AI governance for Hong Kong banks and licensed corporations

Hong Kong financial institutions govern AI through their regulators’ supervisory guidance rather than an AI statute. For banks the reference is now the HKMA’s Supervisory Guidance for Responsible Adoption of Artificial Intelligence, issued on 15 October 2026 under its Accountability, Resilience and Trust (ART) Framework. For brokers, fund managers and other licensed corporations it is the SFC’s circular on generative AI language models of November 2024. Personal data runs through both, under the Personal Data (Privacy) Ordinance. This guide sets out what each asks for in practice and where the work overlaps.

Who expects what

Banks. The HKMA’s guidance applies to all authorized institutions and supersedes its High-level Principles on Artificial Intelligence of 1 November 2019. It sets nine principles under three pillars and asks institutions to uplift their policies, procedures and controls in a timely manner, in proportion to their AI use. It came with a report on the second cohort of the GenAI Sandbox and a Concept Note on the Governance of Agentic AI.

Licensed corporations. The SFC’s circular applies to any licensed corporation offering services or functionality powered by AI language models in its regulated activities, however the model is sourced. It sets four core principles and singles out investment recommendations, advice and research as high risk. On 2 June 2026 the SFC added a circular on AI-enabled cyberattacks, pointing to the speed at which frontier AI models can find and chain vulnerabilities.

Everyone handling personal data. The HKMA guidance tells banks to ensure full compliance with the Personal Data (Privacy) Ordinance, and the SFC tells licensed corporations to have due regard for the PCPD’s AI Model Personal Data Protection Framework.

Start with the inventory and an owner for each use case

Both regulators start from knowing what AI is in use. The HKMA asks for an up-to-date inventory of AI use cases approved for deployment, whether built internally, provided by third parties or embedded in other products. It should generally record purpose, owner, provider and model, risk classification, lifecycle and approval status, dependencies and key controls. Use of anything outside the inventory should be restricted, staff told not to use unapproved tools or approved tools for unapproved purposes, and shadow AI controlled.

An inventory assembled from what project teams declare will miss AI switched on inside software the bank already licenses. It needs a discovery feed, from device management tools, cloud inventories, code repositories and IT asset registers, so new AI turns up as a finding to be reviewed. The shadow AI guide covers where to look.

Every deployed use case also needs an owner. The HKMA makes the owner responsible for defining the system’s authority, overseeing controls and monitoring, and escalating issues. Keep that owner on the inventory record, with its history, so accountability survives reorganisations.

Size the controls to the use case

The HKMA asks for an initial assessment of each use case’s materiality and risk, considering its purpose and data, the impact if it fails and how mature the technology is. That rating decides how much testing and approval the use case gets, and material or higher-risk use cases need senior management sign-off. The SFC takes the same risk-based approach and draws one bright line, around investment advice to clients.

For higher-risk uses the two regulators converge on a familiar set of controls.

• End-to-end testing before deployment and before material changes, with independent validation and adversarial testing for material uses.
• Ongoing monitoring with indicators and escalation thresholds, watching for drift, misuse, safeguard failures and unfair outcomes.
• Human oversight that fits the use case. The HKMA accepts human-on-the-loop or human-in-command arrangements where appropriate, while the SFC expects a human to check high-risk output for factual accuracy before it reaches the client.
• Credible fallback and recovery, and for higher-risk uses a kill switch that can suspend the system and revoke its access.
• Auditable records of the assessment, testing, approvals, deployment and changes, which for higher-risk outputs may include prompts and model versions.

Vendors and concentration

Most AI in a Hong Kong financial institution comes from a small number of model providers. The HKMA says institutions remain responsible for outcomes from external AI services, and asks them to assess concentration, continuity, substitutability and exit risk. Contracts should cover audit rights, data protection, notice of material changes and incidents, and should stop providers reusing customer personal data or confidential information to train models without authorisation. The SFC asks firms to assess a provider’s own model risk management where they can, define who handles which cybersecurity risks, and plan for the AI service being unavailable. The vendor AI risk guide covers due diligence and contract terms.

Customers and transparency

Under the HKMA’s Trust pillar, customers should be told when a service is powered by AI and of the relevant risks, existing disclosure and opt-out expectations continue to apply, and affected people should be able to seek clarification, review or reassessment. The SFC requires a prominent disclosure that a client is talking to AI and that its output may not be accurate.

Agents are next

The HKMA’s agentic AI concept note does not set requirements yet, but it shows where supervision is heading. It discusses tying each agent’s actions to a business intent with an accountable business owner, an inventory of material agents with clear component lineage, a possible industry-level registry of material agents, kill switches and operating boundaries on the tools and data an agent may use. A discussion paper is due. Institutions running agents can get ahead by giving each one its own identity, permissions and owner on the record now. The agents guide sets out how.

Where Aicura fits

Aicura’s AI Register holds every AI system and agent with its owners, provider, model, status and version history, and Discovery surfaces AI in use that is not yet on it. Risk and impact assessments rate each system and run again when it changes. Vendor records hold contracts and due diligence. For agents, the register records what each agent may do and its kill switch, and an agent can be suspended at the AI gateway its traffic passes through. Monitoring raises an alert when a tracked figure crosses a limit you set. Aicura gives guidance and supports the work. It does not assess against HKMA or SFC requirements, and the judgement on your obligations stays with you and your advisers.

A note on this page

This is Aicura’s reading of the HKMA and SFC documents, not legal or regulatory advice. Read the primary sources on the HKMA and SFC websites, and take advice on how they apply to your institution.