How to prepare for the MAS AI Risk Management Guidelines

The MAS Guidelines on Artificial Intelligence Risk Management give Singapore financial institutions a year to get the foundations in place and two years to finish the lifecycle work. They were issued on 7 October 2026 and take effect on 7 October 2027. Sections 3 and 4, on oversight, identification, the inventory and risk materiality, apply from that date. Sections 5 and 6, on lifecycle controls and capability, are due by 7 October 2028. This guide sets out a preparation plan in that order. The detail of the Guidelines themselves is on the framework page.

Decide which track you are on

The Guidelines apply to all financial institutions, proportionately to size and risk profile. Paragraph 2.3 lets an institution apply basic AI governance policies and procedures instead of the full framework, where poor performance or unavailability of its AI is unlikely to have a material adverse impact on the institution, its customers or other stakeholders. MAS gives examples such as AI that helps draft or proofread emails to customers, or summarise documents and meeting notes for internal use.

Make this decision first and write down the reasoning, because it sets the scope of everything that follows. An institution that relies on the basic track still needs clear accountability, permitted and prohibited uses, an approved list of AI tools with a way to request new ones, staff education and regular checks. It also needs to revisit the decision as its AI use grows. One customer-facing model can move it onto the full track.

Name the owners before you build anything

Sections 3 and 4 assign roles before they describe processes. The board approves the governance approach, makes sure AI risk is addressed in the risk appetite framework and keeps enough understanding of AI to challenge it. Senior management implements and reviews the framework, sets an escalation process for AI incidents and breaches of risk thresholds, and keeps the board informed. For institutions using a group framework, local senior management stays accountable for Singapore and must be able to show MAS how.

The role most institutions do not yet have is the designated control function. The Guidelines ask for one to own AI identification and the AI inventory, and a control function to own the risk materiality methodology. Business units can do the work, but the control function keeps independent oversight and decides whether a given use counts as AI. Pick it early, because it will own the next three steps.

Find the AI, including what vendors embed

Identification is the step MAS calls a critical prerequisite. It should reach every business and functional area and, at a minimum, AI incorporated in services from material third-party providers. MAS’s own example is software-as-a-service that is not sold as AI but contains AI features.

A survey of project teams will not find that. Identification needs a standing feed from the places AI arrives, such as device management tools, cloud inventories, code repositories, IT asset registers and vendor due diligence, so new AI shows up as a finding the control function reviews. Where AI cannot be identified, because of shadow AI or poor vendor disclosure, the Guidelines ask for mitigating measures so the residual risk stays within appetite. The shadow AI guide covers where to look.

Build the inventory as a working record

The inventory should be accurate, linked to other inventories such as data and third-party registers, and updated at a set frequency as use cases are added, changed or decommissioned. MAS lists attributes an institution may capture, namely purpose and description, approved scope of use, model type, data used, dependencies, lifecycle status, risk materiality rating, review status, owners and developers, and links to documentation. For AI agents it suggests agent identifiers, the tools and systems the agent can access, its components and its guardrails. The AI system register guide covers how to structure one.

Rate every use case twice

The risk materiality methodology must, at a minimum, consider three dimensions. Impact is the consequence of failure for the institution and its customers, including fairness and consumer protection, and the sensitivity of the data. Complexity covers the technology, the novelty of its use, its data and how explainable its outputs are, and for third-party AI how much visibility the institution has. Reliance is how much the decision depends on AI, including the autonomy granted and the degree of human oversight.

Each use case is rated for inherent risk before controls and residual risk after them, and residual risk must be within appetite before deployment. Calibrate the method on a handful of real use cases before rolling it out, so the ratings separate a drafting assistant from a credit model.

Use the second year for lifecycle controls

By 7 October 2028, controls proportionate to each rating need to cover data management, transparency and explainability, fairness, human oversight, third-party AI, evaluation and testing, cybersecurity, reproducibility and auditability, pre-deployment review, monitoring, change management and decommissioning. High-risk use cases need contingency plans with fallback options such as alternative systems or manual processes, tested regularly. Where an AI system has a kill switch, the activation protocol should be tested too.

Third-party AI gets particular attention. The institution keeps primary accountability, should get visibility of AI changes through its contracts, test vendor AI in its own context and consider concentration risk. The vendor AI risk guide covers contract terms and due diligence.

Where Aicura fits

Aicura’s AI Register holds every AI system and agent with its owners, purpose, model, data, vendor, status and version history, and Discovery surfaces AI in use that is not yet on it, including AI features inside software you already run. Risk assessments start when a system is registered and run again when it changes, and at Pro impact assessments, vendor records, security testing and monitoring sit on the same records. For agents, the register records what each agent may do and its kill switch. Aicura gives guidance and supports the work. It does not assess against the MAS Guidelines, and the judgement on your obligations stays with you.

A note on this page

This is Aicura’s reading of the MAS Guidelines, not legal or regulatory advice. Read the Guidelines on the MAS website, and take advice on how they apply to your institution.