Key Takeaways
- The HKMA announced the ART Framework (Accountability, Resilience, Trust) on 15 October 2026 to govern AI adoption in Hong Kong banking.
- It combines new supervisory guidance, a shared AI training data platform and further agentic AI testing, with a discussion paper on agent governance to come.
- Three ideas lead that paper: intent-anchored accountability, a registry of material AI agents and effective intervention mechanisms.
- Non-banks will feel it through bank vendor reviews, agent-initiated payments and the PCPD's parallel agentic AI guidance.
- Start now: inventory your agents, name owners, tier them by materiality and test the kill switch.
A Hong Kong bank's operations team has an agent in pilot. It checks trade-finance documents, pulls data from three systems and queues payments for approval. The technology works. Then the risk committee asks a question nobody had prepared for: if the agent does something nobody intended, whose intent was it carrying out, and how fast can we stop it?
That question now has a regulator's name on it. On 15 October 2026 the Hong Kong Monetary Authority announced the ART Framework, and the questions it raises reach well beyond banks.
What is the HKMA ART Framework?
The ART Framework is the Hong Kong Monetary Authority's umbrella for responsible AI adoption in banking, announced on 15 October 2026. ART stands for Accountability, Resilience and Trust. It bundles new supervisory guidance, a shared training-data platform and further agentic AI testing, with a discussion paper on agentic AI governance to follow.
According to the HKMA's announcement, the framework sits under the regulator's Fintech 2030 agenda. It is not a single rulebook. It is a set of initiatives that together tell banks what the supervisor expects as AI systems gain more autonomy and more access to data, tools and workflows.
The three initiatives under ART
- Supervisory guidance: the latest Supervisory Guidance for Responsible Adoption of Artificial Intelligence, built on lessons from the second cohort of the HKMA's Generative AI Sandbox.
- Training data platform: an AI training data sharing platform built with Cyberport and the University of Hong Kong, with data quality assurance as a core design principle.
- Agentic AI testing and policy: further trials through the GenAI Sandbox++ and an upcoming discussion paper proposing a supervisory framework for agentic AI governance.
HKMA Deputy Chief Executive Arthur Yuen described the move from generative to agentic AI as an important new stage of innovation for banking. The framework is meant to pair that innovation with clear accountability, resilient operating models and effective human oversight.
Why did the HKMA introduce the ART Framework now?
Because banks have moved from chatbots that draft text to agents that plan and execute multi-step tasks. In the HKMA's second GenAI Sandbox cohort, more than 60% of use cases explored agentic AI. Supervision written for models that only suggest is not enough for systems that act.
A generative model that writes a draft creates a review problem. An agent that reads a ledger, calls an API and books a transaction creates a control problem. The second kind of system can make a chain of small, individually reasonable decisions that add up to an outcome nobody approved.
The timing also reflects how fast the wider market is moving, and how often it stumbles. Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. A regulator that sets expectations early gives banks a reason to build controls before the pilot, not after the incident.
Two other signals arrived within weeks. In August the Privacy Commissioner issued its first standalone guidance on agentic AI, and on 2 November the HKMA launched Project SAGE to explore agentic finance more broadly. Read together, Hong Kong's regulators are converging on one question: who answers for an agent?
How does the ART Framework approach AI agents?
Through three ideas flagged for the upcoming discussion paper: intent-anchored accountability, a registry of material AI agents, and effective intervention mechanisms. In practice, expect that to mean every agent action traces back to a human's stated intent, important agents are inventoried, and someone can stop them quickly.
The discussion paper has not been published yet, so the detail is still open. The three ideas named in the announcement are nonetheless a useful design brief for any organisation running agents today.
Accountability: anchored to intent
An agent cannot be accountable; a person can. Intent-anchoring means recording what a named person asked the agent to achieve, and within what limits, so that every action can be traced to an instruction someone owns. The Privacy Commissioner's guidance makes the same point from another angle: AI agents are not legal persons, so the deploying organisation stays responsible.
Resilience: a registry and the power to intervene
A registry of material agents is the agentic equivalent of an application inventory: what each agent does, which systems and data it can reach, who owns it and how much its decisions matter. Intervention mechanisms are the brakes: kill switches, approval gates for irreversible actions, and a fallback to manual processing when an agent is paused.
Trust: data quality and testing
The training data platform and sandbox testing address the input side. An agent built on poor data or untested workflows will eventually fail in ways that erode customer and supervisor trust, whatever the governance paperwork says.
How is ART different from the PCPD's agentic AI guidance and Project SAGE?
ART sets supervisory expectations for how banks adopt AI. The PCPD's August 2026 guidance explains how the Personal Data (Privacy) Ordinance applies to agentic AI in every sector. Project SAGE explores agentic finance, where agents transact with financial services. They overlap, but each answers a different question.
| Initiative | Issued by | Who it applies to | Core question |
|---|---|---|---|
| ART Framework (15 Oct 2026) | HKMA | Authorized institutions (banks) | How should banks govern the AI they deploy? |
| Agentic AI privacy guidance (Aug 2026) | PCPD | Any organisation handling personal data | How does the PDPO apply when agents process personal data? |
| Project SAGE (2 Nov 2026) | HKMA | Banks, fintechs and technology providers | How can agents transact safely with financial services? |
Project SAGE is worth watching even if you are not a bank. Its four workstreams cover use cases across business-to-consumer, business-to-business (such as procurement and treasury management) and machine-to-machine payments; agent identity, including Know Your Agent practices; interoperability standards; and legal questions such as liability and consumer protection. If your finance team expects to let an agent pay suppliers or manage treasury tasks, these are the rules your bank will be working under.
