MAS AIRG: final expectations, phased dates and evidence
MAS issued the final Guidelines on Artificial Intelligence Risk Management (AIRG) on 7 October 2026. They apply to all financial institutions and all forms of AI. This guide maps discovery, assessment and testing work to the relevant sections and phased dates. The Guidelines set supervisory expectations rather than law.
AIRG paragraphs 5.14, 5.15 and 5.17 address testing and its documentation by 7 October 2028. Evidence should show the scenarios tested, agreed thresholds, observed results and unresolved limitations.
Where the AIRG stands
13 November 2025
MAS opens the AIRG consultation
7 October 2026
MAS issues the final AIRG
7 October 2027
AIRG takes effect. Sections 3 to 4: oversight, identification, inventory and use-case materiality
7 October 2028
Sections 5 and 6: life cycle controls, AI capability and capacity
Paragraph 1.8 sets two implementation phases: Sections 3 to 4 from 7 October 2027, and Sections 5 and 6 by 7 October 2028. Discovery and use-case materiality provide the basis for the later control work.
Who it applies to
All financial institutions regulated by MAS: banks, insurers, capital markets intermediaries, and payment institutions. The Guidelines define AI broadly, covering models, systems, and use cases, while excluding purely rule-based tools.
Under paragraph 2.3, firms apply the full framework unless poor performance or unavailability of their AI is unlikely to have a material adverse impact on the FI, its customers or other stakeholders. Document the impact assessment. Paragraph 2.5 sets out basic policies and procedures for firms meeting that test.
What boards are expected to own
Paragraph 3.4 addresses board approval and regular review of the strategic direction for AI use and the governance approach. All risks from AI use belong in the risk appetite framework. Clear responsibilities and coordination support oversight; a dedicated cross-functional committee is not prescribed. Section 3 applies from 7 October 2027.
Paragraph 3.6 keeps local senior management accountable to MAS when a group framework is adopted. The group framework must meet the AIRG locally. Aivance checks the framework, local reporting, visibility and escalation paths, and records gaps and accountable owners.
The three core systems
Section 4 addresses identification, inventory and use-case materiality from 7 October 2027. These records provide the risk basis for proportionate life cycle controls under Section 5 by 7 October 2028.
01
AI identification
Paragraphs 4.2 to 4.4: find AI across business and functional areas, including AI embedded in material third-party services and shadow AI. Record visibility limits, residual risks and mitigating measures.
02
AI inventory
Paragraphs 4.5 to 4.9: maintain an accurate inventory to the extent possible or practicable. Link use cases to their systems, models, owners, providers and dependencies, with update responsibilities and review records.
03
Risk materiality assessment
Paragraphs 4.10 to 4.13: assess each AI use case, covering its systems and models. Record impact, complexity and reliance, inherent and residual risk, and the materiality rationale. A designated control function retains independent oversight and approves or arbitrates materiality.
Fourteen life cycle control groupings and practical test questions
These fourteen editorial groupings map the final AIRG Section 5 to practical tests. They are Aivance’s grouping of the text. Section 5 applies by 7 October 2028. Each expectation below cites the relevant paragraphs.
Use the questions to scope tests and identify gaps. Assess the complete applicable paragraph and use-case context before drawing a conclusion. Record test results and limitations alongside the findings.
Data management
AIRG expectation
Paragraphs 5.3 to 5.4: manage data quality, suitability, provenance and security through the AI life cycle, proportionate to use-case materiality.
The enforcement question
Can your pipeline reject or quarantine training and inference data that fails quality or provenance checks, before a human notices?
Fairness
AIRG expectation
Paragraphs 5.7 to 5.8: define fair outcomes and assess relevant harmful biases. Record results and mitigation steps for the use-case context.
The enforcement question
Where is the fairness threshold encoded in the system, and what happens at runtime when an output breaches it?
Transparency and explainability
AIRG expectation
Paragraphs 5.5 to 5.6: provide transparency and explainability appropriate to the use case, its materiality and the intended audience.
