Artificial intelligence (AI) technologies are being adopted in financial services at increasing speed and scale, ranging from the use of internal productivity tools to the performance of core functions such as risk management, suspicious transaction monitoring, fraud detection, credit and lending assessments, and customer service. While AI can deliver benefits such as increased efficiency and improved performance, its complexity and probabilistic nature also introduce the possibility of biased or unexpected behaviour, which can exacerbate existing financial, operational, conduct, and other risks.
Many financial regulators globally have therefore highlighted the need for firms to manage AI risks appropriately, particularly in the context of the use of Generative AI tools and AI agents. In this update, we focus on recent key developments in Singapore and Hong Kong.
Singapore and Hong Kong are leading financial centres in Asia which have taken clear steps to regulate the use of AI in the financial sector. The principal distinction between the two jurisdictions’ approaches is one of regulatory architecture. Singapore has introduced a consolidated framework for financial institutions, whereas in Hong Kong the requirements and supervisory expectations must currently be assembled regulator by regulator and, in some cases, use case by use case. In both jurisdictions, however, regulated institutions and their board and senior management remain responsible for their use of AI, and regulators are increasingly focused on ensuring that governance and controls keep pace as AI applications become more material, customer-facing or autonomous. The regulatory frameworks in both jurisdictions operate alongside their respective generally applicable personal data protection regimes.
Singapore
On 7 October 2026, the Monetary Authority of Singapore (MAS) published its Guidelines on Artificial Intelligence Risk Management (the Guidelines) for financial institutions (FIs) in Singapore, which set out MAS’ supervisory expectations in four areas: AI oversight; key AI risk management systems, policies, and procedures; AI life cycle controls; and capabilities and capacity for the use of AI. The Guidelines complement existing legislation and guidelines issued by MAS, and MAS has highlighted that, for example, existing principles on Fairness, Ethics, Accountability and Transparency (FEAT) continue to apply in guiding the use of AI in the financial sector.
The Guidelines take effect on 7 October 2027. FIs should meet the expectations on AI oversight and key AI risk management systems, policies, and procedures from 7 October 2027, and those on AI life cycle controls and capabilities and capacity by 7 October 2028. FIs that are part of a global group may leverage their group’s AI risk management frameworks provided such frameworks meet the expectations in the Guidelines.
Application of the Guidelines
The Guidelines apply to all FIs in a proportionate and risk-based manner, according to their size and risk profile. Where the use of AI is unlikely to have a material adverse impact on the FI, its customers or other stakeholders (e.g., in the event of poor performance or unavailability), such as drafting emails or summarising documents for internal reference (low-impact use), the FI may apply basic AI governance policies only. Such basic policies and procedures would minimally include:
- clear accountability for AI oversight;
- permitted and prohibited uses of AI, together with effective human oversight;
- maintenance of an approved list of AI tools, and a defined process for requesting approval of new AI tools;
- staff education on the FI’s AI-related policies;
- regular compliance checks; and
- periodic and trigger-based review of whether the relevant use cases continue to constitute low-impact use.
Key expectations
In all other situations involving the use of AI, an FI is expected to establish frameworks, policies and procedures to oversee its use of AI, including clear identification of AI use cases, robust risk materiality assessments and an adequate AI inventory; and apply relevant life cycle controls, and establish adequate capabilities and technology infrastructure for the use of AI.
-
Oversight by the board of directors and senior management: The board and senior management should ensure that existing risk management frameworks, policies, and practices across the organisation adequately identify, assess and address risks posed by AI. There should also be consistent standards, clear accountability, and robust coordination across the FI to manage risks. Key responsibilities include the board approving and regularly reviewing the overall governance approach for AI risk management and ensuring that risks from AI use are explicitly addressed within the FI’s risk appetite framework, and senior management establishing an internal escalation process for managing material risks and exceptions from AI use, such as AI incidents or breaches of risk thresholds, and ensuring appropriate and timely actions are taken.
-
Key AI risk management systems, policies, and procedures – identification, inventory and risk materiality assessment: There should be systems, policies and procedures in place to ensure the consistent identification of AI use across all relevant business and functional areas, including by having clear AI definitions and criteria, and a designated control function responsible for the identification process and documentation of such process. In the event of practical constraints in consistently identifying all AI use (e.g., due to inadequate disclosure of AI embedded in third-party services or the use of shadow AI), the FI should identify the risks arising from such constraints and put in place mitigating measures (e.g., uplifting technical controls).
