MAS AI risk management guidelines: who Singapore firms should hire now

Use the toolkit as the job description. On 20 March 2026, MAS and industry published an AI Risk Management Toolkit under Project MindForge, developed by a consortium of 24 banks, insurers, capital markets firms and other partners. It covers traditional, generative and agentic AI, and a workgroup under BuildFin.ai continues the work.

Its Operationalisation Handbook is organised around four areas. Read each one as a staffing question rather than a policy chapter.

Two of these are harder than they look. The AI risk manager needs to understand credit, claims or trading well enough to judge what a failure would cost, which rules out a pure technologist. The validator needs to test agentic systems, where the question is what the agent did, not only what the model predicted. Insurers can borrow the role design behind agentic AI in insurance: the roles insurers need, because the accountability gaps are the same in Singapore as in the US.

Validators and machine learning (ML) engineers come from a different pool from risk managers. Sourcing them through analytics and AI recruitment makes sense. Sourcing your head of AI governance the same way usually does not.

Should the AI risk team sit in risk or technology?

In risk. The team that builds AI should not be the team that decides whether it is safe to use. Put AI risk inside the technology function and it becomes a quality assurance step the delivery team schedules around.

That does not mean staffing it only with risk generalists. The team works when a second line lead with real authority sits over a mix of people: someone from model risk, someone who has shipped a machine learning system, and someone from the business who knows how the product fails customers. A lot of Singapore firms will be tempted to bolt AI onto an existing model risk team. That can work if model risk already reviews machine learning, and it tends to fail if that team has only validated regression models and capital calculations. Candidates for these seats often come through risk and financial advisory recruitment, and the ones worth hiring will ask a sharp question at first interview: can I stop a launch?

Is AI replacing finance jobs in Singapore, and who can unwind an AI failure?

Some tasks are going. Judgement roles are a different story, because every AI control framework needs people who understand what the model is actually deciding. Singapore finance professionals are openly asking whether AI will replace them and whether a short AI course will carry them into a new career. The honest answer is that a bootcamp rarely turns a relationship manager into an ML engineer. It can help a credit analyst, a compliance officer or an operations lead move into AI risk, first line AI control or product ownership of an AI use case, where domain knowledge is the scarce part.

That scarcity is the real risk. If a lending model approves the wrong loans for three months, somebody has to know the credit policy well enough to work out which decisions to reverse. If the domain experts have been let go because the model seemed to handle it, nobody can unwind the failure. The same applies to insurers: the person who can unpick a pricing or reserving error is usually an actuary, and you should not assume you can rehire one quickly after a cut, whatever the market for actuarial jobs in Singapore looks like that year.

Keep a named domain owner for every material AI use case. Write that into the inventory now, before the guidelines make you.

How to make the AI risk manager search possible

A search for an AI risk manager is a problem statement before it is a vacancy. Before anyone sources a single CV, I want to know where the role sits, who it reports to and whether it can stop a launch, because strong candidates ask exactly that and walk away when the answer is vague. Then the practical checks: how big the pool really is, whether pay matches what competitors are offering, whether the working model fits, and whether the criteria are too narrow. Asking for years of agentic AI governance experience describes a person who does not exist yet. Position the role with real authority and a realistic profile, and the search becomes possible.