Project SAGE asks banks to name an owner before an agent acts

Project SAGE asks authorized institutions in Hong Kong to name who owns an artificial intelligence (AI) agent that can plan and act, and to say which of those agents are material enough for the Hong Kong Monetary Authority (HKMA) to see.

On November 2, 2026, the HKMA opened Project SAGE, short for Securing Agentic Finance with Governance and Enablement, and published a Discussion Paper on Agentic AI Governance in the Banking Sector. Section 2.3 records the paper as preliminary views that invite comment: the document creates no new obligations and leaves existing obligations standing, with responses due before February 1, 2027. The decisions that follow are what banks are asked to settle beyond the October supervisory guidance.

Key facts

• November 2, 2026: Project SAGE and the discussion paper were issued together (HKMA press release; discussion paper PDF).
• Comments go to the Banking Supervision Department at all-banks-go-fintech@hkma.gov.hk before February 1, 2027, and the HKMA may publish a submission unless the respondent asks for a name or a passage to be withheld (section 10.2).
• The paper describes the October 15, 2026 supervisory guidance as already requiring lifecycle governance, a dedicated owner for each AI use case, and an inventory of approved use cases (supervisory guidance PDF).
• The Agentic Finance Expert Group has 22 members across banks, payment firms, cloud providers, professional firms and academia (annex).
• On July 22, 2025 the Monetary Authority imposed pecuniary penalties of HK$8.5 million (Hong Kong dollars), HK$4 million and HK$3.7 million under section 21 of the Anti-Money Laundering and Counter-Terrorist Financing Ordinance (Cap. 615) (AMLO) (disciplinary release).
• The paper cites the Financial Stability Board (FSB) report of October 15, 2026, which sets out 12 sound practices for responsible adoption of artificial intelligence (FSB report).

Methodology and sources

Primary documents are the November 2, 2026 press release, the discussion paper and the expert-group annex. The October comparison uses the paper’s account of the supervisory guidance and concept note of October 15, 2026, with the guidance PDF and the October 15, 2026 announcement. Enforcement is the July 22, 2025 release. Comparators are the FSB report, Bank of England Supervisory Statement 1/23 and European Union Regulation (EU) 2024/1689. The window is October 15 to November 2, 2026, with 2025 penalties as the sanction already available. Scope is Hong Kong authorized institutions. The EU and United Kingdom rows show a duty or an expectation that a cross-border group can already face for a neighbouring use of AI, while Project SAGE stays separate from the questions put to banks.

Related:Thai SEC draft lets one company seek derivatives licences

What banks are asked to decide beyond the October supervisory guidance

The October supervisory guidance, on the November paper’s account, already told authorized institutions to govern AI in proportion to risk across the adoption lifecycle, to name a dedicated owner for each AI use case, and to keep an inventory of use cases approved for deployment. A concept note the same day recorded early sandbox observations. The November paper asks banks to settle the agent-specific choices those instruments left open, in a response due before February 1, 2027, as evidence for any later guidance.

Section 3.1 describes agentic AI as a system that plans steps, uses tools, keeps memory and adapts. The operational test is a system with discretion over its actions, including where approval is nominal because volume outruns review. Prescribed automation, a single fixed task, and a drafting assistant that only produces text for a person to send sit outside it. Banks are asked to flag boundary cases, such as queued messages that send themselves or a vendor feature switched on, and question 3B asks which use cases will matter first.

Intent-anchored accountability, in the November 2, 2026 discussion paper, means an authorized institution writes down the business intent an agent may pursue before that agent takes a material action, and leaves accountability with a named person inside the institution. The record covers the objective, the tasks and limits, the data and tools, any data leaving the firm and its recipients, the oversight expected, and when authority is withdrawn. Section 5.3 builds on the October supervisory guidance, which already required a dedicated owner for each use case, and asks whether that owner should normally be the business owner, with separate technical and control role-holders. The control role stays appropriately independent of design, while technology staff keep the build, the logs and the stop mechanism. The practical test is whether the institution can name who answers, for what, and whether that person can withdraw the deployment.

Section 7.1 asks banks to encode risk appetite in an agent harness of permissions, tool lists, deterministic checks and velocity limits, because a prompt can shift once an outsider’s document is in context. A failed check blocks a high-impact action or sends it to a person, and the harness is where the agent is stopped from adding tools at runtime. Oversight counts where the approver sees the substance and has time to reject it. The stop mechanism, also called a kill switch, is meant to be rehearsed at real speed, and question 7C asks what the pre-production record costs when a vendor changes a tool.

How Hong Kong, the EU and the United Kingdom treat an agent that acts

Sources: HKMA discussion paper, November 2, 2026; HKMA disciplinary release, July 22, 2025; Regulation (EU) 2024/1689; PRA SS1/23. The table was last checked on November 3, 2026.