The FI should establish and maintain an accurate inventory of AI use cases, systems and models across the FI to the extent possible or practicable, including having policies establishing the frequency of inventory updates – such policies and procedures should be established by a designated control function. The inventory should capture key attributes (e.g., model type, data used, approved scope of use, dependencies) at an appropriate level of granularity to enable effective governance, oversight and risk management, and its design should be regularly reviewed to ensure that attributes captured take into account newer AI technologies.
An FI should establish an assessment methodology to evaluate the risk materiality of an AI use case based on the nature of its business and the risks arising from the systems and models used – including inherent risk, and residual risk after risk management controls are applied. The assessment should consider various risk dimensions relevant to the FI’s context, minimally covering impact (e.g., in the event of malfunction or poor performance), complexity (arising from the nature of the AI technology used), and reliance (e.g., the extent of autonomy granted to the AI system or model and the degree of human involvement or oversight). A control function should be designated to establish the assessment framework, and to arbitrate or approve the risk materiality of an AI use case.
Where a designated control function is required as noted above, various aspects of the processes (e.g., maintenance of and updates to the AI inventory) may be carried out by the relevant business or functional units. However, the designated control function should retain sufficient and independent oversight. -
Life cycle controls: An FI should plan for and implement robust risk-based controls covering the entire life cycle of an AI use case, system or model (i.e., the evolution from inception to retirement or decommissioning) and assign clear roles and responsibilities for such controls across different business and functional units. Where the AI use case is assessed as high risk, an FI should develop and implement contingency plans outlining fallback options, such as alternative systems or manual processes, to ensure business continuity in the event of AI failure or unexpected behaviour. For AI with “kill switches”, clear contingency activation protocols should be in place and tested regularly.
The expected controls include the following areas:- Data management: To ensure data used across the AI life cycle is fit for purpose and representative, of high quality, and subject to robust data governance. In ensuring that existing data governance controls are adequate to address risks, key areas the FI should consider include data quality, classification, security, and privacy.
- Transparency and explainability: The degree of transparency and explainability required should reflect the context and nature of the AI use case and its assessed risk materiality. For example, AI that is heavily relied upon for credit decisioning, insurance underwriting or other regulated activities with high impact on customer outcomes will require a higher level of transparency and explainability, and the FI should pay greater attention to considerations such as the different features or attributes in data used as inputs and justifications for their use, the ability for users to identify key drivers of the output, and the channels for redress.
- Fairness: An FI should define what it considers “fair” outcomes and have appropriate controls to identify and mitigate harmful biases and discriminatory outcomes across the AI life cycle. Controls should be calibrated to the context and nature of the AI use case and its assessed risk materiality.
- Human oversight: An FI should put in place and regularly review controls to ensure appropriate human oversight over an AI use case, system or model across its life cycle. Key areas for consideration include clear roles and responsibilities, the relevant personnel’s capabilities (including the authority and ability to intervene), design or development of the AI system or model from the outset to enable and facilitate appropriate human oversight (e.g., mechanisms to escalate AI outputs or recommendations for human review where pre-defined conditions are met), and documentation and review.
- Third-party AI management: An FI retains primary accountability for its use of AI developed, owned, operated or provided by external parties. The onboarding, development and deployment controls for third-party AI should be adequate for the risk materiality of the use case, including ensuring that contractual agreements allow the FI to have a risk-appropriate degree of visibility over the introduction of AI and subsequent updates or changes, to facilitate the application of relevant controls (including compensating controls where there are practical constraints or assurance gaps). Relevant considerations include AI provider transparency, supply chain assessments, concentration risks, change management, contingency plans, and legal agreements (setting out clear expectations and responsibilities on the AI provider, such as performance guarantees, data protection, audit rights, and notification when AI is introduced or updated).
objectives and risks of the AI use case; evaluation and testing proportionate to risk materiality; controls to manage technology and cybersecurity risks; documentation of the AI development process to enable reproducibility and auditability; pre-deployment reviews by parties not involved in the development (with formal independent validation for high risk materiality use cases prior to deployment); post-deployment monitoring and review, including defining key metrics to be monitored with acceptable performance thresholds for each metric; and change management.
-
Capability and capacity: An FI should determine and ensure the necessary competence and proper conduct of personnel involved in developing, deploying and using AI, including recruiting the necessary talent and providing appropriate training and support to employees, and conducting regular reviews.
The FI should also ensure that its technology infrastructure is adequate for the AI use cases, systems or models it wants to implement. The underlying hardware and software resources, such as graphics processing units, network infrastructure, and secure data pipelines should be sufficient to meet performance, scalability, and resilience needs.
