AI risk management for risk and compliance

If you sit in risk or compliance, AI risk management is landing on your desk whether or not you built the AI. The good news is that it is an extension of disciplines you already run, not a new one. If your firm already does model risk management, third-party risk, and technology risk, you have most of the muscle. AI stretches each of them rather than replacing them.

Start where you already have muscle

Model risk management already treats a model as a source of risk. The difference with AI is that a learned model's behaviour lives in data rather than in written rules, so validation shifts from checking the logic to probing the behaviour. Third-party risk already governs vendors. The difference is that most of the AI your firm uses it did not build, so third-party AI is now a large part of the surface. Technology and operational risk already watch live systems. The difference is that generative and agentic systems act, not just predict. Extend, do not start over.

Materiality is your throttle

Do not treat every AI use the same. Rate each on impact, complexity and reliance, and put the heavier controls where the risk is highest. The low ratings are the ones to scrutinise, not the high ones, because a low rating quietly tells every downstream control to try less. Materiality can also change without a line of code changing, so it is not a one-time stamp.

What to ask the first line

You do not need to read code to challenge an AI system. You need the right questions and the evidence behind the answers.

• Where is AI used, and is it in the inventory? If it is not written down, it is not governed.
• How material is this use, and who signed off on that rating?
• Who is accountable for the outcome when it is wrong?
• What does "good enough" mean for this task, and what test shows it?
• Is it monitored after launch, and what happens when a quiet change lands, a retrain, a prompt tweak, a vendor update?

Evidence, not vibes

A control is only real if it is clear and has an owner, a threshold that triggers action, and a test that runs. A principle on a wall is not a control. Your single most useful asset is a live inventory: it is the one place where everything about an AI system meets, and it turns "what does this change touch, and what has to be re-checked?" into an afternoon's work instead of an archaeological dig. For how the controls work as one system, see the AIRG in practice.

Work with me

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