Where the Buck Stops: Singapore Holds Financial Firms Accountable for the AI They Build & Buy

Singapore’s financial regulator has told banks, insurers, payment firms and fund managers that they will be responsible for every AI system they use, including what they buy from third-party vendors. The strict stance anchors the Guidelines on Artificial Intelligence Risk Management, released by the Monetary Authority of Singapore (MAS) on October 7.

Stemming from a consultation that began in November 2025, these new rules address critical issues such as third-party blind spots and the need for a human-in-the-loop. Interestingly, no AI tool is exempt now. From admin bots to loan underwriting and insurance pricing, the guidelines blanket everything.

Ho Hern Shin, deputy managing director of MAS, said AI could improve customer outcomes, risk management and productivity. MAS said respondents to its consultation had backed risk-scaled guardrails.

The scope sets Singapore apart from Washington. In April, U.S. banking regulators rewrote their guidance on how banks manage models and left generative AI and AI agents out, describing the technology as novel and rapidly evolving. MAS has brought both within its rules.

The guidelines will take effect on October 7, 2027, with the full set applying from October 2028. They are supervisory expectations, not law, and firms whose AI use poses lower risks can apply lighter controls.

Two days after the rules were issued, MAS managing director Chia Der Jiun turned to AI agents that can initiate and execute transactions for customers. Speaking at the INSEAD Digital Finance and Agentic AI Summit on October 9, he said wider use of such agents could affect how banks manage liquidity and collateral, and he described the safeguards MAS is developing with industry partners.

Key Takeaways from MAS’s New AI Risk Management Guidelines

Boards are in charge	Board and senior management answer for how a firm uses AI and set how much AI risk a company will accept.
Know what you run	Firms should identify AI in use, including vendor-supplied and embedded systems, maintain inventories at an appropriate level of detail and assess the risk of each use case.
Controls must match risk	Basic safeguards for low-risk tools; independent validation is expected before launching high-risk uses.
No hiding behind vendors	Firms stay accountable for third-party AI and should consider restricting, suspending or replacing services if the risks cannot be brought within acceptable limits.
AI agents covered	The guidelines cover AI agents; SAFR adds voluntary checks before an agent acts.

Sources: MAS Guidelines on AI Risk Management (October 7, 2026); SAFR white paper (July 3, 2026).

The Problem with the AI You Do Not Build

Banks have long relied on outside technology suppliers. AI makes that reliance harder to supervise because a bank may know little about how a vendor trained its model, when it last changed it or why it produced a particular answer.

MAS now expects firms to map exactly where AI operates across their businesses, even when it is buried inside the services from key suppliers. This extends to standard software bought for other tasks that happens to have built-in AI features.

The regulator accepts that vendors will not always open their models to inspection. Where information is thin, firms are expected to compensate with extra testing, closer human supervision or independent reviews of the supplier’s controls, such as external certifications, rather than relying on a vendor’s assurances.

Contracts are part of the answer. MAS wants institutions to secure terms on performance, data protection and audit rights, and to be told when a vendor introduces or changes AI in a product. It also asks firms to watch for over-reliance on a handful of AI providers.

If a firm still cannot bring the risk of an outside AI service within its limits, MAS says it should consider restricting, suspending or replacing it.

That shifts part of the burden onto suppliers. Vendors selling to Singapore’s financial institutions can expect requests for test results, explanations of how their models behave and advance notice of updates. Those that can provide them are more likely to win contracts.

A high-performing product will not be enough on its own. A bank (or any financial organisation, for that matter) buying it must show MAS that it controls how the product works within its business.

Boards to Take Charge

The guidelines extend AI responsibility to boards and senior management, beyond the teams that deploy it. Boards and senior management are expected to decide who is accountable, set how much AI risk the firm will accept and ensure governance keeps pace as use grows.

Firms do not need a dedicated AI committee and can use existing structures. But those structures must cover AI across the whole institution, so that individual business units follow consistent risk assessments when adopting systems.

Each firm should keep an inventory of its AI use and assess how much each application matters. That means weighing its effect on customers, operations and regulatory compliance, as well as how far the system can act on its own.

The assessment sets the level of control. A tool that helps staff draft email messages or summarise internal documents may need only basic safeguards if its failure would cause little harm. A model used for credit decisions, insurance underwriting or financial advice warrants more, because errors or bias could affect whether customers get access to finance.

