Singapore Sets AI Risk Rules for Financial Firms, With Deadlines in 2027 and 2028

Two deadlines put AI oversight on the financial sector agenda

Singapore's financial institutions have until 7 October 2027 to meet core supervisory expectations for artificial intelligence oversight, followed by a second deadline on 7 October 2028 for broader controls, capabilities and infrastructure. The Monetary Authority of Singapore (MAS) published its Guidelines on Artificial Intelligence Risk Management on 7 October 2026, covering all financial institutions and all forms of AI technology.

The framework reaches beyond chatbots and software that helps staff draft documents. It covers traditional machine learning, generative AI that produces content, and AI agents capable of making decisions or taking actions through connected tools. MAS warns that agents with tool access could autonomously carry out erroneous or unauthorised actions, making the authority to act as important as the accuracy of an answer.

The regulator's AI risk guidelines establish supervisory expectations rather than a separate law prescribing identical technical requirements for every firm. Their central principle is proportionality: institutions should identify their AI use, assess the importance and risks of each application, and put controls in place that match that exposure.

This creates a broad regulatory boundary without treating every application alike. A system influencing investment recommendations or customer risk assessments warrants a different level of scrutiny from a tool used to proofread an internal email. The distinction depends on potential consequences, including harm to customers and other institutions, rather than simply whether a product carries an AI label.

Who falls within the framework?

The scope includes banks, fund managers, financial advisers, insurers, payment institutions, exchanges, clearing houses, custodians, credit rating agencies and firms dealing in capital markets products. Exempt futures and over the counter derivatives brokers are also explicitly included.

Coverage extends across AI technologies, so the framework is neither a generative AI rulebook alone nor a measure aimed only at the largest banks. MAS allows institutions to tailor implementation to their risk profiles, including the scale and nature of their AI use.

The guidelines also complement existing MAS requirements and principles, including Fairness, Ethics, Accountability and Transparency, known as FEAT. They add an AI risk management framework rather than replacing the wider obligations that already govern financial services.

For institutions deciding how much work is required, the key distinction is between broad coverage and the intensity of controls. Being covered does not mean every firm must build the same governance structure. It does mean each institution needs a defensible understanding of where it uses AI and what could happen if those systems perform poorly or become unavailable.

What changes between now and October 2028?

The publication follows a consultation launched in November 2025. In an August 2026 parliamentary reply, MAS said the proposed guidelines would be finalised soon and confirmed that they covered agentic AI. The October announcement then set out two implementation stages.

• November 2025: MAS launched its public consultation on the proposed AI risk management guidelines.
• 5 August 2026: A parliamentary reply confirmed that the proposals covered all AI use cases by financial institutions, including agentic AI.
• 7 October 2026: MAS published the final guidelines.
• 7 October 2027: The guidelines take effect, with expectations in Sections 3 and 4 applying from this date.
• 7 October 2028: Institutions should meet the expectations in Sections 5 and 6.

The schedule gives firms one year from publication to establish the core oversight, risk management and policy framework, followed by another year for the broader operational arrangements. The later deadline does not postpone the entire framework until 2028.

A brief report describes the effective date as 7 October without specifying the year. MAS's publication and announcement are explicit: the effective date is 7 October 2027, with phased implementation continuing through 2028.

Boards retain responsibility without needing a new committee

MAS expects boards and senior management to provide effective oversight of AI risks. That includes defining responsibilities, setting risk appetite and establishing risk management frameworks, policies and procedures. Risk appetite means the level and types of risk an institution is prepared to accept.

The regulator's publication announcement addresses a practical concern raised during consultation: firms may use existing governance structures if those structures provide adequate oversight and coordination across relevant functions. A dedicated AI committee is not required solely to meet this expectation.

The distinction matters for implementation. The requirement is effective accountability, not a particular committee name or organisational chart. Existing arrangements can meet the expectation, but only if they can oversee AI risks and coordinate the people responsible for them.

Institutions must manage those risks at two levels: across the organisation and within individual applications. An application may appear manageable in isolation while the firm's expanding use of AI demands broader oversight, skills or infrastructure. MAS therefore expects institutions to build their capabilities as adoption grows.

Inventories make the risk assessment concrete

Financial institutions should identify their AI use and maintain inventories at an appropriate level of detail. An inventory provides the basis for assessing which applications matter most and deciding what controls they need. Without that foundation, senior management cannot reliably determine the institution's exposure.

MAS also expects an assessment of each application's risk materiality. In plain language, this means judging whether failures, poor performance or loss of availability could have meaningful consequences for the institution, its customers or other stakeholders, including other financial institutions.

The framework identifies financial, operational, conduct, financial crime and reputational risks. Generative AI also presents security, privacy and risks associated with external providers. These categories make the assessment wider than a check on whether a model usually produces accurate answers.