The materiality test in section 8.1 would decide which agents an authorized institution notifies to the HKMA, beyond the October inventory of approved use cases. Step one asks about significant reach over positions, obligations or operations, using thresholds the bank already keeps and the time needed to stop the agent. Positions include moving funds or deciding credit. Obligations include closing a regulatory alert. Operations include a critical process whose accuracy depends on the agent. Any one of those carries the test on, and a drafting tool that lacks each kind of reach stops. Step two asks whether genuine per-action approval stands in the path. That approval can place the deployment below the line, except for an irreversible action such as payment finality, so a gated wholesale release stays material, and shared platforms are judged together.

The other columns have partly closed the question Hong Kong is still asking. Since August 2, 2026 an EU creditworthiness system for a natural person sits in Article 6(2) and Annex III, point 5(b), fraud detection aside, with Article 26 deployer logs and Article 99(4) fines up to €15 million or 3% of turnover. UK SS1/23, current text effective April 23, 2026, expects model-risk governance for artificial intelligence used in modelling and stays silent on a harness, a Know Your Agent (KYA) check, and a filing of material agents. The same credit assistant can be a high-risk EU system, a London model, and inside or outside Hong Kong’s section 3.1 depending on whether it acts through tools.

“Agentic finance has the potential to transform how financial services are delivered and utilised. By introducing a series of initiatives and facilitating exchanges with industry experts, we aim to foster innovation while reinforcing Hong Kong’s leading role in next generation of financial technology. These initiatives will help us navigate the opportunities and challenges ahead, ensuring that agentic finance serves the best interests of users while safeguarding the stability of our financial system.”

— Eddie Yue, Chief Executive, Hong Kong Monetary Authority (HKMA, November 2, 2026)

Why a failed transaction monitor already carries an AMLO penalty

Exhibit 2 treats an agent that closes financial-crime alerts on its own authority as material, because the closure discharges a regulatory duty. On July 22, 2025 the Monetary Authority reprimanded Indian Overseas Bank, Hong Kong Branch, ordered a look-back, and imposed HK$8.5 million, plus HK$4 million on Bank of Communications (Hong Kong) Limited and HK$3.7 million on Bank of Communications Co., Ltd., Hong Kong Branch, all under section 21 of the AMLO. The findings were weak continuous monitoring: gaps in oversight at Indian Overseas Bank, Hong Kong Branch, and transaction types left out of a shared system at the two Bank of Communications entities.

The discussion paper would put an agent inside that same duty. Closure of alerts on the agent’s own authority is material on obligations reach under section 8.1, and a read-only fraud monitor that is the sole control for a critical operation stays material because a signature leaves a silent failure in place. The July 22, 2025 penalties are the published cost of a monitor that misses a class of transactions, and an agent that repeats the miss at machine speed would face the same section 21 power. Section 8.3 asks whether acting beyond authority, a failed stop mechanism, compromised credentials, a monitor that fails silently, or a gate that lets an action through should count as incidents. Question 4A asks for near-misses, which the HKMA says are still thin.

What the questions mean for banks, payment firms and compliance teams

The near-term product for an authorized institution is the response due before February 1, 2027. Question 5A’s preliminary view builds on the existing framework. A bank with the October owner and inventory is asked whether that pair holds once an agent can choose tools and outrun the queue: the business owner stays accountable, a technical role-holder runs the harness, and a control role-holder challenges both. On a less material deployment, question 5D asks whether one person can hold the first two roles if control stays independent. Retirement still has to clear leftover credentials and memory.

Project SAGE seats firms outside the paper’s perimeter with the banks that are inside it. The annex places Bank of China (Hong Kong), Standard Chartered Bank (Hong Kong) and The Hongkong and Shanghai Banking Corporation alongside Visa, Mastercard, Adyen, Amazon Web Services, Google Cloud, Circle Hong Kong, Clifford Chance, PwC and Professor Allen Huang. They meet the paper when a bank accepts their agent. Section 9 keeps the bank accountable for its channel, and question 8D asks whether shared-model concentration belongs on a central register or an industry one.

Section 9.2 treats a fund or corporate agent as a channel, and Know Your Agent separates identity from a revocable mandate, so a signature still leaves the instruction to be checked against account limits. Hong Kong’s one-time-password reset for operational security is a neighbouring operational-security measure for virtual-asset platforms. Question 8B fits post-deployment notice to most cases and stops short of a general pre-approval. Supervisory data clocks that split hours between banks and authorities show the same argument over who holds the clock.

“Effective transaction monitoring enables timely identification and reporting of suspicious transactions and thus is an essential component of banks’ AML/CFT controls. Bank management should ensure that proper transaction monitoring systems and processes are in place and any identified deficiencies are addressed promptly.”

