Your board does not understand the AI numbers you are showing them

Sit in enough board meetings and you notice the same pattern. An AI paper is presented. It contains a percentage — accuracy, precision, some benchmark — and a cost. Nobody asks what the percentage means, because asking would reveal that they do not know, and the paper is approved. This is not an education problem. It is a reporting problem, and the CFO owns it.

The fix is to stop reporting metrics a board cannot interrogate. A board can challenge cost per transaction, error rate in financial terms, cycle time, and headcount redeployed. It cannot meaningfully challenge an F1 score, and pretending otherwise creates the appearance of oversight without the substance. If a metric cannot be argued with, it should not be in a board pack.

The second omission is stopping conditions. Almost no AI paper I read states what result would cause the organisation to abandon the initiative. Capital projects have always been expected to state this. AI projects have somehow been exempted, with the result that failing pilots persist on the basis that the technology is improving, which is always true and never a reason.

There is an accountability question underneath. In most boards, AI oversight is nobody's specific responsibility — it is distributed across audit, risk and the technology committee, which means it is distributed to nowhere. Naming a single non-executive with the remit is a fifteen-minute decision that changes the quality of every subsequent discussion.

The uncomfortable framing for CFOs: if the board approved something on the basis of a number you presented that they could not evaluate, the governance weakness originated in your pack. That is fixable, and it is fixable by you, this quarter.