For high-risk uses, MAS expects independent validation before launch, continued monitoring and periodic review. It also asks firms to consider switches that can quickly shut down such a system.

The guidelines also address unfair outcomes from biased data, building on the Fairness, Ethics, Accountability and Transparency principles MAS introduced in 2018.

The obligations arrive in two stages. Board oversight, AI inventories and risk assessments under Sections 3 and 4 apply from October 7, 2027. Life-cycle controls and the skills needed to run them, under Sections 5 and 6, follow by October 7, 2028.

What Businesses Should Do Now

Map AI use	Identify AI use across the firm, including features inside vendor software, and build an appropriately detailed inventory before October 7, 2027.
Assign ownership	Name owners at board and senior management level and set the firm’s appetite for AI risk.
Review vendor contracts	Set the standard the AI must meet, protect customer data, secure the right to inspect the vendor’s systems and require advance notice before a vendor adds or changes AI in a product.
Rate and test	Assess each use by risk and line up independent validation for high-risk models.
Prepare for agents	Test pre-execution checks through SAFR pilots and prepare input for MAS’s 2027 consultation.
Build skills	Train staff to run the life-cycle controls due by October 7, 2028.

Sources: MAS Guidelines on AI Risk Management (October 7, 2026); MAS media release on SAFR (July 3, 2026)

When AI Agents Begin Moving Money

Chia’s October 9 speech dealt with a different kind of system: agents that act. Such software could compare products for a customer, find better terms and complete the transaction without the customer approving each step.

MAS has warned that wider use of these agents could change how institutions manage liquidity, move collateral and apply risk controls, particularly as agents begin dealing in tokenised assets and programmable digital money. Banks may eventually compete for deposits and customers through software that keeps searching for better deals.

MAS and industry partners published Safeguards for Agentic Finance at Runtime (SAFR) in July 2026 to prepare for that shift. The framework places checks at the moment an agent proposes an action, in addition to the instructions built into it during development.

SAFR covers three areas. It confirms an agent’s identity and what it is authorised to do, checks each proposed action against the customer’s instructions before it runs and keeps records that can be audited afterwards.

At the summit, Chia said Version 1.1 of the SAFR white paper had been published, along with open-source code that institutions can test and adapt. The Future of Finance Institute, which MAS announced in June 2026, will support adoption through industry pilots and sandbox testing.

SAFR is not binding. It is an industry framework separate from the risk guidelines that already apply to agentic AI. MAS has told firms to review their controls as they use more of these systems, and it plans to consult the industry in 2027 on whether further guidance is needed.

The two tracks reflect current adoption. Institutions already use AI in established functions such as fraud detection and customer service, while the infrastructure that would let agents execute transactions on their own is still being built.

The Cost of Trust

For financial institutions, the rules sharpen a commercial trade-off. AI can cut operating costs, improve customer service and strengthen fraud and financial crime detection. But more sophisticated systems also create new dependencies and hard-to-observe risks, and they require investment in testing, monitoring and specialist oversight.

MAS describes the rules as support for responsible adoption. It is still running a proof-of-value exercise with five banks, the police and the government’s technology agency, using historical transaction data to test whether AI models can detect higher-risk accounts and transactions sooner.

In practice, the cost is more work. AI will have to be managed like credit or market risk, with named owners, set limits and reporting lines that run to the board.

The harder questions lie ahead. When AI only recommends a product, the main risk is bad advice. When it completes the transaction itself, a bank must be able to say who authorised the payment, whether the agent followed the customer’s instructions and who bears the loss if it goes wrong.

Those questions will sharpen as Singapore builds out tokenised markets. In the same speech, Chia said MAS is preparing to settle tokenised MAS Bills, the short-term securities it issues, using wholesale central bank digital currency.

In essence, MAS is setting rules for today’s AI tools, with the next stage in view: a financial system where software agents transact directly on behalf of customers and institutions.

The next markers fall in 2027, when MAS consults on guidance for agentic AI and the guidelines take effect in October. By then, institutions will need to show who controls each AI system they run and who is accountable when it acts for a customer.

Also read:

Singapore Seeks UN Framework to Put Global Brakes on AI, Without Slowing the Race

Wall Street’s Voice-Phishing Wave Tests the Trust Question. Will It Impact MAS Guardrails?

Monetary Authority of Singapore MD On Fostering Responsible AI Adoption