Why the HKMA ART Framework names an owner for each AI use

The HKMA ART Framework, issued to every authorized institution on October 15, 2026, makes a named owner, a live inventory and a workable stop-control the price of putting artificial intelligence into a Hong Kong bank, in the same week the Financial Stability Board published a non-binding toolkit on the same risks. The concept note of that day treats agentic systems as a problem of intent, visibility and intervention, and the HKMA says a discussion paper will follow.

Deputy Chief Executive Arthur Yuen told the Finance + AI Forum on October 16, 2026 that more than 60% of use cases in the second generative-AI sandbox cohort involved AI agents, and that half of those agent use cases were multi-agent applications. Annex A to circular B1/15C replaces the High-level Principles on Artificial Intelligence of November 1, 2019 with nine expectations grouped as accountability, resilience and trust. The sections below set the nine expectations against the FSB's 12 sound practices, the narrower UK model-risk statement, and the US adviser penalties for a false AI claim.

Key facts on the HKMA ART Framework

• Circular B1/15C of October 15, 2026, signed by Carmen Chu, executive director (banking supervision), is addressed to the chief executive of every authorized institution, per the HKMA circular.
• Annex A supersedes the High-level Principles on Artificial Intelligence of November 1, 2019 and sets nine expectations under accountability, resilience and trust, in the supervisory guidance.
• Authorized institutions are expected to uplift policies in a timely manner, and Annex A leaves any fixed compliance date unstated.
• On October 16, 2026 Yuen said more than 60% of second-cohort use cases involved AI agents, half of those were multi-agent, and similar tasks used around 30% less computing power, in his forum remarks.
• Annex C treats agentic-AI governance as preliminary and says a discussion paper will follow, in the concept note.
• The FSB report of October 15, 2026 sets out 12 sound practices as a non-binding toolkit after a June 2026 consultation, per the FSB press release and the final report.

Methodology, time window and primary sources

Primary documents of October 15 and October 16, 2026 anchor the Hong Kong side. The file is circular B1/15C, Annex A, Annex C, the HKMA press release, and Yuen's remarks the next morning. The circular also publishes Annex B, Practical Insights from the Second Cohort of the GenAI Sandbox, described as observations and good practices cross-referred to the guidance. Sandbox performance figures below are taken from the October 16 remarks.

Related:SEC 4-927: when should a tokenised stock venue check who can trade?

Beside that file sit the FSB report of the same day, Prudential Regulation Authority (PRA) Supervisory Statement 1/23, effective May 17, 2024, and the Securities and Exchange Commission (SEC) orders of March 18, 2024 against Delphia (USA) Inc., and Global Predictions Inc. The jurisdictions in view are Hong Kong, the FSB, Britain and the United States.

What the HKMA ART Framework actually requires of a bank

The HKMA ART Framework is the Hong Kong Monetary Authority's October 15, 2026 supervisory package for responsible use of artificial intelligence by authorized institutions, anchored on accountability, resilience and trust. Annex A to circular B1/15C sets nine expectations and says they supersede the High-level Principles on Artificial Intelligence of November 1, 2019. Principle 2 requires an owner for each deployed use case, responsible for the system's authority, the effectiveness of controls and the escalation of problems, while the board and senior management remain accountable for the outputs. Principle 3 requires an up-to-date inventory of approved use cases, covering internally built, third-party and embedded tools, and it tells institutions to keep unapproved tools, and approved tools used for unapproved purposes, under control as shadow AI. Principle 4 scales testing, monitoring and fallback to assessed risk and, for material or higher-risk cases, calls for a kill switch. Yuen said on October 16, 2026 that more than 60% of second-cohort use cases involved AI agents.

The board approves an AI strategy stating approach and risk appetite, and the three lines of defence handle approval before deployment and monitoring after it. Oversight may be human-in-the-loop, human-on-the-loop or human-in-command, and reviewers need competence, information, time and authority to challenge or stop the system. Footnote 3 records the shift from the August 19, 2024 consumer-protection circular, which had called for a human-in-the-loop on early customer-facing generative tools.

Each use case is assessed for purpose, data, actions, harm if it fails, and the institution's competence to run it. Material or higher-risk cases need senior-management sign-off, independent validation and, where appropriate, adversarial testing. Where the system takes actions, a material case needs a kill switch or equivalent that can suspend the model and revoke access, and significant incidents go to the HKMA under prevailing reporting rules.

Personal data use has to meet the Personal Data (Privacy) Ordinance, and contracts are expected to keep a provider from training on customer data except where the institution has authorised it and the law allows it. The institution remains responsible for vendor-model outcomes. Cyber controls are expected to address data poisoning and prompt injection, because model instructions or filters alone fall short of principle 5. Principle 6 expects enough resources and role-specific training. Principles 7 to 9 keep the institution accountable for outputs, including hallucinations, and require the customer to be told that AI is in use and what the risks are.

