Most banks know what software they buy, but fewer know what is running inside it. A customer-service platform, a fraud-detection tool, a document-processing system—each of these may have had AI quietly switched on in a recent product update, without the bank being consulted. Singapore has just decided that it does not matter. If the AI is in software a bank pays for, the bank is responsible for it.
On October 7, the Monetary Authority of Singapore issued its final Guidelines on AI Risk Management, setting out how MAS expects banks, insurers, payment providers and other regulated firms to govern AI. Banks had asked MAS to limit the scope to AI they knowingly procured. MAS declined. “FIs [financial institutions] remain accountable for AI used in the services they deliver,” it said.
What does that mean in practice? Take DBS for example. DBS has deployed AI across more than 430 use cases, including a tool that drafts corporate credit memos by working through more than 70 tasks. From October 2027, DBS and other MAS-regulated institutions will be expected to manage AI risks across their operations, including AI inside the software they buy.
The Guidelines take effect in stages from October 2027. The responsibility of meeting them falls mainly on heads of vendor and third-party risk, chief risk officers, model risk leads, CIOs, and procurement heads at MAS-regulated financial institutions. The software vendors serving these institutions face a parallel task: their product and sales leaders need to be ready with answers when customers start raising questions.
What the Guidelines Ask For
The Guidelines require banks to catalog the AI they use, as far as they reasonably can, including AI embedded in vendor products. MAS gives a specific example: software sold as a business tool, not as an AI product, that contains AI features.
Banks are expected to keep that list current, assess the risk level of each AI application, and apply controls proportionate to that risk throughout the system’s life. Board members and senior leaders carry explicit responsibility under the Guidelines.
Where full identification isn’t possible—for example, because a vendor won’t disclose what’s running inside its product—MAS still expects banks to identify the risks that arise from that gap and put mitigating measures in place.
On vendors more broadly: banks are expected to obtain sufficient assurance that third-party AI is appropriate for the task, and apply additional safeguards where there are gaps. Contracts should require vendors to disclose when AI is added to a product.
The harder case is a critical vendor that cannot explain the AI in its product, where replacing it before the deadline would be costly or disruptive. MAS expects banks to consider restricting, pausing or replacing such a service, but leaves the choice of response to the bank. Banks can press for assurance by other routes, such as independent testing results or audit reports. They can switch off or limit the specific AI feature carrying the risk, or add compensating controls like human review of outputs. They can renegotiate disclosure and audit rights. Alongside all of this, they should begin assessing alternatives early, so that replacement stays an option.
Why This Matters Beyond Singapore
Singapore is not the only regulator in Asia looking at AI in financial services. But the Guidelines do not automatically apply to banks operating only in other ASEAN markets. They reach foreign institutions operating through MAS-regulated entities, which covers a large share of the region’s financial infrastructure.
The practical effect is that Singapore’s standards tend to become the regional floor. A vendor that must answer AI disclosure questions for its Singapore-regulated clients will find it simpler to extend the same disclosure everywhere. A bank that builds an AI inventory for MAS compliance will have it ready when regulators elsewhere ask the same question.
MAS also noted that the Financial Stability Board, the international body that coordinates financial regulation across G20 countries, has since consulted on sound practices for financial institutions’ responsible use of AI. The FSB’s involvement is critical because its recommendations typically become the template that member countries adopt into domestic frameworks.
For multinational banks running shared AI platforms across Southeast Asia: Singapore-regulated operations are expected to meet MAS supervisory expectations regardless of where those AI systems were developed or procured. A platform built and managed out of Hong Kong or Mumbai that serves the Singapore entity will need to comply. For many regional banks, that makes the inventory exercise a group-wide task, not a local one.
What to Do Now, and What’s Coming
MAS began supervisory engagement with the industry from October 7, the day the Guidelines were published. Banks have until October 7, 2027 to meet the core expectations on governance and AI inventory, and until October 7, 2028 for the full framework.
The expectations scale with the size and nature of AI use. A smaller institution whose AI use is unlikely to affect customers or the wider financial system materially can meet the Guidelines through basic policies rather than the full framework. A small payments firm with a handful of AI tools faces a different task than a bank running hundreds of models across lending, trading, and operations.
The order of work depends on who is doing it.
Risk and model governance leads should start with identifying the AI use cases that touch critical operations or consequential customer decisions, such as credit, fraud flags and claims, and mapping where vendor assurance is thin.
CIOs and procurement heads should start by listing the vendors the bank relies on most and checking when each contract is up for renewal. Contracts that renew before October 2027 are a good opportunity to add clauses that require vendors to give notice before switching on new AI features, explain how those features work, and support the bank’s own reviews. For critical vendors whose contracts run past the deadline, there’s no renewal to wait for, so those conversations should start now.
The agentic AI question is not yet fully settled. MAS has said it will consult on AI agents in 2027. The Guidelines issued this month do not yet cover that territory in detail. Banks already deploying AI agents should manage their risks under the current framework rather than wait for additional guidance.
For technology vendors, MAS’s expectations create a commercial reality that does not require waiting for a bank to ask. MAS expects institutions to drop vendors who cannot provide sufficient assurance. That makes AI transparency a condition of doing business with Singapore-regulated financial institutions. Software vendors should prepare before customers ask. Expect questions on where AI sits in the product, what risks it introduces, and what evidence supports it, from testing results to audit reports.
For all the details in the Guidelines, the most useful step now for risk, technology and procurement leaders who want to meet MAS’s expectations is to identify the critical vendors whose AI they still can’t explain. Those are the conversations that take longest, whether they end in better assurance, new controls or a plan to move elsewhere. For vendors, the opportunity is to get ahead of those questions and come to customers with answers ready.
More from Asia Tech Lens
AI Safety Evaluators Are Critical Vendors. Nobody Is Treating Them As One.
The firms that test AI models for safety face no accreditation, liability framework or duty to report failures, yet buyers rely on their results. This looks at how procurement teams should treat evaluators as critical vendors and write breach notification and independent verification into contracts.Agentic AI Can Act. Singapore’s New Rulebook Says: Prove You Can Stop It.
Singapore’s voluntary agentic AI framework is already used as a benchmark in procurement and audits. Argues organizations should give agents only the autonomy they can audit and reverse, with human approval for high-impact actions.AI Middleware Promises Flexibility. In Asia, It Can Create Jurisdictional Lock-In The layer sold to avoid being locked into one model can become a harder dependency to unwind, tied to how data is routed and logged across jurisdictions. Sets out what tech leaders should check on data routing, retention, audit export and exit before committing to a vendor.
Nvidia Is Acquiring Hugging Face. Here Is What That Means for Operators.
A platform many treated as neutral now has an owner with its own strategic interests, and that changes the risk for Asian users who depend on it for open-weight models. Recommends listing which models you depend on and confirming they can be redeployed without the platform.China Is Building a Second Pole in Global AI Governance
The Shanghai-based WAICO bloc sets up a cooperation-led alternative to the EU AI Act, leaving organizations in Asia with two increasingly incompatible frameworks. Argues the best-prepared will be those that can switch models, providers and compliance controls without rebuilding workflows.