On 7 October 2026, the Monetary Authority of Singapore (MAS) published its final Guidelines on Artificial Intelligence Risk Management. The consultation is over. Because their guidelines are final they will set the tone for the finance industry in the rest of the world
8 October 2026 · Bruce Cullen
On 7 October 2026, the Monetary Authority of Singapore (MAS) published its final Guidelines on Artificial Intelligence Risk Management. The consultation has closed. So has the debate about whether banks, insurers and payment firms need a formal AI risk discipline.
It would be easy to treat this as a Singapore story, and that would be a mistake. MAS has produced the most complete supervisory statement so far on how a financial institution should govern AI, including generative AI and autonomous agents. Regulators elsewhere are either catching up or deliberately waiting. In both cases, this text is about to become the reference point that boards, auditors and vendors measure themselves against.
This post covers what the guidelines say, what they mean for firms in Singapore, and why a bank in London, Frankfurt or New York should pay attention.
The guidelines apply to every financial institution MAS regulates and to every form of AI, from classic machine learning to generative AI and multi-agent systems. Simple rule-based tools and RPA that only follow fixed instructions are outside scope. The structure rests on four pillars.
Three details stand out for anyone who has run a governance programme.
The unit of assessment is the use case, not the model. The same foundation model can be low risk when it drafts internal meeting notes and high risk when it informs a credit decision. Inventories have to become linked records that connect each use case to its model, data and vendor, rather than a flat spreadsheet of tools.
Third-party AI got tougher, not softer. Firms stay accountable for AI in their services even when a vendor built, runs or supplies it. Vendor self-attestation is not enough, and MAS looks for independent assessment. If a provider's residual risk can't be brought within appetite, the firm should limit, suspend or replace that service. Dependence on a single provider can even become a board-level risk-appetite metric.
Agents are named and governed directly. MAS expects each agent's identifier, the tools it can reach and its guardrails to be inventoried. Guardrails must be tested, including through red teaming. Reasoning, actions and tool calls should be logged and monitored. High-materiality AI needs a kill switch, and its activation must be tested regularly. MAS has also said it will consult again in 2027 on further guidance for agentic AI.
Proportionality runs through all of it. A firm whose AI use is unlikely to cause material harm can meet the guidelines with basic policies. It must still check periodically that it qualifies for that lighter track.
The clock is now running, in two stages. The consultation proposed a single 12-month transition. The final text splits it so that firms show they know where their AI is before they are judged on how well they control it.
| Date | What applies |
|---|---|
| 7 October 2027 | Sections 3 and 4: oversight, identification, inventory and materiality assessment |
| 7 October 2028 | Sections 5 and 6: life cycle controls, capability and capacity |
Twelve months for the first stage sounds generous until you account for discovery. Most firms will find far more AI than they expect. There will be copilots switched on inside productivity suites, models embedded in CRM and fraud platforms, and tools that teams adopted without asking anyone. Each one needs an owner and a materiality rating with a written rationale.
Formally, the guidelines are supervisory expectations rather than legislation. In practice the distinction matters less than it sounds, because MAS will look for evidence in inspections. A firm that can't produce its inventory, its materiality reasoning and its test results on request will get little credit for a well-written policy.
Regional and global groups face an extra requirement. Oversight can sit at group level, but local senior management in Singapore must still be able to show MAS how it exercises oversight. "Head office handles it" is not an answer.
The 2028 work also has the longest lead time. Independent validation capacity, tested fallback plans and renegotiated vendor contracts all take budget and time. Those contracts should give firms the right to be told when AI is introduced or updated, and the right to audit. These workstreams should start now, not in late 2027.
Supervisors around the world now largely agree on what to govern. They disagree sharply on how, and on how much to write down. That gap is why the MAS text matters globally.
| Jurisdiction | Approach as of October 2026 | Agents |
|---|---|---|
| Singapore | One sector-wide set of AI risk guidelines, phased in over 2027–2028 | Named, with specific controls |
| EU | Horizontal AI Act. After the Digital Omnibus delay, obligations for high-risk uses such as credit scoring apply from 2 December 2027 | Governed by risk tier and role, with no agent-specific regime |
| UK | No new AI rules. Existing model risk rules (SS1/23), Consumer Duty and SM&CR apply | Covered indirectly |
| US | SR 26-2 replaced SR 11-7 in April 2026 | Generative and agentic AI explicitly out of scope |
| Australia | APRA letter of April 2026, applying existing standards CPS 220, 230 and 234 | Flagged, especially the identity of non-human actors |
| Hong Kong | Sandbox-led, with agent guidelines promised for 2027 | Pending |
This comparison is summarised from Luke Soon's six-regime analysis. Positions are moving quickly.
Three consequences follow for institutions outside Singapore.
It is the de facto agent rulebook. If you deploy an AI agent in a US bank tomorrow, federal model risk guidance won't tell you what good looks like. In the UK you have to infer it from older rules. MAS has written it down. Internal audit teams, consultants and other supervisors are likely to borrow its vocabulary, including inventory, materiality, guardrail testing and kill switches, whether or not they hold a Singapore licence.
Global groups inherit it directly. Any group with a Singapore entity must meet these expectations locally, and local management must be able to evidence its oversight. The efficient answer is a single control set built to the highest bar in each area and mapped out to each regime. For agents, the highest bar is currently Singapore's.
Third-party AI is everyone's blind spot. Regulators in Singapore, Australia and the UK have each flagged reliance on a small number of AI providers. MAS has gone furthest in putting the burden on the buyer. Model providers and SaaS vendors that sell into finance should expect contract clauses, independent assurance requests and audit rights to spread well beyond Singapore. Vendors that can show independent evidence will be well placed. Those that offer only self-attestation will struggle.
The regulatory direction is also clear. Hong Kong has committed to agent guidelines, US agencies have signalled further consultation, and APRA has already named the identity gap. Firms that build to the MAS standard now are unlikely to need to rebuild later.
These steps align a firm with the direction every major regulator is taking, whether or not MAS supervises it.
The sequencing is adapted from Luke Soon's commentary on the final text.
MAS has turned a vague question, "are we using AI responsibly?", into one that can be audited. Where is our AI, who owns it, what does it depend on, and what happens when it goes wrong? None of that is new. It is the discipline financial institutions already apply to credit, market and operational risk, extended to systems that can now act on their own.
Firms that treat this as a Singapore compliance project will end up doing the work twice. Firms that treat it as the global baseline will be ready when London, Brussels, Washington and Hong Kong catch up, as they are likely to.
Claude for Government is now generally available. Two-person approval for sensitive admin actions is the standout. Here is what that asks of an organisation, plus the budget, SSO and MDM changes.
OpenAI launched Dots at DevDay 2026, bundling compute for computer-using agents into its subscriptions. What that might mean for the cloud providers underneath.