TL;DR MAS issued AI risk management guidelines for Singapore financial institutions on 7 October 2026. They cover third-party AI, inventories, risk assessment and board oversight, with expectations phased in from 7 October 2027 and 7 October 2028. Vendors should prepare clear evidence about their systems, updates, testing and limits.
What happened
The Monetary Authority of Singapore (MAS) issued its Guidelines on Artificial Intelligence Risk Management on 7 October 2026. The guidelines set supervisory expectations for financial institutions (FIs) using AI, and apply to all FIs and all forms of AI technology. MAS says each institution should tailor its approach to the nature and scale of its AI use and the materiality of the associated risks.
| Topic | What MAS says |
|---|---|
| Who is covered | All financial institutions and all forms of AI |
| First phase | Sections 3 and 4 from 7 October 2027 |
| Later phase | Sections 5 and 6 by 7 October 2028 |
| Third-party AI | Financial institutions remain accountable for AI used in services they deliver, including AI provided by third parties |
| Inventory and risk | Identify AI use, maintain inventories at an appropriate level of detail, and assess use-case materiality |
| Oversight | Boards and senior management should provide effective oversight |
The dates matter for buyers planning governance work with their suppliers. MAS says the guidelines take effect on 7 October 2027. Institutions should meet the expectations in Sections 3 and 4 from that date, and the expectations in Sections 5 and 6 by 7 October 2028.
The guidelines are directed at financial institutions, not their vendors. But MAS says an FI remains accountable for AI used in the services it provides, including AI developed, operated or supplied by a third party. It expects the FI to assess whether third-party AI is suitable for its intended use and to obtain sufficient assurance from the provider. If risks cannot be brought within the FI's risk appetite, MAS says the FI should consider limiting, suspending or replacing the service.
This is the practical connection for a small software or services company selling AI to a bank, insurer, payment firm or other Singapore financial institution: the customer has to manage the risk of the AI it uses, so it needs enough information from the supplier to make and document that decision. MAS does not prescribe a single vendor questionnaire. It does make clear that third-party AI risk is part of the institution's own risk management.
MAS' consultation history helps explain the final shape. It opened a consultation on 13 November 2025, with comments due by 31 January 2026, and published its response on 7 October 2026. The release says respondents asked, among other things, how to manage risks from embedded AI and whether existing governance structures could be used, and that MAS retained the key expectations and refined them to address that feedback. It also says MAS intends to consult the financial sector in 2027 on what additional guidance on agentic AI would be useful.
What it means for a small team that sells AI
The new expectations do not make a small vendor responsible for the bank's governance framework. They do give the FI a reason to ask its supplier for usable evidence. A vendor that can explain what its AI does, what it depends on and how it changes can make that assessment easier for the buyer.
Start by describing the service in the customer's terms. Identify which parts use AI, what inputs they receive, what outputs they produce, and whether the output informs or triggers a decision. Separate the model or AI component from the surrounding product where that distinction is meaningful. If a feature is optional, say how it is enabled and whether the customer can disable it. If you cannot identify a component in a third-party service, state that limitation plainly and describe what you have asked the upstream provider.
Next, define the intended use and boundaries. The customer needs to judge whether the service is suitable for its use case, and the same tool may pose different risks depending on that use. Explain which uses your product supports, which uses you have not evaluated, and where human review is expected. Avoid broad claims such as "safe for financial services" unless you can support them with specific evidence and a clear scope.
Prepare a concise account of how you test and monitor the system. MAS lists testing, human oversight, cybersecurity, monitoring and change management among lifecycle controls for FIs. For a vendor, useful supporting material can include evaluation methods, known limitations, how you detect material performance changes, and what steps a customer can take when the service behaves unexpectedly. Keep the documents tied to the version and configuration the customer uses.
Make changes legible. Change management is on MAS's list of lifecycle controls, so an FI will want to know how your updates reach it. A vendor should be ready to explain its release process, what kinds of changes it can notify customers about, and what information it can provide after an update. The release also says FIs should apply compensating controls where assurance gaps arise, so a customer that cannot see your changes in advance may add monitoring of its own.
Do not assume that a statement from the supplier will always be enough. MAS says FIs should obtain sufficient assurance from third-party providers, assess whether the AI is suitable for its intended use, and apply compensating controls where practical constraints or assurance gaps arise. If the risks cannot be brought within the FI's risk appetite, it should consider limiting, suspending or replacing the service. That gives a buyer a reason to ask for evidence rather than accept a general promise.
This matters for a small vendor that cannot reveal proprietary model details. These are our suggestions, not MAS requirements, but you can still prepare evidence that speaks to the buyer's use: a scoped independent assessment, test results on representative scenarios, documented limitations, and a clear description of controls around customer data. Do not promise access to information you cannot provide. Explain the boundary and offer evidence that answers the risk question as directly as possible.
Keep an inventory of your own AI dependencies, even if the buyer's formal inventory is its responsibility. Record the product feature, model or provider, data involved, intended purpose, owner, deployment status and material dependencies.
A useful vendor response is not a long marketing document. It is a short, current evidence pack that lets the FI answer: what AI is involved, what is it used for, what can go wrong, how will we know when it changes, and what can we verify? MAS says the extent and sophistication of an FI's controls should depend on its risk exposure, so the more consequential the buyer's use case, the more detailed evidence it is likely to want.
For a small supplier, the work can be staged. First, map AI features and dependencies in the product you already sell. Then prepare a standard evidence pack, with a clear owner for updates. Review customer contracts for reasonable change notification and cooperation language. Finally, rehearse answering a buyer's questions without overstating what your tests prove.
Use the related AI vendor due diligence checklist to structure the evidence pack, or the third-party AI risk assessment template to walk through a specific use case. For contract language, see the AI vendor contract redline template. The AI regulation deadline calendar can help track the dates.
Copy and paste: AI vendor evidence request
Send this to your product, engineering and security owners before returning a financial-institution questionnaire.
AI service and scope
- Which product features use AI, including AI embedded in upstream services?
- What is the intended use, and what uses have not been evaluated?
- What input data is processed, and what outputs or actions can the system produce?
- Where does a person review or approve the output?
Evidence and limits
- What testing has been done for this feature and the customer's intended use?
- What are the known failure modes, performance limits and conditions where results may be unreliable?
- What independent assessment, certification or other evidence can we provide?
- What important model or training details cannot be disclosed, and what alternative evidence is available?
Changes and incidents
- Which material changes can affect the feature's behaviour, performance or risk profile?
- How will customers learn about those changes, and what notice can we reasonably provide?
- What monitoring, issue reporting and escalation processes are available?
- What service controls can the customer use to limit, pause or disable the AI feature?
Ownership
- Who owns each answer and supporting document?
- When was the evidence last reviewed, and what product version does it cover?
- Which upstream providers or services does the feature depend on?
This is a preparation list, not a MAS-prescribed questionnaire. Use the buyer's own format when one is provided. Keep each answer specific to the service and intended use, and mark gaps rather than filling them with assumptions.
What we could not verify
The facts above come from MAS's media release of 7 October 2026 and its consultation page, both opened for this article. We did not read the full text of the Guidelines or the response-to-feedback paper, so this article does not describe their detail, such as section contents, inventory granularity or how MAS treats particular products. CNA and The Business Times were not opened.
The release does not set a vendor compliance deadline: the phased dates apply to financial institutions, and the release does not mention a vendor questionnaire. How individual banks and insurers pass these expectations to suppliers, and what evidence they will ask for, is not stated and will vary by customer.