FIs should carefully review the Guidelines and assess which uses of AI would require only the basic governance policies and procedures, and which would be subject to the full framework, before taking steps to comply with MAS’ supervisory expectations by the stipulated timelines. Key steps include mapping all current AI use, including AI embedded in third-party services, and setting up an AI inventory; updating board terms of reference and the risk appetite framework to cover AI risk; and reviewing outsourcing and vendor contracts for AI-related transparency, notification, and audit rights.
MAS has said that the Guidelines will be updated or augmented as necessary, as AI adoption and its associated risks continue to evolve alongside advancements in AI technologies. MAS has also indicated that it will further consult the financial sector in 2027 on additional guidance on agentic AI. FIs should take immediate steps to review and enhance their AI risk management frameworks in view of the Guidelines, while ensuring agility and adaptability as the technology changes and supervisory expectations become clearer.
Hong Kong
Hong Kong has adopted a sector-specific approach to the regulation of AI in financial services. Unlike Singapore, it does not currently have a single consolidated set of AI risk-management guidelines applying across the financial sector. Instead, the Hong Kong Monetary Authority (HKMA), Securities and Futures Commission (SFC) and Insurance Authority (IA) have developed, or are developing, regulatory requirements and supervisory expectations within their respective remits. These are supplemented by cross-sector initiatives and Hong Kong's generally applicable personal data regime.
Although the resulting framework is fragmented, common themes are emerging across the principal regulators. These include senior management accountability, risk-based governance, controls throughout the AI lifecycle, model testing and monitoring, appropriate human oversight, personal-data protection, cybersecurity, transparency and management of third-party risks. As discussed below, the applicable requirements and the intensity of the expected controls depend on the relevant regulator and use case.
SFC-regulated entities
The SFC's principal AI-specific requirements are contained in its circular “Use of generative AI language models” dated 12 November 2024 (the SFC GenAI Circular). The SFC GenAI Circular applies to licensed corporations (LCs) offering services or functionality provided by generative AI language models (AI LMs) or AI LM-based third-party products in relation to their regulated activities. It applies irrespective of whether the AI LM is developed by the LC itself, another group company or an external service provider, or is open-source.
LCs may adopt a risk-based approach to implementing the SFC GenAI Circular, commensurate with the materiality of the impact and level of risk presented by the particular use case. In particular, the SFC regards the use of an AI LM to provide investment recommendations, advice or research to investors or clients as a high-risk use case, given that inaccurate output could lead an LC to recommend unsuitable products or misinform investors.
The SFC GenAI Circular is structured around four core principles:
- senior management responsibilities, including effective policies, procedures and internal controls and adequate oversight and governance throughout the AI model lifecycle;
- AI model risk management, including appropriate segregation of model development and validation functions, validation before deployment and following material changes, end-to-end testing and ongoing monitoring;
- cybersecurity and data risk management, including appropriate controls addressing adversarial attacks, confidential information, data quality and bias; and
- third-party provider risk management, including due diligence, ongoing monitoring, allocation of cybersecurity responsibilities and assessment of supply-chain, data leakage and operational-resilience risks.
Additional safeguards apply to high-risk use cases. These include model validation and ongoing monitoring directed at factual accuracy, human review of AI-generated output for factual accuracy before it is relayed to a user, and testing of output robustness against prompt variations. Where an AI LM is used in a client interface, the LC should prominently disclose that the client is interacting with AI rather than a human and that the output may not be accurate. LCs intending to adopt AI LMs for high-risk use cases are also reminded of applicable notification requirements regarding significant changes in the nature of their business or types of services, and are encouraged to discuss their plans with the SFC at an early stage.
The SFC GenAI Circular should be read together with the SFC's circular “Data risk management” dated 30 March 2023 (the SFC Data Risk Circular). The SFC Data Risk Circular sets out the SFC's expected standards concerning risk governance, controls and monitoring throughout the data lifecycle, covering the collection, classification, use, retention, transfer and disposal of data and the use of third-party service providers. In particular, where a service provider participates in the data lifecycle, LCs should undertake appropriate due diligence and ongoing monitoring to ensure that it is capable of safeguarding the data and complying with applicable legal and regulatory requirements.
The SFC has also addressed risks arising from the use of AI by cyber threat actors. Its circular “Enhanced cybersecurity measures to address evolving risks arising from artificial intelligence-enabled cyberattacks” dated 2 June 2026 (the SFC Cybersecurity Circular) applies to LCs, SFC-licensed virtual asset service providers and their associated entities (collectively described in that circular as licensed firms). It expects licensed firms to assess their preparedness for AI-enabled cyberattacks and enhance their cybersecurity frameworks where appropriate, addressing areas including patching and vulnerability management, access and privilege controls, detection and monitoring, third-party supply-chain risk and incident response and recovery. Senior management, including the Manager-in-Charge of Information Technology, remains ultimately responsible for managing cybersecurity risks.
