In brief: MAS issued its final Guidelines on AI Risk Management on 7 October 2026. They take effect on 7 October 2027 for oversight, identification, inventory and risk materiality (Sections 3–4), and by 7 October 2028 for life cycle controls and capabilities (Sections 5–6). For AI agents, the hardest part is the inventory: MAS says agent records may need agent-specific identifiers, the tools and systems the agent can access, its components and its guardrails. This post sets out a practical plan to have that inventory working before October 2027.
Who this is for: Chief Risk Officers, heads of model risk and technology risk, and Chief AI or Data Officers at banks, insurers and other MAS-regulated financial institutions who must show the board where AI agents are used, how risky each one is, and who is accountable.
Key Takeaways
- The MAS Guidelines apply to all financial institutions and all forms of AI, and MAS names AI agents, including multi-agent systems, as in scope.
- From 7 October 2027, institutions are expected to have board oversight, a consistent way to identify AI use, an AI inventory and a risk materiality assessment for each AI use case.
- For agents, MAS points to extra inventory attributes: agent-specific identifiers, the tools and systems the agent can access, its components and the guardrails imposed.
- Risk materiality is assessed on three minimum dimensions: impact, complexity and reliance. Reliance explicitly includes the autonomy granted to the AI and the degree of human oversight.
- Life cycle controls follow by 7 October 2028, including human oversight, contingency plans and tested "kill switch" protocols for high-risk use cases.
- MAS intends to consult the industry in 2027 on further guidance for agentic AI, so the inventory design should be built to change.
What Are the MAS Guidelines on AI Risk Management?
On 7 October 2026 the Monetary Authority of Singapore issued its Guidelines on Artificial Intelligence Risk Management, following a public consultation in November 2025. The Guidelines set supervisory expectations for financial institutions (FIs) to manage AI risks at both the enterprise level and the individual use case level, and to apply controls in proportion to each use case's risk materiality.
MAS summarises four key expectations:
- Oversight with clear accountability. The board and senior management oversee AI risks, set roles, risk appetite and frameworks. Existing governance structures can be used, and a dedicated AI committee is not required.
- Identify, assess and manage AI risks across the life cycle. Identify AI use, keep inventories at an appropriate level of granularity, assess risk materiality, and apply proportionate controls such as data governance, testing, human oversight, cybersecurity, monitoring and change management.
- Manage third-party AI risk. FIs remain accountable for AI in the services they deliver, including AI developed or operated by third parties.
- Be proportionate. Basic policies and procedures can be enough where poor performance or unavailability of an AI tool is unlikely to have a material impact.
When Do the MAS AI Guidelines Take Effect?
The Guidelines take effect on 7 October 2027, and MAS allows a phased approach:
| Deadline | Sections | What they cover |
|---|---|---|
| 7 October 2027 | Sections 3–4 | AI oversight by the board and senior management; AI identification, AI inventory and risk materiality assessment |
| 7 October 2028 | Sections 5–6 | AI life cycle controls (including data management, human oversight, third-party AI, monitoring) and AI capability and capacity |
In other words, the first year is about knowing what AI you have and how risky it is. That is exactly where agents are hardest to pin down.
Do the MAS Guidelines Cover AI Agents?
Yes. The Guidelines define AI broadly and list AI agents, including multi-agent systems, among the technologies in scope. MAS also explains why agents need attention: an agent with access to tools could take unauthorised or erroneous actions if its goals are translated into actions in ways the institution did not intend, and a compromised agent with access to internal systems could be used to exfiltrate data or run malicious commands at scale.
MAS points institutions to the IMDA Model AI Governance Framework for Agentic AI for agent-specific practices, and says it intends to consult the financial sector in 2027 on what additional guidance on agentic AI would be useful.
What Should an AI Agent Inventory Record Contain?
Paragraph 4.7 of the Guidelines lists attributes an AI inventory may include: purpose and description, approved scope of use, model type, data used, dependencies, life cycle status, risk materiality rating, model review status, key roles and responsibilities, and links to documentation. For AI agents, MAS adds that relevant attributes may include agent-specific identifiers, the tools and systems the agent can access, its components, and the guardrails imposed to manage risk.
We group these into one record per agent, which we call the Agent Inventory Record:
| Block | Fields | Where it comes from in the Guidelines |
|---|---|---|
| 1. Identity | Agent ID, name, version, business owner, technical owner | Agent-specific identifiers; key roles and responsibilities (4.7, footnote 25) |
| 2. Purpose and scope | What the agent is for, approved scope of use (business line, jurisdiction, customer segment) | Purpose and description; approved scope of use (4.7) |
| 3. Components | Underlying models and versions, third-party providers, other agents it calls | Model type; dependencies; components (4.7, footnotes 25 and 26) |
| 4. Tool and data access | Systems and tools the agent can read from or act on; data used; read vs write permissions | Tools and systems the agent can access; data used (4.7, footnote 25) |
| 5. Guardrails | Actions it may not take, approval thresholds, escalation rules, how it can be stopped | Guardrails imposed (footnote 25); kill-switch protocols (5.3) |
| 6. Risk materiality | Inherent and residual rating on impact, complexity and reliance (autonomy and human oversight) | 4.10–4.12 |
| 7. Status and evidence | Life cycle status, last review, links to testing, approvals and decision logs | Life cycle status; review status; links to documentation (4.7); documentation and logs (5.9) |
Two points are easy to miss. First, the inventory should link to the institution's other inventories, for example data inventories and third-party or outsourcing registers (4.6). Second, MAS expects the inventory design itself to be reviewed as new AI technologies appear (4.8), so the record should be easy to extend when MAS consults on agentic AI in 2027.
