AI agents dominate bank trials as governance demands expand
Carmen Chu, Executive Director for Banking Supervision at the Hong Kong Monetary Authority (HKMA), set out AI safeguards on October 16, 2026. At the Finance + A.I. Forum, she reported second-cohort sandbox results and new supervisory guidance.
Less computing, more autonomy
Banks in the second GenA.I. Sandbox cohort achieved more sophisticated outputs using around 30 percent less computing power for similar tasks, helped by open-source models.
The share of multimodal applications doubled.
One bank cut onboarding time by 40 percent; a wealth-management chatbot earned 98 percent positive feedback; a market-risk agent shortened case-specific analysis by 95 percent.
Over 60 percent of use cases involved AI agents, half involving multiple agents.
The HKMA issued guidance and a report drawing on both cohorts.
Its three-pillar ART framework covers Accountability, Resilience and Trust, reflecting guidance from international bodies including the Financial Stability Board.
Agents can plan and execute tasks autonomously.
Autonomy requires operating boundaries
The guidance calls for clear ownership, an updated application inventory and ongoing monitoring as oversight moves toward human-on-the-loop models.
Banks must anchor agents to intended purposes and map dependencies across models, data, tools, infrastructure and providers.
Controls include recovery measures, intelligence sharing and potential kill switches.
The HKMA will explore a central supervisory registry.
Agent boundaries must implement banks' risk appetite; extreme-scenario validation and Sandbox++ trials support testing of guardrails.
Control depends on test data
Chu identified quality data as a prerequisite for effective AI validation.
The HKMA, Cyberport and the University of Hong Kong will establish an AI training data sharing platform for training and testing.
Alongside a forthcoming agentic AI governance discussion paper, this supports the HKMA's Fintech 2030 strategy.