— Raymond Chan, Executive Director (Enforcement and AML), Hong Kong Monetary Authority (HKMA, July 22, 2025)

What Project SAGE will test before any rule is settled

Project SAGE is the enablement track, separate from the discussion paper’s questions to authorized institutions. Opened in the same press release, it lets the market try use cases while governance is still under discussion. Four workstreams cover market development, agent identity, interoperability and legal frameworks, with lanes for consumer payments, corporate treasury and machine-to-machine charging. Agent identity is the industry face of KYA, which is why payment networks sit with three Hong Kong banks, Clifford Chance and PwC.

Responses will be read with sandbox trials, including the generative-AI cohort announced on August 27, 2026, before further guidance is shaped. A SAGE convention still leaves the bank to decide whether to honour an instruction, including where commerce meets the licensing split for stablecoin issuers. By February 1, 2027 the open items are the section 3.1 boundary, the ownership split and the choice of register, while section 21 of the AMLO remains the sanction when a monitor fails.

TL;DR

Project SAGE and the HKMA discussion paper on agentic AI in banking were both issued on November 2, 2026, with comments due before February 1, 2027. Section 2.3 frames the paper as preliminary views, so any binding guidance would be a later instrument. Banks are asked to define which systems count as agents, to anchor each one to a business owner with technical and control role-holders, to enforce limits in an agent harness, and to apply a two-step materiality test covering positions, obligations and operations, including irreversible payments. A 22-member expert group runs the parallel work on payments, identity and liability. On July 22, 2025 the Monetary Authority fined three banks HK$8.5 million, HK$4 million and HK$3.7 million under section 21 of the AMLO for transaction-monitoring failures, the duty an alert-closing agent would sit inside.

Frequently asked questions

What is Project SAGE, and how does it differ from the discussion paper?

Project SAGE, Securing Agentic Finance with Governance and Enablement, is the HKMA industry project launched on November 2, 2026. Four workstreams cover use cases, agent identity, interoperability and legal questions, including consumer payments, corporate treasury and machine-to-machine charging. A 22-member expert group of banks, payment firms, cloud providers, professional firms and an academic supports that work. The discussion paper issued the same day is a separate document. It asks authorized institutions how they would govern their own agents and instructions arriving from a customer’s agent, and it wants those views before February 1, 2027.

What is the legal status of the discussion paper?

Section 2.3 presents the paper as preliminary views and an invitation to comment. Further supervisory guidance, if the HKMA writes it, would be a later instrument shaped by the responses, sandbox trials and international work, including the FSB’s 12 sound practices of October 15, 2026. Submissions go to the Banking Supervision Department before February 1, 2027, and the HKMA may publish them unless a firm asks for a name or a passage to be withheld. Until any later instrument, operative duties remain the Banking Ordinance, the AMLO, the Supervisory Policy Manual and the October supervisory guidance.

What did the October supervisory guidance already require?

On the November paper’s description, the October 15, 2026 guidance already expects lifecycle governance, a dedicated owner for each AI use case, and an inventory of approved use cases. Any registry would sit on that inventory. Fresh decisions are the agent boundary, the business, technical and control split, the harness, genuine approval, two-step materiality, possible notice to the HKMA, and Know Your Agent checks. The October concept note recorded sandbox observations and stopped short of this list.

Which banking agents would count as material?

Section 8.1 assesses significant reach over positions, obligations or operations, using thresholds the bank already keeps, across the time needed to stop the agent. A research assistant that lacks each kind of reach drops out at step one. Closing financial-crime alerts on the agent’s authority is material, because the closure discharges a duty. A wholesale payment release stays material even with a human approval on every payment, because settlement can be irreversible. Agents that share a model can be material as a group.

What does Know Your Agent ask of a bank?

Section 9.3 describes Know Your Agent as a risk-based check, scaled to the interaction, and as distinct from a single technical standard or from customer due diligence. The bank would establish identity over time, the operator, the principal, who granted the mandate, and the mandate’s scope, duration and limits. A credential shows identity and leaves the further question of whether this instruction sits inside the mandate. Authority has to be revocable if it expires, if the customer withdraws it, or if the credentials are compromised.

Who sits on the Agentic Finance Expert Group?

The annex of November 2, 2026 lists 22 members. Financial members include Bank of China (Hong Kong), Standard Chartered Bank (Hong Kong), The Hongkong and Shanghai Banking Corporation, Visa Hong Kong, Mastercard, UnionPay International, Adyen, Ant International and WeChat Pay Hong Kong. Technology members include Amazon Web Services, Google Cloud, Tencent Cloud and Circle Hong Kong. Professional firms are AIFT, Clifford Chance and PwC. Academia is Professor Allen Huang of the Hong Kong University of Science and Technology Business School. A seat supports Project SAGE, and each authorized institution still files its own response to the discussion paper.

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.