MAS Issues Guidelines on AI Risk Management for FIs, Effective from 7 October 2027

Executive Summary

The use of artificial intelligence (“AI“) by financial institutions (“FIs“) can present a multitude of risks, including financial, operational, reputational, security, privacy, third-party, conduct, and other risks. These risks are amplified when generative AI (“Gen AI“) is used due to its greater complexity, unstructured nature, opacity of training data sources, and difficulties in evaluating bias, and when AI agents are granted greater autonomy and access to tools.

As such, the Monetary Authority of Singapore (“MAS“) has issued a set of Guidelines on Artificial Intelligence Risk Management for Financial Institutions (“Guidelines“), which set out its supervisory expectations relating to AI risk management in FIs, and which seek to support responsible AI adoption.

1. The Guidelines apply to FIs in a proportionate manner depending on their size and risk profile. They cover topics such as: (i) AI oversight; (ii) key AI risk management systems, policies and procedures; (iii) AI life cycle controls; and (iv) capabilities and capacity for the use of AI.

2. The Guidelines will take effect on 7 October 2027 and may be implemented in phases, i.e. FIs may meet the expectations set out in Sections 3 to 4 from 7 October 2027, and Sections 5 and 6 by 7 October 2028.

This Update summarises the key points from the Guidelines and their implications for FIs.

Scope of the Guidelines

1. FIs: The Guidelines apply to FIs, as defined in section 2 of the Financial Services and Markets Act 2022.

• • Group application: They apply on a group basis for locally incorporated FIs that are subject to consolidated supervision by MAS or are an owner of a critical information infrastructure.
  • Group guidance: FIs which are part of a global group (“Group FIs“) may take guidance from group AI risk management frameworks so long as these are compliant with the Guidelines.

2. AI: In the Guidelines, AI refers to machine-based systems or models with differing levels of autonomy and adaptiveness, which derive outputs through learned premises such as the data or inputs received. These AI outputs can include predictions, content, summaries, recommendations, or decisions that may influence physical or virtual environments. Calculators or tools whose outputs are solely based on predefined programming logic or rules would not fall within scope.

• • Gen AI and AI agents: The Guidelines set out a list of non-exhaustive, illustrative examples of what would fall within scope (including Gen AI and AI agents) and outside scope.
  • Pilots and third-party AI: They also cover AI that is partially deployed (e.g. pilots or phased rollouts) as well as third-party AI.
  • Ongoing developments: FIs should regularly review the adequacy of their AI risk management efforts against AI developments, and address new or accentuated risks.

3. Proportionate application: The Guidelines apply to all FIs in a proportionate manner, according to their size and risk profile. FIs should implement them in a manner commensurate with the size and nature of their activities, and the extent to which their AI use could lead to material risks to them.

4. Application of basic policies: FIs are permitted to apply the basic AI governance policies and procedures set out, if the poor performance or unavailability of the AI is unlikely to have a material adverse impact on: (i) the FI (including potential financial, operational, regulatory, legal or reputational impact); and (ii) its customers or other stakeholders (including potential impact in relation to fairness, ethical conduct and consumer protection).

• • For example, these criteria would be met if the AI is only used to proof-read and rephrase emails to customers, although humans should still check the outputs before using them.
  • At minimum, these basic policies should include: (i) clear accountability for AI oversight; (ii) permitted and prohibited uses of AI together with effective human oversight; (iii) maintenance of an approved list of AI tools and defined approvals processes; (iv) staff education on AI-related policies; (v) regular compliance checks; and (vi) periodic and trigger-based review.

5. Application of Sections 3 to 6: If FIs are not permitted to apply the basic policies, they should comply with the expectations set out in Sections 3 to 6 of the Guidelines.

• • Sections 3 and 4 set out the frameworks, policies and procedures for oversight of AI use that should be established. The scope of such oversight, policies and procedures may be proportionate to the size and nature of the FI’s activities, its use of AI, and its risk profile.
  • Sections 5 and 6 set out the life cycle controls, capabilities and technology infrastructure for AI use that should be applied. These may be calibrated based on relevance to the AI use case, system or model, and may be implemented based on the risk materiality of the AI use case.

