Summary

  • HKMA’s ART package places updated AI guidance, training-data quality and agentic-AI testing on the same banking-supervision path.
  • More than 60% of second-cohort sandbox use cases explored agentic AI. That is an exploration share, not a measure of banks or production systems.

Analysis

The Hong Kong Monetary Authority (HKMA) announced its Accountability, Resilience and Trust framework on 15 October. The release describes a sequence of measures for banks adopting AI, while leaving the detailed agentic-AI supervisory proposal for a forthcoming discussion paper. It does not announce a new AI statute or say that the proposed agent framework is already in force.

The package has three current tracks. HKMA says it has issued updated supervisory guidance for responsible AI adoption, drawing on the second cohort of its Generative AI Sandbox. It also announced a training-data sharing platform with Hong Kong Cyberport Management Company Limited and the University of Hong Kong, naming data-quality assurance as a design principle. And it plans more agentic-AI testing through GenAI Sandbox++. The release gives no platform launch date, access terms or technical requirements, and does not set a timetable for the discussion paper.

The sandbox figure signals interest, not deployment. HKMA says more than 60% of the second cohort’s use cases explored agentic AI, including automated planning and multi-step tasks within defined boundaries and controls. The release does not state the denominator, report how many trials passed evaluation, or say that any use case reached production. In 2025 HKMA said it had selected 27 use cases from 20 banks and 14 technology partners for the second cohort. That earlier selection count is useful context, but the 2026 release does not confirm that the same 27 cases formed the percentage’s calculation base.

The distinction matters because the 2019 AI circular already placed accountability for AI-driven decisions with bank boards and senior management. It also called for defined responsibilities across the three lines of defence, proportionate explainability, good-quality data, model validation, audit records and oversight of third-party providers. The new release says updated guidance has been issued, but it does not link the document or describe which expectations changed. The 2019 circular remains listed as current in HKMA’s regulatory repository; the announcement does not say it has been replaced.

Agentic systems can plan and carry out several steps using data, tools and workflows. That makes the move from a bounded test to a live banking process a governance transition: someone must own the process, know what the agent can do, review its evidence and be able to stop or constrain it. HKMA says its future discussion paper will explore intent-anchored accountability, a register of material agents and intervention mechanisms. Those are agenda items, not yet a published control standard.

For now, ART is best read as a staged adoption program: guidance and testing are under way, data infrastructure is announced, and the proposed agent-specific supervisory detail remains ahead.

Sources