Singapore's AI Risk Guidelines: A Domain 2 Blueprint

On October 7, the Monetary Authority of Singapore (MAS) published its final Guidelines on Artificial Intelligence Risk Management, almost a year after consulting on a draft. They apply to every financial institution MAS supervises, from banks and insurers to payment firms. Even if you'll never deal with a Singapore regulator, the document is useful. It sets out an AI risk management process step by step: how to find AI, record it, rate it and decide how much control it needs. That sequence is the core of AAISM™ Domain 2.

What MAS Published

The guidelines cover all forms of AI, and MAS names generative AI and AI agents explicitly. They are organized into four areas: board and senior management oversight; systems, policies and procedures for AI risk management; controls across the AI life cycle; and staff capability and technology infrastructure. They take effect on October 7, 2027. The first two areas are expected from that date, and institutions have until October 7, 2028, to meet the life cycle and capability expectations.

Proportionality runs through the whole document. If poor performance or an outage of an institution's AI tools would be unlikely to cause material harm to the firm or its customers, a basic governance baseline is enough. MAS gives examples such as drafting customer emails or summarizing meeting notes, with a person checking the output. That baseline includes a named senior manager accountable for AI, rules on permitted and prohibited uses, an approved-tools list with a request process for new ones, staff training and periodic compliance checks. Anything beyond that triggers the full framework.

The Risk Process, Step by Step

• Identify AI, including AI you didn't buy as AI: identification has to reach AI built into services from material third-party providers, such as SaaS products that added AI features without marketing them that way. A designated control function makes the final call on whether something counts as AI. Where some use will inevitably go undetected, for example shadow AI, the institution must still keep the residual risk within its appetite, using measures such as data loss prevention and network monitoring.
• Inventory it with links to existing registers: the AI inventory should record purpose, approved scope, data used, dependencies, owners and risk rating. It should also share identifiers with the existing data inventory and the third-party register, so the three can be cross-referenced. For agents, MAS suggests also recording which tools and systems each agent can reach and what guardrails constrain it.
• Rate materiality on three dimensions: impact (harm to the firm and its customers, including the sensitivity of the data involved), complexity (how novel or explainable the technology is, and how much the vendor discloses) and reliance (how much autonomy the AI has and how much a human is involved). Each use case is rated both before and after controls, and the residual rating must sit within risk appetite before deployment.
• Calibrate controls to the rating: high-materiality use cases need formal independent validation before go-live, regular revalidation, tested contingency plans and, where relevant, a tested mechanism to switch the AI off quickly. Lower-rated cases can rely on peer review by someone outside the build team.

The Third-Party AI Line Worth Remembering

The guidelines leave no room to shift accountability onto the vendor. The decision to use third-party AI belongs to the institution, and so does primary responsibility for its effects on customers and on regulatory compliance. In practice, that means:

• Test the vendor's AI in your own context, using your own data, and run extra testing to cover anything the provider won't disclose.
• Prefer independent evidence: where vendor transparency is limited, look for certifications or assessments by competent independent parties, not the vendor's self-attestation.
• Contract for visibility: seek clauses requiring notice when AI is introduced or updated, or your consent before it is added, along with audit rights and performance commitments.
• Watch concentration: count how many material use cases depend on one AI provider, directly or through other vendors.
• Know your exit: if residual risk can't be brought within appetite, limit the service, suspend it or replace the provider.

None of this is unique to Singapore. The same themes run through the EU AI Act, the NIST AI RMF and new US state laws requiring third-party AI audits. MAS states them more concretely than most, though, and for a risk manager building a program, the identify, inventory, rate and calibrate sequence is a reasonable template even without a regulator requiring it. Expect exam questions in this area to test the same order of operations: you can't rate a risk you haven't found, and you can't set proportionate controls for an AI use case that was never rated.

Related Reading

For the full exam breakdown, see the AAISM™ Certification Study Guide. On AI risk arriving through a vendor, see The Vercel Breach: An AI Vendor Risk Case Study. On the AI use an inventory tends to miss, see Shadow AI Governance Gap. On independent assurance becoming a legal requirement, see Third-Party AI Audits Are Becoming Law. On another regulator turning AI controls into obligations, see California's New AI Laws Turn Controls Into Duties.

Sources

Monetary Authority of Singapore — Guidelines on Artificial Intelligence Risk Management (October 7, 2026)
Monetary Authority of Singapore — Consultation paper on the proposed guidelines (2025)
Baker McKenzie — Summary of the final guidelines and phased timeline