MAS Guidelines on Artificial Intelligence Risk Management

What it is

The Guidelines on Artificial Intelligence Risk Management set out the Monetary Authority of Singapore’s supervisory expectations for how financial institutions manage AI risk. MAS consulted on them from 13 November 2025 to 31 January 2026 and issued the final Guidelines on 7 October 2026, alongside a response to the feedback received.

They cover four areas. AI oversight by the board and senior management. Key systems, policies and procedures for identifying AI, keeping an AI inventory and assessing risk materiality. Controls across the AI lifecycle. And the capability and capacity an institution needs to use AI safely. They complement existing MAS rules, and MAS’s 2018 FEAT principles continue to apply.

Who it applies to

All financial institutions as defined in Singapore’s Financial Services and Markets Act 2022, in a proportionate way according to their size and risk profile. For locally incorporated institutions under consolidated supervision or owning critical information infrastructure, the Guidelines apply on a group basis. An institution that is part of a global group can use its group’s AI risk framework, provided it meets the Guidelines, and local senior management stays accountable.

They apply to all forms of AI, including machine learning, generative AI and AI agents, and to AI developed, operated or provided by third parties. Rule-based tools and robotic process automation that follow fixed instructions are not treated as AI.

Institutions whose AI use is unlikely to have a material adverse impact, such as AI that helps draft emails or summarise internal notes with a person checking the output, may instead apply basic AI governance policies. Those cover accountability, permitted and prohibited uses, an approved list of AI tools with a request process, staff education, regular checks and periodic review.

What it requires

Oversight. The board, or a committee it delegates, approves the governance approach for AI, makes sure AI risk is explicitly addressed in the risk appetite framework, sets roles and keeps enough understanding of AI to challenge it. Senior management implements and reviews the framework. A dedicated AI committee is not required.

Identification. Consistent processes to identify AI across the business, including AI embedded in services from material third-party providers. A designated control function owns the process and is the final arbiter on whether something counts as AI. Where AI cannot be fully identified, for example because of shadow AI or poor vendor disclosure, the institution should keep the residual risk within appetite through mitigating measures.

Inventory. An accurate inventory of AI use cases, systems or models, linked to other inventories such as data and third-party registers. Attributes may include purpose, approved scope of use, model type, data used, dependencies, lifecycle status, risk materiality rating, review status, owners and documentation. For AI agents MAS suggests agent identifiers, the tools and systems the agent can access, its components and its guardrails.

Risk materiality assessment. A consistent methodology that rates each use case on, at minimum, three dimensions. Impact covers the consequences of failure for the institution and its customers. Complexity covers the technology, its novelty, its data and its explainability. Reliance covers how much the decision depends on AI, including the autonomy granted and the degree of human oversight. Both inherent and residual risk are assessed, and residual risk must be within appetite before deployment.

Lifecycle controls. Controls proportionate to risk materiality covering data management, transparency and explainability, fairness, human oversight (including automation bias and decision fatigue), third-party AI, selection, evaluation and testing, technology and cybersecurity, reproducibility and auditability, pre-deployment reviews, post-deployment monitoring, change management and decommissioning. High-risk use cases need contingency plans.

Capability and capacity. Competent personnel and adequate technology infrastructure for the AI in use.

Key dates

• 13 November 2025. Consultation paper published.
• 31 January 2026. Consultation closed.
• 7 October 2026. Final Guidelines issued.
• 7 October 2027. Guidelines take effect. Sections 3 and 4 are expected from this date.
• 7 October 2028. Sections 5 and 6 are expected by this date.

MAS has also said it will consult industry in 2027 on what further guidance on agentic AI would help.

How it relates to other frameworks

The Guidelines refer institutions to IMDA’s Model AI Governance Framework, the AI Verify Foundation’s work and the Cyber Security Agency of Singapore’s guidance on securing AI systems, and to the Model AI Governance Framework for Agentic AI for agents. MAS’s FEAT principles remain the reference for what fair outcomes mean. The industry’s Project MindForge has published AI risk management handbooks, hosted on the MAS website, to help with implementation.

Outside Singapore, the structure is close to ISO/IEC 42001 and the NIST AI RMF, and to prudential expectations elsewhere, such as APRA’s in Australia. Groups that already run an AI inventory, a risk tiering method and lifecycle controls will find most of the structure familiar. The detail that tends to need work is the designated control function, the inherent and residual ratings, and the depth of third-party AI assurance.

A note on this page

This is Aicura’s reading of the Guidelines, written to help you understand them. It is not legal advice and it is not the Guidelines themselves. Read the primary source from MAS, and take advice on how it bears on your institution.