Opening Remarks at Finance + A.I. Forum

Introduction

1. Good morning, everyone. A very warm welcome to the Finance + A.I. Forum co-hosted by the HKMA and Cyberport.
2. We are also gathering here to celebrate the successful completion of the second cohort of the GenA.I. Sandbox.
3. The second cohort shows how quickly A.I. technology is advancing. Compared with the first cohort, for similar use cases, banks are able to achieve more sophisticated outputs with around 30% less computing power.  Open-source models, which are increasingly capable, have contributed to these gains.
4. The share of use cases with multimodal capabilities, including sound, image and video, has also doubled from the first cohort.
5. More importantly, these advances are delivering tangible results. For example,   
   • more efficient and advanced fraud detection capability enabled a bank to reduce customer-onboarding time by 40%;
   • a wealth-management chatbot achieved a 98% positive-feedback rate by providing more personalised customer support; and
   • a market-risk agent helped shorten the cycle of case-specific risk analysis by 95%.
6. These are impactful outcomes. They show that, when developed and deployed responsibly, A.I. can bring real value to banks, their customers and the wider financial system.
7. The Sandbox has also provided valuable practical insights on the risks and controls associated with A.I. adoption. These observations have informed the HKMA’s iterative development of our supervisory expectations.

Supervisory Guidance

8. We’ve just issued our latest supervisory guidance on A.I. adoption, together with an insight report that shares good practices and lessons learnt from the first two cohorts of the Sandbox.
9. It also aligns with the latest guidance by international bodies including the Financial Stability Board, to which the HKMA is a member contributing actively to the discussion.
10. Our latest guidance is built around the “ART” Framework, comprising three pillars: Accountability, Resilience and Trust. It provides a practical basis for responsible A.I. adoption.
11. For example, the guidance calls for clear ownership of each A.I. use case, supported by an up-to-date inventory of approved applications. It also recognises a progressive shift from “human-in-the-loop” to “human-on-the-loop” oversight.
12. As A.I. use cases become more sophisticated, effective oversight need not mean human involvement in every step. But it must mean clear accountability and on-going monitoring.
13. This will become even more important in the light of another key development and observation from the second cohort: the emergence of agentic A.I.

Agentic A.I.

14. In the second cohort, more than 60% of use cases involved A.I. agents, and half of these were multi-agent applications.
15. Unlike earlier A.I. applications that primarily generate insights, content or recommendations, agentic A.I. can plan and undertake multi-step tasks, use tools, access data and systems, and most importantly, act with a high degree of autonomy.
16. But greater autonomy also brings particular characteristics that we need to address:
    • First, agency – the ability to plan and act;
    • Second, complexity – across models, tools and data sources; and
    • Third, novelty – in outputs and behaviours.
17. The “ART” Framework provides a structured way to address these characteristics. Let me briefly elaborate.

Accountability

18. First, accountability. For agentic A.I., governance must be anchored to intent. Banks should be clear about what the agent is designed to achieve, and ensure that such intent is carried through the entire process – from its design and data, to the tools that an agent can use and actions it can take.
19. Such an intent-anchored accountability is a prerequisite for effective intervention and the progression towards “human-on-the-loop” oversight. As agentic interactions are expected to become faster and more complex, “A.I. versus A.I.” can also be a useful strategy to support meaningful human accountability.

Resilience

20. Second, resilience. Banks need to map their A.I. dependencies. This means maintaining clear lineage across the components that support an A.I. application: models, data, tools, infrastructure and third-party providers.
21. Resilience also means robust response and recovery capabilities. Financial institutions should have corrective and restrictive measures at hand, supported by timely intelligence sharing.  Where appropriate, these measures may include “kill-switch” mechanisms to suspend agents before an issue escalates.
22. We will also explore whether a centralised supervisory registry could support this work across the banking sector.

Trust

23. Third, trust. Trust starts with securing the boundaries around an agent – often referred to as the agent harness.
24. In financial services, the agent harness has a particular purpose. It should reflect and operationalise the institution’s wider risk-management framework.  This means translating risk appetite into clear boundaries for an agent.
25. Trust can also be established through observation, including by validating behaviours under extreme scenarios. The GenA.I. Sandbox++ will continue to support financial institutions in assessing the practical application of appropriate guardrails as the technology develops.
26. Based on all the Sandbox trials so far, one consistent lesson learnt is that effective validation depends on quality data.
27. To address this, we’re happy to announce our collaboration with Cyberport and the University of Hong Kong to establish an A.I. training data sharing platform. This platform will provide the data required for effective training and testing of A.I. systems.

Conclusion

28. Taken together, the latest supervisory guidance, the Sandbox practical insight report, our forthcoming discussion paper on agentic A.I. governance, and the new A.I. training data sharing platform, will all help drive responsible adoption of A.I. as part of the HKMA’s Fintech 2030 strategy.
29. Our goal is clear: to capture the opportunities of A.I. through:
    • Accountability for agency;
    • Resilience amidst complexity; and
    • Trust in the face of novelty.

30. This is the spirit of “ART”: creativity requires discipline; innovation demands responsibility.
31. Let’s bring “ART” to action, and further promote responsible A.I. adoption to the next level. Thank you.