AI in action · Regulation
Hong Kong's Monetary Authority Issued AI Guidance Built on "Accountability, Resilience, Trust" and Reported That 60% of Its Sandbox Use Cases Are Now Agents
The Hong Kong Monetary Authority released new supervisory guidance on AI adoption on October 16, 2026, alongside an insight report from its generative-AI sandbox, Carmen Chu, its executive director for banking supervision, told the Finance + A.I. Forum. The second sandbox cohort produced more sophisticated results with about 30% less computing power than the first, more than 60% of its use cases involved AI agents and half of those were multi-agent; one bank cut customer onboarding time 40% and another cut market-risk analysis cycles 95%. The guidance asks for an owner for every use case, an inventory of approved applications, mapped dependencies and response measures "possibly including kill-switches," and the HKMA will explore a central supervisory registry of agents.
Hong Kong has gone from sandbox to rulebook faster than any other financial centre. On October 16, 2026 the Hong Kong Monetary Authority issued new supervisory guidance on the adoption of artificial intelligence by banks, built on what it calls the ART framework, for Accountability, Resilience and Trust, and published alongside it an insight report drawing on two cohorts of its generative-AI sandbox, run with Cyberport. Carmen Chu, the HKMA's executive director for banking supervision, set out the contents in her opening remarks at the Finance + A.I. Forum the same day. The guidance is aligned, she said, with the Financial Stability Board's sound practices published the previous day, and it recognises a gradual move from "human-in-the-loop" to "human-on-the-loop" oversight, which is the regulator's way of saying that agents acting without per-step approval are coming and need a different kind of supervision.
The sandbox figures are the most specific any regulator has published on what banks are actually building. In the second cohort, banks achieved more sophisticated outputs for similar use cases with about 30% less computing power than in the first; the share of multimodal use cases, handling sound, image and video, doubled; more than 60% of use cases involved AI agents, and half of those were multi-agent applications. The results the HKMA chose to cite were operational: a fraud-detection application cut one bank's customer-onboarding time by 40%, a wealth-management chatbot ran a 98% positive-feedback rate, and a market-risk agent cut the cycle time for case-specific risk analysis by 95%. The speech gave no figure for the number of banks or use cases, and none of the examples involves an agent trading or moving client money.
The guidance's demands follow the three words. Under accountability, every AI use case needs a clear owner and an up-to-date inventory of approved applications, with governance tied to each agent's intended purpose so that intervention and human-on-the-loop oversight are possible. Under resilience, banks should map their AI dependencies, models, data, tools, infrastructure and third parties, and have response measures ready, "possibly including kill-switches"; the HKMA will explore a centralised supervisory registry, which would give the regulator its own list of what agents are running where. Under trust, risk appetite should be translated into explicit boundaries for each agent and behaviour validated under extreme scenarios, work the next sandbox, GenAI Sandbox++, will support. A discussion paper on agentic-AI governance is to follow, and an AI training-data sharing platform is planned with Cyberport and the University of Hong Kong. "Trust starts with securing the boundaries around an agent," Chu said.
Set beside the month's other announcements, Hong Kong's is the operational one. Singapore's guidelines, finalised on October 7, take effect in a year and promise a consultation on agents in 2027; the UK's FCA published examples and asked for near-miss reporting; the Fed, OCC and FDIC asked questions. The HKMA has told banks to inventory their agents, name an owner for each, map what they depend on and prepare to switch them off, and has said it may keep its own register. For a bank running sixty per cent of its AI projects as agents, that is not guidance about a future technology; it is a description of the audit that is coming.
Sources: Hong Kong Monetary Authority — Opening Remarks by Carmen Chu at the Finance + A.I. Forum (Oct 16, 2026); Financial Stability Board — Final report: sound practices for responsible AI adoption by financial institutions (Oct 15, 2026)