The Hong Kong Monetary Authority on 15 October announced its “ART” framework for artificial intelligence in banking, built on accountability, resilience and trust. Most of the package consolidates work already under way: updated supervisory guidance, a shared training data platform and further sandbox testing. The item with new regulatory substance is a promised discussion paper on supervising agentic AI, which the HKMA says will explore “intent-anchored accountability, a registry of material A.I. agents, and effective intervention mechanisms”. Those three ideas, if they become rules, would put the HKMA ahead of Singapore in setting specific expectations for AI agents, since MAS plans only to consult on the subject in 2027.
The framework has three working parts
The first is the latest Supervisory Guidance for Responsible Adoption of Artificial Intelligence, which the HKMA says draws on the second cohort of its Generative AI Sandbox. The second is an AI training data sharing platform built with Cyberport and the University of Hong Kong, with data quality assurance as a core design principle. The third is more testing of agentic applications through the GenAI Sandbox++ and the upcoming discussion paper.
The HKMA co-hosted a “Finance + A.I. Forum” with Cyberport on 15 and 16 October to mark the end of technical trials in the second sandbox cohort. More than 60% of the use cases in that cohort explored agentic AI, which the HKMA describes as including automated planning and execution of multi-step tasks within defined boundaries and controls.
That figure is the reason the discussion paper exists. If most of what banks want to test involves software that plans and acts on its own, supervision written for models that produce a score or a draft for a human to review will not be enough.
A registry would treat agents like regulated functions
The HKMA has not published the discussion paper, so the details are unknown. The three phrases in the announcement still indicate the direction.
“Intent-anchored accountability” suggests that a bank would have to define, in advance, what an agent is meant to achieve and within what limits, and that a named person would answer for outcomes against that stated intent. A “registry of material A.I. agents” suggests banks would keep, and perhaps report to the HKMA, a list of agents above a materiality threshold, much as they maintain inventories of models and outsourcing arrangements. “Effective intervention mechanisms” points to kill switches and override procedures that work quickly enough to stop an agent before harm compounds.
For a bank, the registry is the idea with the largest operational cost. It requires a definition of what counts as an agent, a materiality test and a process to keep the list current as business units deploy new tools.
By pairing innovation with clear accountability, resilient operating models and effective human oversight, banks can fully realise the benefits of A.I. while maintaining the trust and stability for Hong Kong’s financial system.
Arthur Yuen, Deputy Chief Executive, HKMA, 15 October 2026
Singapore has finished its rules and left agents for 2027
The comparison with Singapore is instructive. The Monetary Authority of Singapore issued final AI risk management guidelines on 7 October, applying to all financial institutions with phased effect from October 2027. MAS requires an inventory of AI use cases, systems or models, but has said only that it intends to consult in 2027 on what additional guidance on agentic AI would be useful.
Hong Kong is moving the other way round. Its general guidance takes the form of supervisory guidance updated through sandbox experience, and its first formal policy document on agents is already promised. A bank operating in both centres could face an AI inventory under MAS rules and a separate agent registry under HKMA rules, with different definitions, unless the two regulators align their terms.
The data platform is aimed at smaller banks
The training data sharing platform with Cyberport and the University of Hong Kong addresses a different problem. Large banks can draw on their own customer and transaction data to train and test models, while smaller institutions have far less to work with. A shared platform with quality controls could lower the cost of building and validating models for those banks, though the HKMA has not said what data it will hold or who may use it.
The framework sits under the HKMA’s Fintech 2030 strategy, which Yuen cited in his statement.
What to watch
- Publication of the HKMA discussion paper on a supervisory framework for agentic AI governance.
- The definition of “material” agents and whether registry entries are reported to the HKMA.
- Further results from the GenAI Sandbox++.
- Any alignment of terms between HKMA and MAS on AI inventories and agents.
Primary sources
- HKMA, HKMA announces “ART” Framework to promote responsible A.I. adoption, 15 October 2026
- MAS, MAS sets out supervisory expectations on responsible AI adoption by financial institutions, 7 October 2026
Related: MAS gives financial institutions until October 2028 to put AI life cycle controls in place
Photo (cropped): Suicasmo, CC BY-SA 4.0 via Wikimedia Commons
_20170808.jpg){kind=link}