AI Oversight

The board and senior management of an FI should ensure the following:

1. Effective oversight: Ensure that existing frameworks, policies, and practices for managing risk adequately identify, assess, and address AI-related risks. They should also maintain effective oversight, foster the appropriate risk culture, and ensure that the FI’s AI use does not conflict with its ability to meet regulatory requirements or other supervisory expectations.

2. Robust coordination: Ensure consistent standards, clear accountability and robust coordination to manage risks across all relevant functions. This can be done by creating a new, centralised function to manage these risks, or by managing incremental risks via existing risk management functions (e.g. establishing structures to coordinate AI risk management in a cross-functional manner, such as cross-functional committees or communication lines between functions).

The board or a committee delegated by the board bears strategic responsibilities.

1. These include: (i) approving and regularly reviewing the overall governance approach for AI risk management (e.g. strategic direction for AI use, key frameworks, etc); (ii) ensuring that the risks from AI use are explicitly addressed within the FI’s risk appetite framework (including qualitative statements and quantitative measures for such risks, as appropriate); (iii) setting clear roles and responsibilities for the board and senior management; (iv) ensuring it has an adequate understanding of AI to provide effective oversight and challenge; and (v) ensuring regular review to accommodate newer AI developments and changes in risk profile and business strategies.

2. For Group FIs: This can include the relevant regional or global committee with oversight over the FI in Singapore. For non-Group FIs, this refers to the board, or a board-level committee which is delegated by the board.

Senior management bears operational responsibilities.

1. These include: (i) effective implementation and regular review of AI-related risk management frameworks; (ii) instituting strategies and controls to manage risks across the AI lifecycle; (iii) establishing clear roles and responsibilities across different business lines and functions, and internal escalation processes; (iv) providing timely updates to the board; and (v) ensuring the necessary competence of personnel and allocating adequate resources.

2. For Group FIs: Local senior management remains accountable to ensure: (i) that group frameworks address local considerations and regulatory requirements; (ii) clear escalation paths for Singapore-specific issues; and (iii) that they have adequate visibility and input on matters affecting Singapore operations and can demonstrate how they discharged their oversight responsibilities even when relying on group frameworks.

The three lines of defence model may be applied to the roles and responsibilities if appropriate. For example:

1. First line functions could comprise model developers and AI users and may be responsible for assessing risks from AI use cases or the application of appropriate human oversight.

2. Second line functions could provide effective and independent challenge to first line functions and may be responsible for independent validation and overall compliance with applicable regulations and supervisory expectations.

3. Third line functions (i.e. internal audit) could provide independent assurance on the effectiveness of the AI risk management framework.

Key AI Risk Management Systems, Policies and Procedures

An FI’s AI risk management framework should cover key systems, policies and procedures to identify, inventorise, and assess the risk materiality of AI. FIs should apply: (i) consistent and robust approaches for each process; and (ii) controls that are proportionate to the assessed risk materiality.

1. AI identification: FIs should:

• • Establish framework: Establish systems, policies and procedures to ensure the consistent identification of AI use across all relevant business and functional areas. Clear AI definitions, criteria and processes, supported by tools and systems as appropriate, should be implemented to facilitate this. If AI is incorporated in third-party services, identification should minimally cover AI in services provided by material third-party service providers.
  • Residual risks: Ensure that residual risks from unidentified AI use remain within their risk appetites. If there are practical constraints in consistently identifying all AI use (e.g. due to inadequate disclosure of AI embedded in third-party services, or use of shadow AI), the risks arising from such constraints should be identified and mitigating measures established (e.g. clear guidelines on allowed / disallowed AI, or technical controls to detect unauthorised use).

2. AI inventory: FIs should:

• • Establish framework: Set up and maintain an accurate inventory of AI use cases, systems or models across the FI as much as possible or practicable, updated at an appropriate frequency, to support governance, oversight and risk management throughout the AI lifecycle. This can be done by: (i) enhancing existing inventories to include AI use cases, systems or models; or (ii) establishing a dedicated AI inventory. There should be clear linkages between the AI inventory and other relevant inventories to facilitate cross-referencing.
  • Regular review: Capture key attributes at an appropriate level of granularity in the AI inventory, and regularly review its design to ensure that the attributes captured account for newer AI technologies (e.g. for AI agents, there may be additional relevant attributes).

