Singapore's AI risk management guidelines were formally issued by the Monetary Authority of Singapore (MAS) on 7 October 2026 under the title Guidelines on Artificial Intelligence Risk Management. The framework had been consulted on as paper P017-2025 from 13 November 2025 to 31 January 2026, and this release is the final text together with MAS's response to feedback. It governs financial institutions in Singapore. It is not a general AI law for every industry, and it is not a rulebook other countries can simply adopt as their own.
Its legal nature needs stating plainly. This is a set of supervisory guidelines setting out what MAS expects of the firms it regulates, not a newly enacted statute, and the release materials list no schedule of fines. Guidelines are not optional reading either: firms face supervisory scrutiny, and boards and senior management must be able to explain their AI risk appetite, who owns which responsibility, and how controls fit together. Existing governance structures may be reused where they are adequate; MAS says no dedicated AI committee has to be created solely for this purpose.
Two deadlines: governance in 2027, full lifecycle by 2028
The timetable has two stages. Sections 3 and 4 are to be met by 7 October 2027, covering board and senior management oversight, AI identification and inventory, risk materiality assessment, and the supporting systems and procedures. Sections 5 and 6 are due by 7 October 2028, covering controls across the full AI lifecycle, including data governance, testing, human oversight, cybersecurity, ongoing monitoring, change management, third-party AI management and staff capability.
The guidelines apply to all MAS-regulated financial institutions, including banks, insurers, payment firms and capital markets firms, and to all forms of AI, whether built in-house, generative, or embedded inside a business system. Controls scale in proportion to how much AI a firm uses and how much harm a given use could cause. Low-impact uses whose failure would not materially affect the firm, its customers or other stakeholders can be covered by basic policies and procedures; drafting internal emails or summarising meeting notes are the kind of use meant here, not a reason to build high-risk-grade paperwork for everything.
The easiest AI to miss is the AI hidden inside software
The final text singles out a blind spot. Where a service comes from a material third-party provider, the identification process should at least capture AI inside it, even when the product was never sold as an AI product. An automatic summary inside a CRM, smart classification inside a ticketing system, or a generate button inside a document platform can all be AI of this hidden kind. Inventories are expected to record purpose and description, approved scope of use, model type, data used, dependencies, lifecycle status and a materiality rating.
That rating decides how much effort follows. It should turn on impact, complexity and reliance, not novelty: a use that directly affects credit decisions, claims, trading or customer rights should not receive the same testing and approval as one that only fixes staff spelling. European institutions are facing a similar timed stocktake under their own regime, but the jurisdictions and subject matter differ, so one exercise cannot substitute for the other.
When third-party AI fails, accountability does not leave with the vendor
The heaviest line in the guidelines is that a financial institution remains accountable for AI used in its services even when a third party developed, operates or provides it. In practice, firms must obtain sufficient assurance from providers, judge whether the AI suits their intended use, and add compensating controls where disclosure is thin or practical constraints bite. A vendor's own statement is not enough on its own; the guidelines look for independent assessments or certifications as external evidence. Firms should also test third-party AI on their own use cases and their own data rather than rely on a vendor demonstration.
Contracts are part of the control set: visibility over when AI is introduced and later changed, advance notice and impact assessment of changes, contingency plans for failures, unexpected behaviour or withdrawn support, and audit arrangements. If the risk of an external AI service cannot be brought within the firm's own risk appetite, the options MAS names are blunt: consider limiting it, suspending it, or replacing it. A vendor who will not cooperate, or a model described as a black box, is not a defence under this framework.
Agents get a separate follow-up track
MAS also signalled that it plans to consult in 2027 on whether agentic AI needs additional guidance. Systems that can execute trades, update records or face customers on their own plainly cannot be handled with a generic inventory alone. Institutions already running agents in trading, treasury or customer workflows should separately map tool permissions, reasoning traces and human stop mechanisms now, rather than wait for that consultation.
Three moves are available today for firms operating in Singapore. First, build the inventory now, covering in-house, bought-in and SaaS-embedded AI, with a materiality rating for each entry; that is the hard core of the 2027 stage. Second, fix vendor evidence packs and contract terms early: independent assessment, change notification, audit rights, failure handling and the ability to switch off a single AI feature should be written in at signing, not renegotiated against the 2028 deadline. Third, complete lifecycle controls in risk order, starting data, testing, monitoring and human oversight with high-impact uses, and leave basic policies for low-impact ones so compliance effort is not spread evenly and thinly.
Vendors are not directly regulated by the guidelines, but the message reaches them first through due diligence. Whether a supplier can explain which upstream models it relies on, where its data comes from, what testing it has done, what independent assessment it holds, how version changes are notified and how a feature can be shut down will decide whether it stays on a Singapore financial institution's vendor list. Singapore financial AI compliance has moved from voluntary principles to dated supervisory expectations, and the party that starts late usually has to rebuild not one document but trust across a whole supply chain.
