Questions to ask your first line about AI

These are the five questions I would put to a first line about its AI - the oversight duties from the MAS AI risk management guidelines (AIRG), read as plain questions the second line puts to the business, and to itself. I wrote the AIRG and led the thematic review of how banks actually manage AI model risk, so these are the ones that separate a real programme from a theatre of it. For each, the common question I hear from risk and compliance teams, and the one I would put instead.

The areas are the MAS'. The interpretations here are wholly mine, so not regulatory guidance. Blame me if you disagree.

Enforce one approach, do not design the org chart

What operating model should we set up - a central AI function, hub-and-spoke, or a capability in every unit?

Risk and compliance teams spend a lot of energy on the shape. I would resist. A shape that fits one firm fails the next, and the structure is the business's call, not yours. The AIRG says the same: consistency is required, the shape is not. Your job is not to design the box on the org chart. It is to make sure one approach governs AI risk the same way wherever AI runs, because you are the function that has to hold that line.

So the question I would put is whether one approach governs AI risk consistently across the whole firm - and whether you can actually enforce it, since you are the control function that arbitrates what counts as AI and how material it is.

Own the inventory, then go to the low-rated

How do we know we have found all of our AI?

You cannot prove completeness, and chasing it will exhaust you. But the inventory and the materiality methodology are yours to own, not the business's. So build the net, and then probe your own ratings. The cheapest way a first line makes a lot of risk disappear is to rate it low, and a low rating quietly tells every control downstream to try less. So the low ratings are the ones to pull on, not the high ones.

So the question that tells you most is whether you know where your AI is and have rated each use honestly - and whether the low ratings your first line gave you survive a second look.

Challenge, without owning the outcome

Should risk and compliance own AI risk, and sign off that it is acceptable?

Being handed the whole thing feels like authority. It is a trap. The business owns the risk it takes; your job is to challenge and validate, not to mark its homework and not to mark your own. The financial sector settled this long ago with the three lines of defence, and AI slots straight into it. If you both run the control and sign off that the residual risk is fine, no one independent has actually said yes.

So the question I would put is whether someone in the business is accountable for each AI outcome, and whether you can challenge it independently, across the three lines - not whether AI risk has been dumped on your desk to own end to end.

Demand evidence, not a demo

Do we need to understand the models to challenge them?

Your job is to challenge, not to build. You already challenge credit and market-risk models without having built one, because the validation, the limits and the reporting are laid out in a form you can interrogate. AI is the same. The failure is rarely that you cannot understand it. It is being handed a polished demo instead of a validation report, test results including the failures, and a clear statement of what "good enough" means for the task.

So the question that matters is whether your first line can show you evidence you can interrogate, including what good enough means for the task and the test that proves it, not whether your team can read the model's weights.

Put your effort where the risk is

How does a small second line review everything the business is shipping?

You cannot, and you should stop trying. Spreading yourself evenly is not diligence, it is thin everywhere, so thin where it matters. Match your effort to materiality, the same way you expect the first line to match its controls to it. The material systems get deep challenge and independent validation; the trivial ones get a light touch, or a check that they were rated trivial for good reason.

So the question that lasts is whether you are spending your limited second-line effort where the risk actually is.

Work with me

I train risk and compliance teams on turning AI risk management into a working system. See the courses and workshops, read more on AI risk management, or get in touch.