Scope: What This Guide Covers
On Oct. 7, the Monetary Authority of Singapore (MAS) issued its Guidelines on Artificial Intelligence Risk Management for financial institutions (FIs). These guidelines outline expectations for managing risks from AI use. FIs have until Oct. 7, 2027, to comply.
This guide is for compliance officers and risk leads who need to integrate these expectations into a functional compliance program. It details the structural requirements, how to build towards them, and common pitfalls.
Note: The guidelines are supervisory guidance, not a prescriptive control catalog. Treat any control-level detail here as a design choice until you've confirmed it against the published text. Where specific mechanisms are suggested, it will be noted.
Key Concepts and Definitions
Enterprise Level vs. Use-Case Level: The guidelines require managing AI risk at two levels. The enterprise level involves governance, oversight, and policy, while the use-case level addresses risks of specific deployments, like a credit-decision model or a customer-service chatbot. Both levels must be addressed.
Proportionality: FIs must tailor their approach based on their risk profile, including the scale and nature of AI use. A bank with a few low-impact tools and one embedding models in lending, fraud, and trading should have different programs. Proportionality requires justification of your program's depth.
Risk Profile: This is your assessment of AI usage, locations, and consequences. It anchors every tailoring decision, so document it thoroughly.
Four Key Elements: MAS organizes expectations around four key elements that FIs should incorporate. Confirm the exact names and scope in the published guidelines and map your program to them directly.
Requirements Breakdown
Consider the regulatory obligations in three tiers:
Tier 1: Governance and Accountability (Enterprise): Show who owns AI risk, how it's escalated, and how it connects to existing risk governance. Integrate AI into your current three-lines model.
Tier 2: Risk Identification and Assessment (Both Levels): Identify AI in your organization, including vendor AI and shadow use, and assess each use case consistently. This is your AI slice of the Risk Universe.
Tier 3: Lifecycle Controls (Use-Case): Controls should scale with the risk of the use case. Consider development standards, validation, monitoring, change management, and retirement. Depth depends on materiality.
The Oct. 7, 2027, deadline provides a timeline. Start early, as inventory and risk-rating work can take longer than expected.
Implementation Guidance
1. Build the AI Inventory First: Identify all AI use cases across business lines, procurement records, and vendor contracts. Capture purpose, data inputs, decision impact, owner, and whether the model is in-house or third-party.
2. Define a Materiality Rating: Create a simple, repeatable scale based on failure consequences: customer harm, financial loss, regulatory exposure, and human involvement. Tie control depth to the rating to operationalize proportionality.
3. Run a Policy Gap Analysis: Compare existing model risk, outsourcing, data governance, and information security policies against the four key elements. Extend rather than replace policies where possible.
4. Perform Control Objective Mapping: Map each AI expectation to an existing control objective and flag gaps. This keeps AI governance within your current framework.
5. Assign Ownership Across the Three Lines: First line owns the use case and controls. Second line sets standards and challenges ratings. Third line plans assurance coverage. Document ownership to avoid confusion.
6. Extend Third-Party Oversight: Include vendor-supplied AI in your Vendor Risk Profile and Ongoing Vendor Monitoring. Ensure contracts support transparency and testing access.
7. Track Deviations Formally: Maintain a Policy Exception Registry for AI use cases that can't meet standards, noting owner, compensating control, and expiry date.
Balancing Innovation and Control: Don't require every pilot to undergo a full review. Use a tiered intake: a light-touch path for low-materiality experiments and a full assessment for high-impact projects. This approach keeps AI work visible and manageable.
Common Pitfalls
- Treating AI as an IT or Data Science Issue: The guidelines cover FI-wide risk management. Compliance, risk, legal, and business owners all play roles.
- Inventory by Survey Only: Self-reporting misses embedded vendor features and informal tools. Cross-check with procurement and network data.
- One-Size Controls: Applying maximum rigor everywhere stalls delivery; minimum rigor ignores use-case specifics. Both approaches fail proportionality.
- Parallel Governance: A separate AI committee without links to existing risk forums creates duplicated reporting and unclear authority.
- Back-Loading the Work: Waiting until 2027 leaves no room to mature monitoring, which needs time to prove itself.
- No Evidence Trail: A control without evidence is treated as absent by supervisors. Decide now where assessments, approvals, and monitoring outputs will be stored; a GRC or IRM Platform can be configured for AI use cases.
Quick Reference Table
| Item | Detail |
|---|---|
| Issuer | Monetary Authority of Singapore (MAS) |
| Instrument | Guidelines on Artificial Intelligence Risk Management |
| Issue Date | Oct. 7 |
| Compliance Deadline | Oct. 7, 2027 |
| Applies to | Financial institutions |
| Risk Layers | Enterprise level and individual use-case level |
| Tailoring Basis | FI risk profile, including scale and nature of AI use |
| Structure | Four key elements to incorporate into operations (confirm names in published text) |
| First Action | Build AI inventory and materiality rating |
| Key Artifacts | AI inventory, materiality scale, Policy Gap Analysis, Control Objective Mapping, Policy Exception Registry |
| Third-Party Angle | Fold vendor AI into Vendor Risk Profile and Ongoing Vendor Monitoring |
| Biggest Risk | Late start and unclear ownership across the three lines |
Bookmark this guide and revisit it as your program matures. The deadline is distant enough to build well, but start your inventory now to stay on track.