Of particular relevance as firms increasingly deploy AI systems connected to other tools and workflows, the SFC Cybersecurity Circular expects licensed firms to treat external and untrusted inputs, including content retrieved from applications, emails, documents and webpages, as potentially adversarial; prevent such inputs from directly altering system instructions or triggering privileged actions; apply least-privilege principles to connectors and tool permissions; and apply maker-checker controls to high-impact actions.
HKMA-regulated entities
The HKMA has developed its supervisory expectations through a series of high-level principles and use-case-specific circulars, most recently supplemented by its “ART” Framework for responsible AI adoption.
The foundation of the regime is the HKMA's circular “High-level Principles on Artificial Intelligence” dated 1 November 2019 (the HKMA AI Principles). Authorised institutions (AIs) are expected to take the HKMA AI Principles into account when designing and adopting AI and big-data-analytics applications, with the principles capable of proportionate application having regard to the nature of the relevant AI application and the level of risk involved.
The HKMA AI Principles address, among other matters, board and senior management accountability, appropriate expertise, explainability, data quality, model validation, auditability, third-party vendor oversight, ethical and fair outcomes, transparency, ongoing monitoring, data protection, cybersecurity, and risk mitigation and contingency planning. In particular, boards and senior management remain accountable for AI-driven decisions. AIs should maintain appropriate risk-mitigating controls, including human-in-the-loop mechanisms, and contingency arrangements capable of suspending an AI application and reverting to human intervention or conventional processes where necessary.
For customer-facing generative AI (GenAI), the HKMA's circular “Consumer Protection in respect of Use of Generative Artificial Intelligence” dated 19 August 2024 (the HKMA GenAI Circular) expects AIs to apply and extend the guiding principles in the HKMA's earlier circular “Consumer Protection in respect of Use of Big Data Analytics and Artificial Intelligence by Authorized Institutions” dated 5 November 2019 to GenAI, adopting a risk-based approach commensurate with the risks involved. The HKMA GenAI Circular focuses on four areas: governance and accountability; fairness; transparency and disclosure; and data privacy and protection.
Boards and senior management remain accountable for GenAI-driven decisions and processes. AIs should clearly define the scope of customer-facing GenAI applications, establish appropriate policies, procedures and controls, and properly validate the relevant models. In particular, during the early stages of deploying customer-facing GenAI applications, the HKMA expects AIs to adopt a human-in-the-loop approach. Customers should, as far as practicable, be given the option to opt out of using GenAI and request human intervention in GenAI-generated decisions; where an opt-out cannot be provided, AIs should provide channels through which customers may request review. AIs should also disclose their use of GenAI to customers and communicate the purpose and limitations of the relevant models in an accurate and understandable manner.
The most recent development is the HKMA's announcement on 15 October 2026 of its “ART” Framework, organised around “Accountability, Resilience and Trust” (the ART Framework). As part of the ART Framework, the HKMA stated that it had issued its latest “Supervisory Guidance for Responsible Adoption of Artificial Intelligence” (the HKMA AI Supervisory Guidance), drawing on insights obtained from the second cohort of the GenA.I. Sandbox.
The HKMA has indicated that the HKMA AI Supervisory Guidance calls for clear ownership of each AI use case, supported by an up-to-date inventory of approved applications. It also recognises a progressive shift from “human-in-the-loop” to “human-on-the-loop” oversight: as AI use cases become more sophisticated, effective oversight need not involve human participation in every step, but must continue to involve clear accountability and ongoing monitoring.
The ART Framework also looks ahead to the governance of agentic AI. The HKMA has announced further testing of agentic AI applications through the cross-sector GenA.I. Sandbox++ and an upcoming discussion paper on a proposed supervisory framework for agentic AI governance. The areas identified for further consideration include intent-anchored accountability, a registry of material AI agents and effective intervention mechanisms.
The HKMA has separately issued use-case-specific requirements and guidance. In relation to anti-money laundering and counter-financing of terrorism (AML/CFT), its circular “Use of Artificial Intelligence for Monitoring of Suspicious Activities” dated 9 September 2024 required AIs with significant operations in Hong Kong to give due consideration to adopting AI in their money laundering and terrorist financing monitoring systems, undertake a feasibility study and formulate an implementation plan, with board-level sign-off.