How Hong Kong, the FSB, Britain and the US split the same problem

Sources: HKMA document 20261015-4-EN, the FSB final report page, the PRA SS1/23 PDF, and SEC press release 2024-36, checked on October 24, 2026.

Hong Kong's October 15, 2026 expectations, the FSB's 12 sound practices of the same date, and the PRA's SS1/23 statement effective May 17, 2024 divide supervision of bank models into three different legal bites. Annex A tells every authorized institution to name an owner, keep an inventory and scale controls to materiality, and it uses the verb expected. The FSB executive summary strongly encourages boards to use its 12 practices, labels them non-prescriptive, and allows smaller firms to apply only the relevant practices or a modified version. SS1/23 brought five model-risk principles into effect for UK firms that already hold internal-model permission, a narrower class than Hong Kong's authorized institutions, and the SS1/23 text uses neither the phrase artificial intelligence nor the phrase machine learning. On March 18, 2024 the SEC's combined civil penalties in the Delphia and Global Predictions orders were US$400,000, a figure neither B1/15C nor SS1/23 contains.

The FSB case study describes central inventories at large banks and kill switches that can pause an agent's access, while Annex A asks every authorized institution for an inventory and places a kill switch on material or higher-risk use cases. A toolkit a firm may trim still differs from an expectation addressed to every chief executive. The same week carried the FSB letter on cross-border vulnerabilities to the G20. Private markets are already selling Know Your Agent checks on AI agents, while Annex C only explores an industry-level registry.

"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. Such an intent-anchored accountability is a prerequisite for effective intervention and the progression towards “human-on-the-loop” oversight."

— Arthur Yuen, Deputy Chief Executive, Hong Kong Monetary Authority (HKMA remarks, October 16, 2026)

Where false AI claims have already drawn a cash penalty

On March 18, 2024 the Securities and Exchange Commission announced settled charges against Delphia (USA) Inc., and Global Predictions Inc., for false and misleading statements about their use of artificial intelligence. Combined civil penalties were US$400,000: US$225,000 under Advisers Act Release IA-6573 for Delphia, and US$175,000 under Release IA-6574 for Global Predictions, plus censure and cease-and-desist relief. The Commission's press release 2024-36 summarises orders filed as IA-6573 and IA-6574.

On the Commission's account, Delphia's statements from 2019 to 2023 claimed that client data would make its artificial intelligence smarter at picking companies and trends. The order finds those statements false and misleading because the firm lacked the capabilities, and it charges the Marketing Rule bar on an advertisement with an untrue statement of material fact. Global Predictions used "first regulated AI financial advisor" and "[e]xpert AI-driven forecasts" in 2023. The orders bind those two advisers. Principles 7 to 9 apply that conduct test inside a bank, and B1/15C leaves a separate fine tariff unstated.

The 40% onboarding cut, the 98% feedback rate and the 95% shorter cycle are participant examples, beside the cohort agent-share above 60%. One wording gap is internal to the HKMA file: the October 15 press release quotes Yuen giving the middle pillar as "Resiliency", while Annex A and the October 16 remarks both use "Resilience", and the Annex A spelling is the one attached to the expectations.

What the package means for banks, vendors and compliance teams

For an authorized institution the centre of the package is the inventory plus the owner. The board strategy states risk appetite, each deployed use case has a named owner who can escalate, and the three lines of defence have a path for approval and later monitoring. Where the use case can act, the runbook names who may throw the kill switch and how the fallback runs. Footnote 6 lists indicators, escalation thresholds and the risks of drift, misuse and unfair outcomes.

Footnote 3 lets a bank move some customer-facing generative tools from the August 19, 2024 human-in-the-loop stance towards human-on-the-loop or human-in-command, provided reviewers still have competence and authority. Firms outside this circular are already widening fraud checks on AI-agent checkout. If an authorized institution's model takes an action, oversight covers that path, and principle 9 still requires the customer to be told that AI is in use. The institution stays responsible for vendor outcomes, and training on customer personal data needs an authorisation under the Personal Data (Privacy) Ordinance.

A securities intermediary comes inside B1/15C only where it also holds authorized-institution status, and fund managers, exchanges and crypto-asset platforms stay in other regimes when that status is absent. The point that a technology change leaves older operational-resilience duties standing is tracked in GC26/7 on SYSC 15A operational-resilience duties. Footnote 7 still asks for model versions, prompts and the evidence behind material generative outputs.

"The twelve sound practices are intended as a flexible toolkit, rather than as minimum standards. They can be adapted to the size, business model and risk profile of an institution, as well as to the nature and scale of its AI adoption."

— John Schindler, Secretary General, Financial Stability Board (FSB remarks, October 21, 2026)

What is still open on agentic AI and the data platform

Annex C describes agentic AI as systems that plan multi-step tasks and reach data and tools. Intent-anchored accountability puts the owner with the business line that holds the intent. The HKMA is exploring an industry-level registry of material agents. Oversight moves towards human-on-the-loop, and a risk-integrated agent harness turns risk appetite into limits on tools and data.

