The Hong Kong Monetary Authority has updated its approach to supervising artificial intelligence in banks, asking institutions to assign responsibility for each application as systems gain more freedom to act. Its 15 October announcement introduced the ART framework, standing for Accountability, Resilience and Trust. The package combines new supervisory guidance with lessons from banking trials and a preliminary note on governing autonomous AI agents. That last document opens a policy discussion, including a possible industry registry, rather than establishing the proposed registry.
At the Finance + A.I. Forum on 16 October, banking supervision executive director Carmen Chu said more than 60 per cent of applications in the second sandbox cohort involved AI agents. Half of those were multi-agent applications. Such systems can plan and carry out several connected tasks, using tools and accessing data or other systems. Chu said effective oversight need not put a person into every individual step, but must preserve accountability and continuing monitoring. The supervisory question therefore extends beyond whether a chatbot's final answer sounds correct.
Who answers for an AI system?
The updated guidance replaces principles issued on 1 November 2019. Boards should approve an AI strategy and define the institution's risk appetite. Every deployed use case needs an owner, while an inventory should include outsourced AI and AI embedded in other products. Staff should not use unapproved tools or approved tools for unapproved purposes. The guidance draws on international work, including the FSB's AI oversight practices covered by Dalimss News.
Testing should cover the complete application before deployment and significant changes, with greater scrutiny for material or higher-risk uses. For those systems, a kill switch or equivalent should allow prompt suspension and withdrawal of access. Customers affected by AI outputs should be able to seek clarification or reassessment, with human reviewers empowered to correct outcomes. Institutions should tell consumers when a service uses AI and explain relevant risks. The document calls for timely improvements to controls without specifying a calendar deadline.
The accompanying sandbox report shows why a reviewer's presence alone may be insufficient. It describes the risk of staff trusting an automated answer too readily or merely approving it without examination. Practices discussed include making the basis for an answer visible and recording interactions so reviewers can inspect what happened. One bank restricted agents to decision support, requiring human review before recommendations, customer communications or action-support outputs were finalised. These are reported pilot arrangements, not proof that every participating institution has the same controls.
The trials also examined using one AI system to check another. In one case study, adding more reviewer models in a sequence did not consistently improve performance: similar models could share blind spots on difficult cases. The report describes a separate approach using two independent review chains, with disagreement or uncertainty sent to a person. Agreement between models is still not conclusive proof of correctness. These findings help explain why supervisory assurance has to examine the checking process itself, rather than count how many models approved an answer.
A registry remains under discussion
The separate agent-governance concept note expressly says it does not yet create specific supervisory requirements. It proposes connecting an agent's permitted actions to its underlying business purpose. The business function responsible for that purpose would generally be accountable for outcomes; the technology team would not automatically own that responsibility. HKMA is exploring an industry-level register of material agents to improve visibility over accountability and shared dependencies. A bank's own inventory of AI applications under the issued guidance should not be confused with that possible sector-wide register.
The note also considers limits on which tools, systems and data an agent may access, and intervention before actions with material consequences for customers. Mapping connections between models and outside providers could help identify where an incident began. HKMA plans a discussion paper seeking industry views on possible supervisory requirements, informed by further sandbox testing. It gives no publication date for that paper. The distinction between testing safeguards and imposing duties also runs through the separate UK parliamentary inquiry into AI safety.