MAS permits basic policies and procedures where poor performance or unavailability of AI services is unlikely to have a material impact. More sophisticated controls should follow greater exposure. The practical test is the potential effect of failure, not the novelty of the technology.

Controls must continue after an AI system is introduced

The guidelines call for proportionate controls throughout the AI life cycle, meaning the stages from development and testing through deployment, use and subsequent changes. MAS names data governance, testing, human oversight, cybersecurity, monitoring and change management among those controls.

Data governance concerns how information is managed and used. Testing examines whether a system is suitable for its intended purpose. Monitoring checks performance during operation, while change management addresses modifications that could alter behaviour or risk. These are connected responsibilities rather than a single approval exercise before launch.

Human oversight is also part of the framework. For an AI system that can influence or execute a financial decision, the issue is not merely whether a person can read its output. Institutions must consider how oversight fits the risks of the application and the actions it can take.

MAS expects controls to be reviewed regularly as AI use expands and technologies evolve. An arrangement designed for a tool that drafts text may need reassessment if that tool later gains access to systems through which it can take action. The guidelines specifically connect regular reviews to the increasing use of autonomous agents.

Buying AI does not transfer accountability

Financial institutions remain accountable for AI used in the services they deliver, including systems developed, operated or provided by external companies. Outsourcing the technology does not outsource responsibility for the risks it creates.

MAS expects firms to obtain sufficient assurance from providers and assess whether their AI is suitable for the intended use. Where practical constraints or gaps in assurance arise, institutions should apply compensating controls. These are measures intended to reduce risks when the preferred assurance or control is unavailable.

The regulator also sets out a response when those measures are insufficient. If the risks cannot be brought within an institution's risk appetite, it should consider limiting, suspending or replacing the AI service. Continued use is therefore not justified simply because a provider is commercially attractive or the system is already embedded in operations.

This addresses another consultation concern: AI embedded in products and services. The relevant question is how the institution uses that functionality and manages its consequences, rather than whether it built the model itself. The assessment must connect the provider's assurances to the firm's actual intended use.

Agentic AI is covered before further guidance arrives

Agentic AI can do more than generate a response. It may use tools to pursue tasks and act with greater autonomy. In financial services, that makes permissions, intervention and records of decisions central to managing the risk of erroneous or unauthorised actions.

In the August parliamentary reply, Gan Kim Yong, Deputy Prime Minister, Minister for Trade and Industry and Chairman of MAS, confirmed that the proposed guidelines applied to all AI use cases by financial institutions, including agentic AI. The reply answered a question from Mariam Jaafar, Member of Parliament for Sembawang GRC, about autonomous agents and possible supervisory requirements.

Gan also described practical resources developed with industry. Under Project MindForge, the industry developed an AI Risk Management Toolkit to help institutions implement the guidelines. The Safeguards for Agentic Finance at Runtime framework, known as SAFR, sets out a potential approach to authorising agent actions, activating human oversight and recording what happens at each consequential decision.

Those resources serve a different purpose from the supervisory framework. The toolkit supports implementation, while SAFR describes a potential operational approach. The parliamentary reply did not announce that SAFR itself would become mandatory.

MAS intends to consult the financial sector in 2027 on what additional guidance for agentic AI would be useful. It has not given a specific consultation date or announced the content of any resulting guidance. Agents are already within the guidelines' scope, so that planned consultation is not a reason to postpone managing their risks.

Regulatory clarity, with limits on acceptable risk

MAS presents the framework as support for responsible adoption rather than a ban on AI. Consultation respondents supported the approach based on principles and proportionate controls, while seeking clarification about existing governance structures, embedded AI and when basic policies would suffice. The final framework retains the main expectations while addressing those concerns.

Ho Hern Shin, Deputy Managing Director of MAS, connected the potential benefits of AI to the need for institutions to understand increasingly capable systems. She said:

Realising these benefits sustainably requires financial institutions to understand and manage the risks that come with increasingly capable AI systems.

Her statement identifies the balance at the centre of the guidelines. Institutions can pursue productivity gains, stronger risk management and new services, but they must remain accountable for the systems used to deliver them.

The next fixed milestones are the October 2027 and October 2028 implementation deadlines. The additional agentic AI consultation is planned for 2027, while MAS has also said it will continue reviewing and updating supervisory expectations where necessary. No separate deadline for new agentic AI requirements has been announced.

Key Points

• MAS published its AI risk management guidelines on 7 October 2026, covering all financial institutions and AI technologies.
• Core expectations in Sections 3 and 4 apply from 7 October 2027; Sections 5 and 6 should be implemented by 7 October 2028.
• Boards and senior management must provide effective oversight, but a dedicated AI committee is not required.
• Institutions should inventory AI use, assess materiality and apply controls proportionate to potential harm.
• Firms remain accountable for AI supplied by external providers and should consider restricting or replacing services with unacceptable risks.
• Agentic AI is already covered. MAS plans a further consultation on possible additional guidance in 2027.