The enforcement question
Can you reproduce the explanation for a specific past decision, with the exact model version and inputs that produced it?
Human oversight
AIRG expectation
Paragraph 5.9: define proportionate human oversight, clear authority to intervene and records of decisions. Consider automation bias and decision fatigue.
The enforcement question
If the designated reviewer were unavailable right now, would the system halt and wait for ratification, or proceed without it?
Suspended Handoff State →Third-party AI
AIRG expectation
Paragraphs 5.10 to 5.11: the FI retains primary accountability. Test provider AI in the FI use case and document suitability. Independent, competent external assessments are a permitted way to address transparency limitations. Where residual risk exceeds appetite, consider limiting, suspending or replacing the provider.
The enforcement question
Which controls on vendor AI do you operate yourself, and which exist only inside the vendor’s assurances?
Model and feature selection
AIRG expectation
Paragraphs 5.12 to 5.13: document selection in light of use-case objectives, risks and complexity. Consider domain or user review where relevant.
The enforcement question
What technically prevents an unapproved model version from serving production traffic?
Evaluation and testing
AIRG expectation
Paragraphs 5.14 to 5.15: set measurable thresholds and test relevant failure modes. For generative AI and agents, test the effectiveness of associated guardrails.
The enforcement question
Does a failed evaluation technically block promotion to production, or does it produce a report someone may read?
Technology and cybersecurity
AIRG expectation
Paragraph 5.16: apply technology and cybersecurity controls proportionate to the materiality of the use cases the system supports.
The enforcement question
Do your AI service accounts hold IAM permissions wider than the actions your governance policy permits?
Reproducibility and auditability
AIRG expectation
Paragraph 5.17: document development, evaluation and testing with enough detail for an independent reviewer to understand and potentially replicate the work.
The enforcement question
Is your audit trail append-only and tamper-evident, or a log that a privileged user could edit after the fact?
Pre-deployment review
AIRG expectation
Paragraphs 5.18 to 5.22: review the use case and underlying systems before deployment. High-materiality use cases undergo formal independent validation. Record findings and actions.
The enforcement question
Is sign-off a deployment precondition enforced in the pipeline, or a meeting minute filed alongside it?
Human Ratification Gate →Post-deployment monitoring
AIRG expectation
Paragraphs 5.23 to 5.24: monitor deployed AI with risk-based intensity. Record thresholds, findings and remediation, and conduct reviews or re-validations as needed.
The enforcement question
Do threshold breaches halt or escalate execution automatically, or surface on a dashboard reviewed weekly?
Incident management
AIRG expectation
Paragraph 5.23(b) to (d): report, track, escalate and resolve incidents and issues. Assign accountable owners and retain monitoring and remediation records.
The enforcement question
Can you halt a misbehaving AI system mid-execution, and is a named person required to do so under defined conditions?
Override Architecture →Change management
AIRG expectation
Paragraph 5.25: assess the significance of changes by use-case risk. Review and re-approve significant changes; apply compensating controls when prior review is impracticable.
The enforcement question
Could a prompt or configuration change reach production today without leaving an approval record?
Decommissioning
AIRG expectation
Paragraph 5.26: retire AI with dependencies, data retention, secure removal and stakeholder notifications addressed.
The enforcement question
When a model is retired, what technically prevents downstream systems from continuing to call it?
Where GenAI and AI agents raise the bar
Paragraph 5.15 addresses testing the key failure modes of generative AI and AI agents, including the effectiveness of their guardrails. Test relevant hallucination, data leakage, adversarial and unauthorised-action scenarios. Retain the configuration, thresholds, results and limitations under paragraph 5.17.
Under paragraph 5.9, define human oversight according to use-case materiality. Aivance tests configured approval, intervention and escalation paths, including the outcome when a reviewer is unavailable. IMDA MGF v1.5 and SAFR inform test scenarios as references. Firms are not supervised against SAFR or the MindForge Toolkit and Handbook.