One-day decision-register workshop
Bring one AI agent or agent-assisted workflow. In one day we fill in its Agent Inventory Record together: identity, scope, components, tool access, guardrails and a draft risk materiality rating against impact, complexity and reliance.
Book the one-day workshopHow Should Risk Materiality Be Assessed for an AI Agent?
MAS asks FIs to assess risk materiality for each AI use case on at least three dimensions (4.12):
- Impact: the consequences of failure, malfunction or poor performance for the institution, its customers and other stakeholders, including the sensitivity of the data processed.
- Complexity: the nature and novelty of the technology, the data it uses and how explainable its outputs are. For third-party AI, how much visibility the FI has.
- Reliance: how much the decision or output relies on AI, including the autonomy granted to the AI and the degree of human involvement or oversight.
For agents, reliance usually moves the most. An assistant that drafts a reply for a relationship manager and an agent that executes a payment change through a tool can use the same model, yet their reliance and impact ratings are very different. The assessment should rate both inherent risk (before controls) and residual risk (after controls), and residual risk must sit within risk appetite before deployment (4.11).
Illustrative example: a bank runs a dispute-handling agent that can read card transaction data and propose refunds. If it can only propose, reliance is moderate because a human approves each refund. If it can issue refunds below a threshold on its own, reliance rises, and the record should show the threshold, the escalation rule and how the agent can be stopped.
What Is a Practical Plan to Be Ready by October 2027?
This is our view of a workable sequence. Timings will vary by institution.
- Agree the definition and the owner. MAS expects a designated control function to oversee AI identification, the inventory and risk materiality assessment, and to be the final arbiter on what counts as AI use (4.3, 4.9, 4.13). Decide who that is and confirm that agents built on low-code platforms or embedded in vendor software are in scope.
- Find the agents. Look beyond major AI projects: AI embedded in software-as-a-service, agent features switched on in productivity tools, and shadow AI. Where you cannot see everything, MAS expects you to identify the resulting risks and put mitigating measures in place (4.4).
- Write one Agent Inventory Record per agent. Start with the agents that can act on systems (write access, payments, customer communications) before read-only assistants.
- Rate risk materiality. Apply the impact, complexity and reliance dimensions consistently and have the control function approve the ratings.
- Report to the board. Make sure AI risk appears in the risk appetite framework. MAS gives examples such as thresholds on AI-related incidents or the number of material use cases depending on a single provider (footnote 18).
- Prepare for 2028. For high-risk agents, start designing human oversight, contingency plans and tested stop mechanisms now, because these are expected under Section 5 by October 2028.
How Does Decision Tracing Help With the MAS Expectations?
An inventory says what an agent is allowed to do. Supervisors and internal audit will also want to see what it actually did. MAS asks FIs to keep logs and records from deployment, operation, monitoring and incident management, to review human oversight decisions and interventions, including near misses (5.9). Under ongoing monitoring, it adds that enhanced documentation for generative AI or AI agents could include logging prompts and model responses, with model versions and reasoning processes (5.23).
A decision trace is one way to meet this: a record for each consequential agent action of the inputs and context used, the policy checks applied, any human approval, the action taken and the result. ElixirData Context OS is designed to produce these traces and to check an agent's actions against its approved scope and guardrails before they execute. You can read more in How Decision Traces Build Trust in AI Agents and How AI Agent Audit Trails Support Bank Compliance.
Frequently Asked Questions
What are the MAS guidelines for AI risk management?
They are supervisory guidelines issued by the Monetary Authority of Singapore on 7 October 2026. They set expectations for board oversight, AI identification, an AI inventory, risk materiality assessment, life cycle controls and AI capabilities at financial institutions, applied in proportion to risk.
Do the MAS AI guidelines apply to generative AI and AI agents?
Yes. The Guidelines apply to all forms of AI technology and name generative AI and AI agents, including multi-agent systems. Rule-based systems, tools whose outputs come from manually specified formulae, and robotic process automation that follows fixed instructions would not ordinarily be considered AI under the Guidelines.
Do we need a separate AI committee?
No. MAS says existing governance structures can be used where they provide adequate oversight and cross-functional coordination, and an FI need not set up a dedicated AI committee only to meet this expectation.
Can we extend our existing model inventory instead of building a new one?
Yes. MAS allows FIs to enhance existing inventories or create a dedicated AI inventory, as long as there are clear links to other relevant inventories such as data inventories and third-party registers.
Can we keep agent records and decision traces in our own environment?
Yes. Context OS can be deployed as managed SaaS, in your own VPC or on-premises, so inventory data and decision traces can stay inside your environment. See Context OS deployment models.
Next Step
October 2027 is close for a full inventory of AI agents across business lines. Start with one agent that can act on systems and build its record properly; the rest follow the same pattern.
Book a one-day decision-register workshop
Related Reading
- AI Agent Governance for Banks in 2026: A Practical Guide
- How AI Agent Audit Trails Support Bank Compliance
- AI Agent Identity and RBAC: Machine-Grade Access Control
- What Is AI Policy Enforcement for Enterprise Agents?
Related on ElixirData: Banking & Financial Services · Risk & Compliance
Sources
- Monetary Authority of Singapore, "MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions", media release, 7 October 2026. mas.gov.sg
- Monetary Authority of Singapore, Guidelines on Artificial Intelligence Risk Management, 7 October 2026 (paragraphs 1.8, 1.11, 3.4, 4.2–4.13, 5.3, 5.9, 5.23 and footnotes 18, 25, 26). mas.gov.sg
- Infocomm Media Development Authority, Model AI Governance Framework for Agentic AI. imda.gov.sg
This article is general information, not legal or regulatory advice. Check the full Guidelines and your own obligations with your compliance team.
