Quaintitative

For the second line

AI risk management for risk and compliance

You do not build the AI. You make sure the first line's AI is identified, rated and controlled.

There is plenty written for the board, which sets the direction, and some for the supervisor, who judges from the outside. There is far less for the person in the middle: the second line, which runs the risk-management system day to day, challenges a first line that knows the models far better, and answers to both the board above and the supervisor outside. It is the seat where the system actually runs. The board meets on it a few times a year and the supervisor visits once in a while; the second line is in it every day.

Under the MAS AI Risk Management Guidelines (AIRG), the second line gives the firm independent oversight of its AI. You hold independent oversight of how the first line identifies, inventories and rates its AI, and you are the final arbiter of what counts as AI and how material a use is. You provide independent validation and effective challenge of the models and systems in use, run the ongoing monitoring that catches drift and degradation, and own the vendor checkpoint so that bought AI is held to the same standard as built AI. This is your existing model risk, third-party risk and technology risk discipline, extended to AI rather than replaced.

Six questions I would put to any first line

I wrote the AIRG, and before that I led the 2024 thematic review of how banks actually manage AI model risk. Here is where I usually start with a second line: six questions, drawn from the AIRG but read from your seat. For each one, the question I tend to hear from the business, and how I think about it.

Consistency first, org chart second

"What operating model should the firm set up for AI risk: a central function, hub-and-spoke, or a capability in every unit?"

Weigh in on the operating model, but the first principle matters more: one approach has to govern AI risk consistently wherever AI runs. The AIRG requires consistency, not a particular shape. A clean org chart with three different standards inside it is worse than a messy one with a single standard. Focus on whether one approach governs AI risk the same way across the whole firm, not on how nice the org chart looks.

Own the net to catch AI, then check the low ratings

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

You cannot know for certain. AI arrives through features switched on in tools you already license, vendors adding AI to products they sold you years ago, and teams using public chatbots when the sanctioned route is slow. Focus on the net, not the list: wire identification into procurement and vendor onboarding, and reconcile against the registers you already keep. Then watch the rating, because a low rating is the cheapest way to make risk disappear. Focus on whether the firm has a good net to catch AI as it enters, and whether each low rating would survive a second look.

Make the inventory the command centre

"We keep an AI inventory and the business updates it every quarter. Isn't that enough?"

It depends what it does all day. Most AI inventories are a spreadsheet for auditors: accurate the week before an inspection, stale the rest of the year, read by nobody. The inventory the AIRG has in mind is a command centre: the ratings attach to it, the controls are determined by it, monitoring writes back to it, and a change to a system triggers re-review from it. Two questions tell you which kind you have. Does anything depend on it, and does anything write back to it?

Test to your task, not their benchmark

"The vendor's model scores near the top of the public benchmarks and the demo went well. Is that good enough?"

No. A benchmark is someone else's exam, and the questions may have leaked into the training data. A demo is the one path the vendor chose to show you. Evidence has to be based on your task, in your context: before anything ships, the first line should say what a good and a bad answer look like, set a threshold, and show the results against it on your data, including the cases where the system did badly. Focus on whether the first line can show you what good enough means for the task in your context.

Run a system, not a checklist

"We bought a control checklist that maps to every major framework. Aren't we covered?"

A checklist covers you about as well as a car manual prevents a crash. Controls do not work as a list, they work as a system: identification feeds the inventory, the inventory holds the ratings, the rating sets the intensity of every lifecycle control and how hard you challenge each use, and monitoring feeds back. The material systems get deep validation and the trivial ones a light touch, by policy rather than random judgement, which is how a small second line covers a large portfolio. Focus on whether the parts reinforce each other, not on how many rows the checklist has.

Decide where fast is safe

"The business ships AI faster than we can review. How do we keep up?"

Not by reviewing faster, but by deciding in advance where speed is safe. Two things help: expedited approval paths for genuinely minor, pre-agreed changes, so a routine retrain does not queue behind a high-stakes deployment; and clear boundaries for pilots, with limits on time, users and production data, so an experiment does not grow into an ungoverned system. Focus on whether the firm has deliberately decided where it can move fast, not on how to make every review faster.

These six are where I start. The fuller treatment, what the second line actually does, where risk hides from you, the three lines, and proportionality as the one tool that lets you cover a portfolio you could never validate in full, is in the book: a short primer, and the full AI Risk Management for Risk and Compliance.

Where to start

I developed the AIRG while leading AI risk supervision at MAS, and now advise second-line functions independently. I am publishing more on AI risk management; subscribe to get it, and future updates on the AIRG.

Subscribe for updates