3. AI risk materiality assessment: FIs should:

• • Establish framework: Establish an assessment methodology to evaluate the risk materiality of AI use cases, based on: (i) the nature of the business; (ii) risks arising from the systems and models used; and (iii) dependencies between AI systems or models that support or interact with an AI use case where these could affect the overall risk profile of the same.
  • Risk considerations: Consider in the assessment: (i) the inherent risk materiality before risk management controls are applied; and (ii) the residual risk materiality after such application, and ensure that the latter meets the FI’s risk appetite prior to deployment.
  • Risk dimensions: Consider also the risk dimensions relevant to the FI’s context, which may evolve over time, and which should minimally cover: (i) impact (including the potential consequences of failure, malfunction or poor performance of the AI on the FI, its customers or other stakeholders, and the nature and sensitivity of the data processed by the AI); (ii) complexity (including the nature and novelty of the AI technology used, the data used, the explainability of outputs, and the level of visibility where third-party AI is concerned); and (iii) the level of reliance on AI (including third-party AI) for the associated decision or output (including the degree of AI autonomy and human involvement or oversight).

FIs should assign clear roles and responsibilities and designate a control function to be responsible for: (i) AI identification systems and processes; (ii) establishing AI inventorisation policies and procedures; and (iii) establishing risk materiality assessment frameworks and arbitrating or approving the risk materiality of AI use cases. While operational aspects may be carried out by the relevant business or functional units, there are aspects that should always remain with the designated control function (i.e. retaining sufficient and independent oversight, being the final arbiter or approver, and ensuring clear documentation and regular reviews and updates).

AI Life Cycle Controls

FIs should plan for, implement and regularly review robust controls covering the entire life cycle of an AI use case, system or model, and assign clear roles and responsibilities for such controls across different business and functional units.

Such controls may be implemented in a proportionate manner, so long as they are appropriate to the assessed risk materiality of the AI use case. For example, pilots or phased rollouts may require adjustments to lifecycle controls designed for full deployments, and FIs should establish clear policies and procedures to govern any deviations from these standard lifecycle controls.

For high-risk AI use cases, FIs should also develop and implement, and regularly review and test, their contingency plans outlining fallback options (e.g. alternative systems or manual processes) to ensure business continuity in case of AI failure or unexpected behaviour. Especially for AI with “kill switches”, clear contingency activation protocols should be in place and tested regularly.

1. Data management: FIs should establish data management controls to ensure that the data used across the AI life cycle is fit for purpose, sufficiently representative (i.e. encompasses data used for training and testing the AI use case, system or model, across a range of real-world conditions, including stressed conditions), of high quality, and subject to robust data governance, including data classification, data security, data privacy, and auditability. 

2. Transparency and explainability: FIs should establish controls to ensure an appropriate level of: (i) transparency (i.e. disclosures and explanations to individuals or groups relating to AI use that affects them); and (ii) explainability (i.e. methods for facilitating understanding of AI-generated outputs or decisions), for an AI use case. The degree and manner of providing transparency and explainability should reflect the context and nature of the AI use case, its assessed risk materiality, and the needs of the intended audience (e.g. customers, internal users, other stakeholders). 

3. Fairness: FIs should define what they consider to be “fair” outcomes and may take reference from relevant frameworks and principles. They should also have appropriate controls to identify and mitigate harmful biases and discriminatory outcomes across the AI life cycle, which should be calibrated to the context, nature and assessed risk materiality of the AI use case. Fairness assessments should be conducted where fairness considerations are relevant. 

4. Human oversight: FIs should establish and regularly review controls to ensure appropriate human oversight over an AI use case, system or model across its life cycle. The need for and degree of oversight required should take into account the objective of using AI, automation bias and decision fatigue, and be proportionate to the assessed risk materiality. Key areas for consideration include:

• • Roles and responsibilities: Clearly allocating of roles and responsibilities for human oversight, including processes for escalation and decision-making.
  • Capabilities: Enabling competent personnel who are assigned to monitor AI use with the requisite capabilities, including the necessary authority and ability for intervention.
  • Design: Designing and developing AI to enable and facilitate appropriate human oversight, including escalation mechanisms for human review should pre-defined conditions be satisfied.
  • Documentation and review: Establishing processes to document logs and other records generated throughout lifecycle activities, and regularly reviewing human oversight decisions and interventions to assess their effectiveness.