This was followed by the circular “Supporting Adoption of Artificial Intelligence in Fighting Financial Crime” dated 22 June 2026, which reported further developments in AI-enabled AML/CFT, including use cases concerning dynamic risk assessment, facial watch lists and corporate mule detection. The HKMA stated that AIs with significant operations in Hong Kong that had submitted implementation plans should keep those plans under review and report progress and new use cases to the HKMA when called upon.
IA-regulated insurers and insurance intermediaries
The IA's AI-specific regulatory framework remains less developed than the corresponding SFC and HKMA regimes, but is evolving.
The IA has publicly described its AI strategy as having two components: enhancing regulatory clarity and promoting the development of AI use cases. Its principal developmental initiative is the AI Cohort Programme, launched in August 2025 to facilitate responsible AI development and collaboration within the insurance sector. In June 2026, the IA reported that ten insurers were participating as core participants in the programme.
More significantly from a regulatory perspective, the IA stated in April 2026 that it aimed to roll out a new guideline on insurers' AI governance frameworks later in 2026. The IA indicated that the proposed guideline would adopt a risk-based approach covering governance and accountability, data protection, cybersecurity, disclosure arrangements and operational resilience, with the objective of encouraging innovation while strengthening fairness, accountability and transparency. At the time of writing, the guideline has not yet been issued.
Existing IA requirements are nevertheless directly relevant to certain AI-related risks. In particular, the revised Guideline on Cybersecurity (GL20), issued under section 133 of the Insurance Ordinance (Cap. 41) and effective from 1 January 2025, sets the minimum cybersecurity standards expected of authorised insurers within its scope and incorporates the Cyber Resilience Assessment Framework (CRAF). Under CRAF, insurers assess their inherent cyber risk and the maturity of their cybersecurity controls against prescribed control principles.
GL20 requires, among other matters, an appropriate cybersecurity strategy and framework, governance, risk identification, assessment and control, continuous monitoring and response and recovery arrangements. These requirements assume particular importance where AI is deployed in core operational or customer-facing processes or materially changes an insurer's technological risk profile.
The IA also operates the Insurtech Sandbox, launched on 29 September 2017 for authorised insurers and later extended to licensed insurance broker companies. It allows eligible firms to run pilot trials of technology initiatives and collect enough data to show whether those initiatives broadly meet the IA’s supervisory requirements.
Separately, the HKMA, SFC, IA and Mandatory Provident Fund Schemes Authority (MPFA) issued the “Joint Circular on the Expansion of Generative Artificial Intelligence Sandbox” dated 5 March 2026 (the Joint Sandbox Circular), expanding the original banking-sector GenA.I. Sandbox into the cross-sector GenA.I. Sandbox++. The GenA.I. Sandbox++ is open to, among others, authorised insurers and licensed insurance broker companies, as well as AIs, LCs, MPF approved trustees and principal intermediaries and stored value facility licensees. It offers targeted supervisory guidance from the participant's respective regulator, together with technical support and access to computing resources provided in collaboration with Cyberport.
The first cohort of the GenA.I. Sandbox++, announced on 27 August 2026, focuses on agentic AI. Testing covers end-to-end processes including customer onboarding, payments, insurance claims and customer interactions, and examines the use of AI to dynamically oversee actions taken by other AI applications. While the initiative signals the areas the Hong Kong financial regulators are currently exploring, it does not itself impose regulatory requirements.
Insurers should therefore monitor publication of the IA's proposed AI governance guideline closely. In the meantime, AI deployments should be assessed under the regulatory requirements applicable to the underlying insurance activity, together with relevant governance, cybersecurity, outsourcing, conduct and personal data requirements.
What should financial institutions do now?
Hong Kong financial institutions should first identify the regulator-specific requirements applicable to the relevant entity and underlying regulated activity, and then calibrate their controls to the materiality, customer impact and autonomy of each AI use case.
As a practical priority, firms should:
- establish appropriate ownership and inventories of AI use cases;
- classify use cases by risk and potential customer or investor impact;
- implement proportionate validation, monitoring and human-oversight arrangements;
- review customer disclosures and mechanisms for human intervention or review;
- assess personal data, cybersecurity and operational resilience risks;
- review third-party arrangements, including dependencies and concentration risk; and
- monitor forthcoming regulatory developments, particularly the HKMA's proposed approach to agentic AI and the IA's proposed AI governance guideline.
This article is provided for general information only and is not legal advice. Please contact your usual Norton Rose Fulbright contact for advice on your specific circumstances.