The PCPD guidance is non-binding, but legal commentary such as Bird & Bird's analysis treats it as the practical benchmark for how the Commissioner reads existing PDPO duties.
Does the ART Framework matter if my company is not a bank?
Yes, indirectly. Banks will push ART expectations down to their technology vendors and outsourcing partners, other regulated sectors tend to follow HKMA thinking, and any company letting agents touch payments will meet these controls through its bank. ART is also a ready-made governance template for firms without a sector regulator.
Consider three Hong Kong scenarios.
- A logistics group uses an agent to reconcile freight invoices and trigger supplier payments. Its bank will increasingly ask how those payment instructions are authorised, and whether a person or an agent is initiating them.
- A professional services firm sells AI-assisted services to banks. Expect procurement questionnaires to ask for an agent inventory, a named owner per agent and evidence of human override, because the bank's supervisor will ask the bank the same thing.
- A property management company has no sector regulator of its own. It can adopt ART's three principles as an internal standard instead of writing a governance framework from scratch.
The pattern is familiar. When the HKMA set technology risk and outsourcing expectations in the past, they reached suppliers through contracts long before most people outside banking read the original circulars. Agentic AI governance is likely to travel the same way.
What do enterprises get wrong about agentic AI governance?
The most common mistake is treating an agent like software that only needs an IT security review. Agents make decisions, so they need a business owner, defined limits and a way to stop them. The second mistake is waiting for final rules, because the controls ART points to are useful whatever the final wording.
- Governance by policy document only. A policy stating that people remain accountable means little without logs linking each agent action to an instruction and an owner.
- No inventory. Many organisations cannot list every agent in production, including agent features embedded in SaaS tools. You cannot register what you have not found.
- Kill switches that were never tested. An intervention mechanism nobody has exercised is a hope, not a control. Rehearse pausing an agent and falling back to manual work.
- Treating materiality as binary. An agent that drafts internal summaries and an agent that releases payments need very different oversight. Tier them.
What should enterprise leaders do in the next 90 days?
Build a register of every AI agent in use, name an accountable owner for each, record what each agent is permitted to do, and test that you can stop it. Then rank agents by materiality, so oversight effort goes to the agents that move money, personal data or customer outcomes.
A four-step starting point
- Inventory: list every agent, including features inside existing platforms, with the systems and data each one can reach.
- Anchor intent: for each agent, document the business objective, the limits and the person who owns the instruction.
- Tier by materiality: separate agents that inform from agents that act, and agents that act reversibly from those that act irreversibly.
- Prove intervention: run a drill. Pause your highest-tier agent and time how long it takes to stop it and continue the work manually.
This is also your board conversation. A one-page agent register showing owners, tiers and the result of the last intervention drill answers most of what a risk committee, an auditor or a bank's vendor review will ask.
Accountability is the design brief, not the afterthought
The ART Framework does not ban anything, and it does not set new limits on what agents may do. It sets a direction: autonomy is acceptable when someone owns the intent, the agent is on a register and it can be stopped. Organisations that build those three habits now will find the HKMA's discussion paper, their bank's vendor questionnaires and the PCPD's expectations far easier to meet.
The hard part is not the technology. It is deciding who owns each agent, and keeping the register current as agents multiply.
Empowering AI. Centered on you.
Put an owner and a register behind every AI agent
If you are planning AI agents for operations, finance or customer service, build the governance in from day one. Our team starts with one role and one workflow, keeps data in Hong Kong with PDPO in mind, and we'll walk you through every step, from a free AI Staff assessment to go-live and KPI tracking.
Book a Free AI Staff Assessment
Frequently Asked Questions
Is the HKMA ART Framework legally binding?
The ART Framework is a supervisory initiative, not legislation. Banks are expected to follow HKMA supervisory guidance, while the agentic AI governance proposals will first appear as a discussion paper before any final expectations are set.
When will the HKMA publish its agentic AI discussion paper?
The HKMA called the paper upcoming in its 15 October 2026 announcement but gave no date. Watch the HKMA press release page and your bank's compliance updates.
Who is responsible when an AI agent makes a mistake in Hong Kong?
The organisation that deploys it. The PCPD's 2026 agentic AI guidance states that AI agents are not legal persons, so PDPO duties stay with the data user.
What does Know Your Agent (KYA) mean?
Know Your Agent means verifying an AI agent's identity and the authority delegated to it before it takes part in financial activity, much like Know Your Customer checks for people. The HKMA is exploring it under Project SAGE.
What should a register of material AI agents record?
Each agent's purpose, the systems and data it can reach, its accountable owner, its materiality tier and how it can be stopped. The HKMA has named such a registry as one idea for its agentic AI governance framework.