5. Third-party AI management: The decision to onboard and use third-party AI is ultimately the FI’s decision. An FI retains primary accountability for such use, including the possible impact on its stakeholders and on its ability to comply with regulatory requirements. Thus, an FI should ensure that: (i) its onboarding, development and deployment controls for third-party AI are adequate for the risk materiality of the use case; and (ii) the residual risk materiality is within its risk appetite, failing which, it should consider limiting or suspending the service or replacing the provider. Third-party AI should also be tested in the context of the FI’s own use cases and data. If there are testing data limitations, these may be mitigated by conducting pilots or phased rollouts to obtain more representative data. Key areas for consideration include:

• • Third-party AI transparency: FIs may review disclosures or available documentation from providers. If there are transparency limitations in the same, FIs may review certifications or external assessments of the providers by independent and competent parties that cover the key risks from AI use, or implement compensating controls.
  • Supply chain assessments: The extent to which key third-party and open-source AI models, datasets, and dependencies have undergone supply chain assessments and validations.
  • Concentration risks: The extent to which the use of third-party AI may contribute to concentration risks and over-reliance on key providers.
  • Change management: The extent to which the FI is notified of post-deployment changes to third-party AI, and is able to manage and assess the impact of the same.
  • Contingency plans: Whether robust contingency plans are present to address potential failures or unexpected behaviour of third-party AI, or discontinued support from providers.
  • Legal agreements: Whether agreements set out clear expectations and responsibilities of the provider, e.g. clauses on performance guarantees, data protection, audit rights, notifications when AI is introduced and updated, or seeking the FI’s agreement before incorporating AI.
  • Capabilities: The level of awareness and capabilities of the staff involved in procuring, developing, deploying and using third-party AI.
  • Complexity: The need to conduct more detailed assessments for complex third-party AI services that the FI may have less experience with.

6. Selection: When selecting AI systems, models or features to use, an FI should: (i) consider the objectives and risks of the AI use case; (ii) document its considerations in a manner commensurate with the associated risk materiality, noting that it may refer to research and accepted industry practice, and should consider incorporating reviews by domain experts or users; and (iii) carefully weigh the benefits of deploying more complex systems or models which are less understood, against new or heightened risks and its risk mitigation capabilities. 

7. Evaluation and testing: FIs should conduct evaluation and testing proportionate to the risk materiality of the AI use case. Assessments should take into account the context of AI use, take reference from best practices, and assess performance under a range of plausible conditions. FIs should also establish appropriate controls and guardrails prior to deployment, to mitigate the risks and limitations identified during development. For example, for Gen AI and AI agents, evaluation and testing should cover their key failure modes and effectiveness of associated guardrails. Key areas for consideration include:

• • Evaluation measures: Define appropriate evaluation measures aligned with the AI’s objectives and establish acceptable performance thresholds.
  • Testing approaches: Employ relevant testing methods and representative testing datasets.
  • Overfitting mitigation: If possible, especially for more complex AI, implement techniques to prevent overfitting.

8. Technology and cybersecurity risks: FIs should ensure that the AI system is secure, well-governed, and underpinned by appropriate controls to manage such risks, commensurate with the risk materiality of the AI use cases supported by the system. Key areas for consideration include:

• • Security: Deployment should occur in secured Information Technology environments.
  • Access controls: These include tightly-controlled access to AI components and infrastructure during deployment, role-based access, multi-factor authentication, robust privileged account access management, and separation of duties.
  • Third-party: If third-party components or services are used, controls should be implemented to govern use, restrict data exposure and monitor for security, compliance or operational risks.

9. Reproducibility and auditability: FIs should document the AI development process to enable reproducibility and auditability. The extent and rigour of documentation should: (i) be commensurate with the risk materiality of the AI use case; (ii) be sufficiently detailed for an independent party to understand and replicate the implementation of the AI system or model and its evaluation and testing process; and (iii) cover the entire development process. 

