If a board can't reconstruct a decision, it isn't overseeing it

Boards are being made accountable for AI they don't understand. It is happening quietly, and most directors haven't clocked the exposure yet.

The duty hasn't changed. A board is responsible for the decisions of the organisation: for oversight, for asking the question that should have been asked, for not signing off on what it didn't examine. That duty is old and well understood. What has changed is how the decisions reaching the board get made.

The AI arrived without a policy

Somewhere in your organisation, a decision is being shaped by an AI chatbot right now. The board will never know it happened. That's not a scandal. It's just the new normal, and it's worth thinking through before it matters.

A manager is weighing a real call: a supplier, a market, a restructuring. They open a chatbot, describe the situation, and get a fluent, confident, well-structured recommendation. It's genuinely helpful. It shapes their thinking. By the time the decision reaches anyone senior, it arrives as the manager's considered view, and it is, sincerely. The chatbot's role has quietly dissolved into it.

Multiply that across an organisation and you have decisions increasingly influenced by a tool that, asked the same question twice, might have answered differently, and nobody upstream can see it, question it or reproduce it.

"Should we allow AI?" is no longer the question. That decision already got made, informally, from the bottom up, by every manager with a login.

Better papers should worry you more

Sit on any board and you've felt the shift. The papers are sharper. The analysis is tighter. The options are better structured than they were three years ago. Management has quietly got more capable, almost overnight.

You know why. So does everyone in the room, though nobody says it. For most of what a board handles, this is pure upside: better papers, better decisions, less wasted time.

But sit with the consequence for the decisions that actually matter: the irreversible ones, the ones your name is on.

A board's protection is reconstruction

A board's core protection has always been reconstruction. When a decision goes wrong, you need to show how it was reached: what was considered, what was weighed, why this and not that. That reconstruction is how a board demonstrates it exercised oversight rather than just nodding things through.

You cannot reconstruct reasoning that happened inside an AI session nobody kept, from a process that, asked again, might have reasoned differently. The recommendation is excellent and its origins are unrepeatable. You're being asked to approve, under the same fiduciary duty as always, something you can no longer fully trace.

The uncomfortable question every board should be asking: if a decision this AI helped shape goes wrong, and we're asked how we exercised oversight, what do we point to?

"The model suggested it" is not an answer a director wants to give.

If the honest answer is "we'd be relying on the recommendation itself, because we can't see behind it", that's not oversight. That's trust. And trust has never been an acceptable substitute for oversight, whatever the quality of what you're trusting.

Run it twice

Here is a test worth running at your next board meeting. Take the AI-assisted analysis behind a big recommendation. Run it again. See if you get the same answer.

You often won't. Conversational AI generates fresh every time. Ask it the same question in two sessions and the reasoning shifts: a different emphasis, a softer caveat, sometimes a different conclusion. That variation is by design. In a conversation, it's a feature. In the analysis underneath a board decision, it's a problem hiding in plain sight.

It means the recommendation in front of you isn't a fixed position. It is one draw from a distribution: a confident, well-argued draw, but not the only one the same inputs could have produced. Approve it and you've approved a snapshot, presented as a settled view.

A board can live with that on small things. On the decisions that don't reverse (the acquisition, the market entry, the capital you can't recall) it's a real exposure. An analysis that changes when you re-run it fails at the first question a sharp regulator, litigator or fellow director would ask: would this have held if you'd run it again?

Reproducibility is a governance property

This is why I keep saying reproducibility is a governance property, not a technical one.

If you can't reproduce it, you can't govern it. You can only trust it, and trust is not oversight.

A verdict a board can stand behind has to give the same result from the same evidence, whoever asks and whenever. If it moves, you don't have a verdict. You have an opinion that happened to be generated on the day you asked, and you've attached your fiduciary duty to it.

What to insist on

The fix isn't to ban the tools or slow the papers down. None of this means keeping AI out of the analysis.

It means insisting that consequential decisions, the ones that reach your table and carry your accountability, run through something reconstructable: the same answer from the same evidence, whoever asks, whenever, regardless of what informal help fed into them along the way.

The narrower, sharper question for a board is this: for the decisions we're accountable for, can the reasoning still be reconstructed and defended?

Boards spent a decade learning to govern data and cyber risk. AI decision-making is the next thing they'll be held to, and most are earlier on that curve than they'd like to admit.

The better your management's AI-assisted recommendations get, the more this matters. Fluency was never the thing a board was there to provide. Scrutiny was.

One question to take into the next meeting: of the AI-shaped recommendations you've approved this year, how many would survive being run twice? If you don't know, that's the finding.

I write this as an IOD-certified independent director as well as an adviser. My board profile is here.