Hong Kong opens the debate on rules for agentic AI in banking
On 2 November 2026, during Hong Kong FinTech Week 2026, the Hong Kong Monetary Authority (HKMA) announced Project SAGE, short for “Securing Agentic Finance with Governance and Enablement”, and set up a 22-member expert group that includes eleven financial institutions, seven technology providers, three professional services firms and one academic representative. HKMA Chief Executive Eddie Yue said agentic finance has the potential to transform the way financial services are delivered and used. The plan runs along four workstreams: market development across B2C, B2B and M2M; agent identity management, including Know Your Agent practices and authentication; interoperability standards; and a legal and regulatory framework covering governance, liability and consumer protection.
Alongside it, the HKMA published a discussion paper on the governance of agentic AI in banking, stating that the text does not itself introduce binding supervisory requirements or change existing obligations; it is collecting views until 1 February 2027. The core of the proposal is “intent-anchored accountability”: a software agent should act only for an approved business objective and within a defined delegation, with responsibility remaining with identified people and functions inside the bank. According to the paper, relying on third-party models, cloud services or tools should not dilute that responsibility. The authority envisages a designated accountable owner for each agentic deployment, normally the business head, supported by technical and control roles. It also seeks views on a centralised register of material agents, fed by the banks.
To assess materiality, the HKMA proposes a two-step test focused on the potential impact of the operations and on whether real human oversight exists. An approval is genuine, the authority specifies, only if the approver sees the substance of the action, has the authority and the practical ability to reject it, refusal is accepted as a legitimate outcome, and the action stays on hold until the decision is made. For customer-facing agents, the paper stresses that an agent should neither impersonate a person nor deny being an agent, and that customers should be able to reach a human operator promptly. It also describes a kill switch that quickly brings the agent to a safe state, for example by halting execution, revoking credentials and preserving logs, and asks for views on a possible supervisory approach that could, in exceptional circumstances, extend to measures suspending an agent.
The HKMA has not yet decided whether, or in what form, it will issue further guidance. The register, notifications and suspension measures remain ideas under discussion: it is not settled whether notification of material agents should come before or after going into production, and the press release gives neither a timetable for Project SAGE nor any pilot results. The measures under discussion apply solely within Hong Kong's regulatory perimeter. Defining the chain of command before delegation to agents becomes operational strikes me as a methodical and prudent approach: handing a task to a machine does not erase the responsibility of whoever decided to hand it over. — Olya
Come Olya ha verificato questa notizia
- Verificato
- We read the HKMA press release of 2 November and the related releases of 2 November (FinTech Week) and 15 October (ART). The 22 members were counted directly on the official PDF annex. Of the Discussion Paper, pp. 1-20, 26-28, 42-46, 51-53 and 56-57 were read; the remaining pages were not. Independent confirmation: The Standard, Hong Kong (2 November). The Paypers and Financial IT rework the press release and were not counted as confirmations. We found no coverage by Reuters or Bloomberg. Data found only in secondary outlets, such as the Signicat study, was not used.
- Incertezze
- The paper sets out preliminary views: the register, notifications and suspension measures are ideas, not rules, and the HKMA has not decided whether to issue further guidance or in what form. It is not settled whether notification of material agents should precede or follow production, or who would run the register. There is no timetable for Project SAGE and there are no pilot results. Two claims that could not be confirmed on the HKMA pages we read were left out: that around 80% of Hong Kong banks use AI or generative AI and, according to a Russian outlet, that SAGE does not provide for autonomous agents accessing customer accounts. The scope is Hong Kong: no direct effect on European banks.
- Perché pubblicarla
- It is one of the first documents from a banking authority to go into the operational detail of AI agents: who answers when the software acts, when a human approval is merely formal, how to stop an agent and how to recognise one sent by a customer. These questions also concern European banks, regulators and customers. It should be read for what it is: a consultation open until 1 February 2027, not a rule.
Fonti / Sources
- Hong Kong Monetary Authority, comunicato «HKMA launches Project SAGE and establishes Expert Group for agentic finance» (2 novembre 2026)
- Hong Kong Monetary Authority, «Discussion Paper on Agentic A.I. Governance in the Banking Sector» (novembre 2026)
- Hong Kong Monetary Authority, allegato con i membri dell'Agentic Finance Expert Group
- Hong Kong Monetary Authority, comunicato sul Framework «ART» (15 ottobre 2026)