10. Pre-deployment review:

• • Independent review: Prior to deployment, FIs should subject the AI use case, including its underlying systems or models, to reviews by parties not involved in its development, to ensure that the relevant controls have been adhered to. The scope of review and degree of independence should be proportionate to the assessed risk materiality of the AI use case. Those assessed with high risk materiality should undergo formal independent validation before deployment, which should: (i) be conducted by independent and competent persons; (ii) provide effective challenge to developers; and (iii) cover areas relevant to the AI use case. For other AI use cases, other forms of documented review on key aspects of development and deployment may be undertaken. Review findings, including identified limitations, necessary remediation actions, or conditions for use, should be reported to the relevant approval body and actioned appropriately.
  • Technology and cybersecurity review: These should be conducted to ensure the AI can be deployed into the production environment in a controlled and secure manner. Key areas for consideration include secure design and access controls, and testing and threat mitigation.

11. Post-deployment monitoring and review: FIs should: (i) develop and implement comprehensive and robust controls for the ongoing monitoring of all deployed AI, including third-party AI; and (ii) conduct re-validations or reviews if necessary to ensure that deployed AI continues to be used appropriately and associated risks remain appropriately managed. The frequency and intensity of monitoring, and the need, frequency and rigour of re-validations or reviews, should be proportionate to the assessed risk materiality. AI use cases assessed as having high risk materiality should be subject to regular re-validations by independent parties. Key areas for consideration include:

• • Monitoring measures: Based on the risks of each AI use case, define key metrics for monitoring and acceptable performance thresholds for each such metric.
  • AI incident and issue management: Set up robust processes for reporting, tracking, escalating, and resolving issues or incidents, if breaches or anomalies arise. For AI of high risk materiality, consider implementing “kill switches” or override mechanisms to rapidly deactivate the AI system should it exceed risk tolerances.
  • Roles and responsibilities: For deployed AI, allocate clear roles and responsibilities for ongoing monitoring and incident management.
  • Documentation: Keep clear records of monitoring activities, results, issues or incidents, and subsequent remediation actions, for auditability and ongoing risk management. Consider enhanced documentation where relevant for Gen AI or AI agents.
  • Training and awareness: Equip those responsible for monitoring with appropriate training for effective monitoring and incident management, and end-users with training and awareness to report unintended behaviour of deployed AI.

12. Change management: FIs should develop and implement: (i) comprehensive, robust controls for managing changes to deployed AI; and (ii) clear controls for retirement or decommissioning if the AI is no longer needed or exceeds risk tolerances. Key areas for consideration include:

• • Significant changes: Set up a framework for assessing the importance of changes, based on the nature of the change and assessed risk materiality of the AI use case. Significant changes should trigger appropriate reviews and re-approvals prior to implementation. Compensating controls should be implemented for high materiality use cases where significant changes may occur without prior review.
  • Change control: Implement change control mechanisms with appropriate human oversight to prevent unauthorised changes to AI systems or models, and to ensure that changes to AI do not cause unintended results.
  • Enhanced controls: Implement these for AI that are designed to be automatically updated.

Capabilities and Capacities for the Use of AI

1. Personnel: FIs should: (i) determine and ensure the necessary competence and proper conduct of personnel involved in developing, deploying and using AI; and (ii) conduct regular reviews to ensure that they are equipped with adequate capabilities and capacity for effective AI risk management. Adequate human, technological and financial resources should be allocated proportionate to the risk profile of the AI use case and the underlying systems or models.

2. Infrastructure: An FI should ensure that its technology infrastructure is adequate for the AI use cases, systems or models it wants to implement, including that underlying hardware and software resources are sufficient to meet its needs, including performance, scalability, and resilience.

Key Insights

The Guidelines are significant in clarifying at least three aspects of AI risk management for FIs. First, that a new, centralised function need not be established to manage AI risks, and that existing risk management functions may do so in a cross-functional manner. Second, that FIs continue to bear primary accountability for their use of third-party AI, and the steps expected for FIs to address the risks associated with the same. Third, that the basic AI governance policies and procedures may be applied if the poor performance or unavailability of the AI is unlikely to have a material adverse impact on the FI, its customers or other stakeholders, and the minimum content of such basic policies and procedures.

FIs are encouraged to embark on early implementation of the Guidelines ahead of the phased implementation timelines stipulated, and we are able to advise on these aspects.

If you have any queries on the above, please reach out to our team set out on this page.

For regional Financial Services Regulatory and Technology, Media & Telecommunications matters, please see Rajah & Tann Asia’s Financial Services Regulatory Practice and Technology, Media & Telecommunications Practice for more information.