What to complete for 2027 and 2028
7 October 2027
Build the foundations
For Section 4, produce a linked inventory and documented use-case materiality assessments. Include material vendors, shadow AI, discovery limitations, owners and mitigating measures.
7 October 2028
Test controls against the enforcement questions
For Section 5, test relevant third-party AI controls and guardrail effectiveness. Record scenarios, thresholds, findings, residual risks and remediation. For Section 6, document the capabilities, resources and infrastructure needed to operate the framework.
7 October 2027
Make oversight reportable
For Section 3, document board and senior management responsibilities, the strategic direction for AI use, risk appetite and escalation paths. Include local accountability when using a group framework.
Common questions
Is the MAS AIRG in force?
MAS issued the final AIRG on 7 October 2026. Paragraph 1.8 states that it takes effect on 7 October 2027. Sections 3 to 4 apply from 7 October 2027; Sections 5 and 6 apply by 7 October 2028.
Who does the AIRG apply to?
The AIRG applies to all financial institutions and all forms of AI, including generative AI and AI agents. Under paragraph 2.3, firms apply the full framework unless poor AI performance or unavailability is unlikely to have a material adverse impact on the FI, its customers or other stakeholders. Paragraph 2.5 sets out the basic policies for firms meeting that test.
What is a risk materiality assessment under the AIRG?
Paragraphs 4.10 to 4.13 address assessment per AI use case, covering its systems, models and relevant dependencies. Assess inherent and residual risk across impact, complexity and reliance, with a documented rationale and control-function oversight. Section 4 applies from 7 October 2027.
How does the AIRG relate to FEAT and other references?
The AIRG sets MAS supervisory expectations rather than law. FEAT, Veritas, IMDA guidance, ISO 42001, NIST and other relevant materials retain their own scope and status. The MindForge Toolkit and Handbook are industry practice references. SAFR is an MAS publication on runtime safeguards for agentic finance. Firms are not supervised against MindForge or SAFR.
Does the AIRG cover generative AI and AI agents?
Yes, including multi-agent systems. Paragraph 5.15 addresses testing key failure modes and the effectiveness of associated guardrails. Section 5 applies by 7 October 2028. Retain the test scenarios, results and limitations under paragraph 5.17.
What should financial institutions complete for each phase?
By the first phase on 7 October 2027, establish oversight under Section 3 and discovery, inventory and use-case materiality under Section 4. By 7 October 2028, complete the relevant Section 5 life cycle controls and Section 6 capability and capacity work. Record owners, findings and evidence for each gap.
Is an independent third-party AI assessment mandatory?
Paragraph 5.11 permits an FI to consider assessments by independent, competent parties to address provider transparency limitations. It does not mandate an external assessor for every provider. Vendor self-attestation is distinct from that independent evidence. The FI retains primary accountability. Section 5 applies by 7 October 2028.
Can a Singapore operation use a group AI framework?
A group framework can be used only if it meets the AIRG. Paragraph 3.6 keeps local senior management accountable to MAS, with local visibility, input and escalation paths. Section 3 applies from 7 October 2027. Aivance checks the group framework against the AIRG and documents local gaps.
Update log
14 October 2026: Updated to the final AIRG issued on 7 October 2026, including both phases, impact-based proportionality, use-case materiality and guardrail testing.
12 June 2026: Guide first published.
This guide supports planning and assessment against the final AIRG. It does not constitute legal advice. Refer to the final Guidelines directly at mas.gov.sg and your legal counsel. The MindForge Toolkit and Handbook are industry practice references; SAFR is an MAS publication on runtime safeguards for agentic finance. They inform Aivance’s approach and are not supervisory assessment targets.
Scope the discovery, assessment or testing your institution needs.
Start with a 30-minute Authority & Control Review. The written scoping note identifies the use cases, applicable AIRG phase, deliverables and evidence required.