Singapore · MAS
The Guidelines on Artificial Intelligence Risk Management, issued on 7 October 2026, apply to every financial institution MAS licenses and cover generative AI and AI agents. Board oversight, a quantified appetite, an inventory and a materiality rating are expected from 7 October 2027; lifecycle controls by 7 October 2028; controls for high-risk uses as soon as possible.
What applies
The Guidelines set supervisory expectations, which MAS takes into account in supervising an institution. The notices behind them are already in force.
| Instrument | What it asks | What it means for evidence |
|---|---|---|
| MAS Guidelines on Artificial Intelligence Risk Management, 7 October 2026 | Apply to every financial institution MAS licenses, including generative AI and AI agents. Board oversight, AI risk appetite, an AI inventory and a materiality rating for each use from 7 October 2027. Lifecycle controls and capability by 7 October 2028. Controls for high-risk uses as soon as possible. | An inventory and materiality rating that reach AI supplied by your group and by third parties, and an appetite with numbers in it. |
| MAS Notice 658, Management of Outsourced Relevant Services for Banks | Outsourcing agreements for material services carry specified terms, including audit rights, from the date each agreement is entered into or next extended (para 14.2). A bank's own offices outside Singapore count as service providers (para 2.1). | AI terms ride the renewal cycle. An agreement renewed in 2027 without them may not come round again before October 2028. |
| MAS Notice FSM-N05, Technology Risk Management | Notify MAS of a relevant incident within one hour of discovery, and report the root cause within 14 days. | A record of what the AI did has to exist at the time, not be reconstructed afterwards. |
| Guidelines on Individual Accountability and Conduct | Senior managers are identified by their responsibility for the institution's operations in Singapore. | Named people who will want to point to the same evidence the group holds. |
Where MAS is specific
Each is a point where a self-attestation from a provider is not enough, or where the record has to exist before something goes wrong.
Para 3.4
AI risk appetite should include quantitative measures where appropriate, such as thresholds on AI incidents and the number of material AI uses that depend on a single provider.
Paras 5.10, 5.11(a)
External assessments of a provider should be by independent and competent parties, not self-attestations. Keep documentation of how you tested third-party AI for your own use.
Para 5.11(f)
Audit rights, and notification when AI is introduced and when it is updated. Where residual risk is outside appetite, consider limiting, suspending or replacing the provider.
Paras 5.3, 5.23
For high-materiality AI, consider kill switches or override, with activation tested. Access to AI systems and configuration controlled and logged; logging of prompts and responses may be considered.
The seven questions, in MAS terms
"No reasonable reach" is a documented conclusion, not a gap. The determination at each step is yours and your counsel's.
A working timeline
Two items set the pace. Contract terms that miss a 2027 renewal may not be available again before October 2028. And monitoring records for 2027 and 2028 can only be produced in 2027 and 2028.
| Window | What MAS expects | What the institution does |
|---|---|---|
| Nov–Dec 2026 | Guidelines issued 7 October 2026. Transition running. | Gap assessment against sections 3 to 6. Name the locally accountable senior manager. Decide whether to adopt the group framework (para 1.5), and record why it meets the Guidelines. |
| Jan–Mar 2027 | A control function as final arbiter of what counts as AI (para 4.3). An inventory with the para 4.7 attributes, including AI supplied by the group and by third parties. | |
| Mar–May 2027 | A materiality rating for each use on impact, complexity and reliance, inherent and residual (paras 4.11, 4.12). | |
| Apr–Sep 2027 | Contracts renewing before October 2028 | Audit rights and notice of AI changes in each renewal (para 5.11(f)). Independent assessments of providers (para 5.11(a)). Your own testing documented (para 5.10). |
| May–Jun 2027 | Board, or a committee it delegates, approves AI in the risk appetite, with quantitative measures (para 3.4). | |
| Jul–Sep 2027 | High-risk uses: controls as soon as possible | Lifecycle controls applied early to high-materiality uses. A readiness review by the second line or internal audit. |
| 7 Oct 2027 | Sections 3 and 4 expected | Oversight, identification, inventory and materiality in place, with evidence. |
| 2027, date not set | MAS consults on further guidance for AI agents | Agent identifiers, tool access and guardrails already recorded in the inventory. |
| Oct 2027–Jun 2028 | Sections 5 and 6 build | Kill switches or override for high-materiality AI, with activation tested (paras 5.23(b), 5.3). Logged access to AI systems and configuration (para 5.23(d)). |
| 7 Oct 2028 | Sections 5 and 6 expected | The full framework in place, with a year of monitoring records, incidents and remediation on file. |
The sequence and windows are Triodian's estimates for a large institution, not dates set by MAS, except where a MAS date is shown.
For branches of foreign groups
A branch is not locally incorporated, so the Guidelines apply to it directly. It may adopt its group's AI risk management framework if that framework meets the Guidelines (para 1.5), and local senior management stays accountable for that choice (para 3.6). Any gap in the group framework becomes a gap MAS can point to with paragraph numbers.
One framework, many supervisors →Regulatory position as at 26 October 2026. Not legal advice: every legal determination is yours and your counsel's.
This page summarises published texts for orientation. Your counsel grades what is a requirement and what is an expectation for your institution.