The note states that it "does not constitute specific supervisory requirements at this stage." A discussion paper is promised soon, and a close date is unstated in the October 15 documents. Trials continue through the GenAI Sandbox++, and the press release announces a training-data platform with Hong Kong Cyberport Management Company Limited and the University of Hong Kong, with an operating date unstated. 24-hour incident reporting under the Cyber Resilience Act states an hour-count, while Annex A principle 4 sends significant AI incidents back to prevailing HKMA rules. The circular ties the package to the Fintech 2030 strategy and the Fintech Promotion Blueprint of February 3, 2026.

TL;DR

The HKMA ART Framework is the October 15, 2026 package in which the Hong Kong Monetary Authority told authorized institutions to govern artificial intelligence through accountability, resilience and trust. Circular B1/15C and Annex A replace the November 1, 2019 principles with nine expectations: a board strategy, an owner for every deployed use case, a live inventory against shadow AI, controls scaled to materiality, and a kill switch on material or higher-risk cases. Policies are to be uplifted in a timely manner, and a fixed compliance date is unstated. Annex C sketches intent-anchored accountability and a possible registry of material agents, with a discussion paper promised soon. On October 16, 2026 Arthur Yuen said more than 60% of second-cohort sandbox use cases involved AI agents. The FSB's 12 sound practices of the same day are framed as a non-binding toolkit.

FAQ

What is the HKMA ART Framework for authorized institutions?

The HKMA ART Framework is the Hong Kong Monetary Authority's responsible-AI package built on accountability, resilience and trust. It was issued on October 15, 2026 as circular B1/15C to the chief executive of every authorized institution. Annex A sets nine expectations and supersedes the High-level Principles on Artificial Intelligence of November 1, 2019. The same circular publishes sandbox observations as Annex B and a concept note on agentic AI as Annex C. Annex A spells the middle pillar Resilience, which is also the spelling in Yuen's October 16 remarks.

Which firms does circular B1/15C actually address?

Circular B1/15C is addressed to the chief executive of every authorized institution. The expectations apply proportionately, drawing on governance the firm already has where that governance fits. A securities intermediary comes inside this package where it also holds authorized-institution status. Exchanges, fund managers and crypto-asset platforms remain in their own regimes when that status is absent. SS1/23 is a different perimeter again, covering UK internal-model firms from May 17, 2024 rather than the Hong Kong authorized-institution population as a whole.

What should the AI inventory record for each use case?

Principle 3 of Annex A says the inventory should generally cover purpose, owner, provider and model, risk classification, lifecycle and approval status, dependencies, and the key controls for risk, access and monitoring. The institution defines what counts as AI use and assigns someone to keep the record current. Internal tools, third-party tools and embedded tools all belong in the stock-take. Staff rules are expected to catch unapproved tools and approved tools used outside their approval, which is how the guidance treats shadow AI. Footnote 7 adds prompts and model versions for higher-risk generative output.

How has the HKMA changed its view of human oversight?

Footnote 3 of Annex A records the August 19, 2024 consumer-protection circular, which called for a human-in-the-loop while customer-facing generative tools were at an early stage. The October 15, 2026 guidance says a human-on-the-loop or human-in-command arrangement can fit some use cases as the technology matures. Principle 2 still requires reviewers to have competence, information, time and authority to challenge, override or stop the system, and the board remains accountable for outputs. Yuen repeated that direction on October 16, tying lighter oversight to clear accountability and ongoing monitoring.

What did the second generative-AI sandbox cohort report on agents?

On October 16, 2026 Deputy Chief Executive Arthur Yuen said more than 60% of second-cohort use cases involved AI agents, and that half of those agent use cases were multi-agent. For similar use cases, banks achieved more sophisticated outputs with around 30% less computing power than in the first cohort, and the share of multimodal use cases, covering sound, image and video, had doubled. Three participant examples were a 40% cut in onboarding time, a 98% positive-feedback rate on a wealth chatbot, and a 95% shorter market-risk analysis cycle. The agent share is the cohort-level statistic, unlike those three examples.

What remains open after the concept note on agentic AI?

Annex C sets preliminary considerations and says a discussion paper with options for possible requirements will follow soon. The ideas include intent-anchored accountability, visibility over material agents, an emergency stop, and operating boundaries described as a risk-integrated agent harness. The HKMA is exploring an industry-level registry of material agents. A close date for that paper is unstated. Further trials sit with the GenAI Sandbox++, and the announced Cyberport and University of Hong Kong data platform still has an operating date unstated.

This article is informational analysis only and does not constitute legal, regulatory, tax, or investment advice. Regulatory frameworks change frequently and interpretation depends on facts and circumstances; primary documents and official regulator guidance always supersede summaries. Firms should consult qualified legal counsel and their relevant supervisory authority before taking any action based on the analysis above.