The Pulse
Hong Kong Plans Agent Registry and Kill Switches for Banking AI
Hong Kong’s monetary authority has outlined proposed controls for banking AI agents, including clearer accountability, an inventory of material agents and human intervention mechanisms. The ideas are preliminary; the HKMA says a forthcoming
AI.info Team · Written with AI ·
“The progression from generative to agentic A.I. represents an important new stage of innovation for the banking sector.”
Arthur Yuen, Deputy Chief Executive of the Hong Kong Monetary Authority
Hong Kong’s monetary authority is weighing new supervisory controls for banks’ AI agents, including a possible registry of significant systems and mechanisms to halt their actions. The proposals sit inside a wider responsible-AI initiative announced on 15 October, but the Hong Kong Monetary Authority (HKMA) stresses that its agent-governance concepts are preliminary—not binding requirements.
Accountability follows the business purpose
The HKMA’s concept note describes agentic AI as systems that can plan and carry out multi-step tasks, access data and tools, and interact with internal or external systems. Some deployments may involve multiple agents handling connected tasks. The authority sees possible uses in risk management, fraud prevention and customer experience, while warning that more autonomy requires governance arrangements to adapt.
One proposed principle, “intent-anchored accountability,” would connect an agent’s actions to the business objective it was assigned and identify the person or function responsible for its use. The authority says the business function behind that objective would generally be expected to own the agent’s outcomes; technology teams may support the system without being its accountable owner. The approach is intended to help banks judge whether an agent stayed within its authorized purpose.
A registry is under consideration, not yet a rule
The HKMA is exploring an industry-level registry for material AI agents used by authorized institutions. An inventory could record which agents perform significant functions, operate with high autonomy or access sensitive information and important systems. The authority says visibility could help banks monitor systems, clarify accountability and identify shared technology dependencies.
Those dependencies matter because agents can draw on interconnected models, data, tools, infrastructure and third-party services. If an incident occurs, that web of components can complicate efforts to trace its cause and impact. The concept note therefore points to clear component lineage as part of building operational resilience.
Human oversight is another proposed safeguard. The HKMA describes a move from “human in the loop” toward “human on the loop,” with monitoring and intervention scaled to the activity’s potential impact. Banks may need to define when a person must approve an action, suspend an agent’s access to a tool or system, or activate an emergency stop before an action with material customer consequences.
The broader ART guidance is already issued
The agent-governance ideas form part of the HKMA’s broader “ART” framework: Accountability, Resilience and Trust. The authority issued updated supervisory guidance on responsible AI adoption on 15 October, replacing its 2019 high-level AI principles. It directs authorized institutions to apply the principles proportionately and update their policies, procedures and controls in a timely manner.
The updated guidance calls for a named owner for each deployed AI use case, an up-to-date institutional inventory and risk-based human oversight. For material or higher-risk applications, it also calls for more stringent review and approval, ongoing monitoring, incident response and auditable records. The guidance says banks remain accountable for AI-generated outputs, including actions, even when they rely on external providers.
Sandbox trials show banks are testing agents
The HKMA said more than 60% of use cases explored in the second cohort of its Generative AI Sandbox involved agentic AI. The trials included systems designed to plan and execute multi-step tasks within defined boundaries and controls. The authority also announced an AI training-data sharing platform with Hong Kong Cyberport and the University of Hong Kong, with data-quality assurance as a core design principle.
The next step for agent oversight is a discussion paper that the HKMA says it will issue soon, setting out options for possible supervisory requirements and seeking views from banks and other stakeholders. Until then, the registry, intervention mechanisms and intent-based accountability remain proposals for industry discussion; the separate ART guidance is the document that sets out the authority’s current supervisory expectations.