is not an argument for delay. On AI the evidence may never be conclusive, and waiting has its own costs. The board’s task is not to remove the uncertainty before it decides. It is to know, when it decides, what remains uncertain, what would show it was wrong, and when it will look again. The uncertainty should still be visible after the decision is taken.

In my work on risk appetite I have seen the same pattern many times: boards agree on risk without sharing a view of it. With AI the gap is wider, because the evidence is thinner and the expertise is unevenly spread. As Andrea Durrant put it in her note on the article, a board of excellent directors is not the same as an excellent board. On AI, the difference is made in the conversation, and the chair runs it.

So here is a question worth asking before the next board discussion of AI: