Hong Kong’s banking regulator has set out how it wants banks to govern AI that can act on its own. On October 15, the Hong Kong Monetary Authority announced its “ART” Framework, short for Accountability, Resilience and Trust, to support responsible AI adoption in banking.
The framework starts with new supervisory guidance on responsible AI adoption, drawn from the second cohort of the HKMA’s Generative AI Sandbox. The regulator is also building an AI training-data sharing platform with Cyberport and the University of Hong Kong, with data quality assurance as a core design principle. It will run further tests of agentic AI in its GenA.I. Sandbox++ and then publish a discussion paper on supervising it. That paper will explore intent-anchored accountability, a registry of material AI agents and ways to intervene when something goes wrong.
More than 60% of the use cases in the second cohort explored agentic AI, the HKMA said. “The progression from generative to agentic A.I. represents an important new stage of innovation for the banking sector,” said Arthur Yuen, the HKMA’s Deputy Chief Executive.
Three participating banks published their results the same day. Bank of China (Hong Kong) validated an investment information assistant and an AI voice service that understands mixed Chinese and English speech. In testing, the voice system could also identify AI-synthesised voices, which could help flag suspicious calls. BOCHK has been selected for the Sandbox++.
CCB (Asia) tested a “Smart RM Assistant”, a multi-agent system that retrieves customer profiles and past interactions and makes service recommendations based on product guidelines and internal policies. An “AI vs AI” review step flags unsupported statements, omitted conditions or misleading wording. The bank reported no performance figures.
Ping An Digital Bank worked with OneConnect on an anti-fraud platform that checks user selfies in real time for forged or manipulated faces. It reported deepfake detection accuracy above 96% in sandbox testing.
Herald View
Two parts of the HKMA package will matter more than the framework’s name. The first is the proposed registry of material AI agents, which would give supervisors an inventory of systems that act on customer data rather than simply advise. The second is timing: more than 60% of the second-cohort use cases were agentic, so the guidance is being written from live bank testing, not ahead of it. For Hong Kong lenders, the near-term work is to assign an accountable owner to each agent and document how it can be overridden, before the discussion paper is published. Banks elsewhere testing similar tools, such as the relationship-manager assistant CCB (Asia) trialled, should treat the paper as an early read on where their own supervisors are likely to go.
