Five questions a board should ask before approving an AI model

Most boards I advise feel under-equipped to oversee AI. They are not. Oversight does not require knowing how a model works. It requires knowing what it decides, who owns it, and what happens when it is wrong. Before approving any model that touches customers, money or regulatory obligations, I suggest five questions.

What decision does it make or influence?

Be specific. “Supports underwriting” is not an answer. “Sets the price offered to small-business customers, within a band agreed by the pricing committee” is.

Who is accountable when it is wrong?

A named executive, not a team. If accountability sits with the vendor, it sits nowhere.

How will we know it has drifted?

Models degrade quietly as customers and markets change. Ask what is monitored, how often, and who sees the results before the board does.

If nobody can answer the fallback question, the model is not ready.

What is the fallback?

If the model is switched off tomorrow, what happens to the decisions it was making? A manual process nobody has tested is not a fallback.

Could we explain a decision to a customer or a supervisor?

Not the mathematics, the reasoning. If the answer is no, the institution is carrying a risk it cannot describe.

None of these questions are technical. All of them are uncomfortable when the answer is vague, which is exactly why they work.