HKMA Project SAGE is the Hong Kong Monetary Authority’s new governance programme for autonomous AI agents operating in financial services. Launched at Hong Kong FinTech Week 2026 on November 2, it is the first formal banking-regulator programme in the world focused on agents that can move money, negotiate contracts and instruct other AI systems without human approval at every step. The launch accompanied a discussion paper seeking industry feedback on how banks should govern these systems, including what a practical Know Your Agent, or KYA, standard might require.
The initiative matters because agentic AI is not simply a faster version of a chatbot. A conventional generative AI tool can draft an email, summarise a document or suggest an answer, after which a person remains in the loop. An agent can instead receive a goal, plan a sequence of actions, call payment or database systems, observe the result and revise its next move. In a multi-agent design, one instruction may trigger a chain of actions across several institutions, currencies and time zones.
What HKMA Project SAGE Covers
Project SAGE, short for Securing Agentic Finance with Governance and Enablement, is organised around four areas of work. The first covers market development and use-case exploration, including business-to-consumer agentic commerce and everyday payments, business-to-business procurement and treasury management, and machine-to-machine payments for digital services. The second examines agent identity and authorisation. The third considers interoperability standards, while the fourth addresses legal and regulatory questions involving governance, liability and consumer protection.
The Hong Kong Monetary Authority also established a 22-member Agentic Finance Expert Group drawn from financial institutions, technology service providers, professional service firms and academia. Members include Bank of China (Hong Kong), HSBC, Standard Chartered Bank (Hong Kong), Visa and Mastercard. The group is intended to test practical questions with the industry rather than leave design to a rulemaking exercise conducted in isolation.
Eddie Yue described the programme as a way to support useful applications while reinforcing governance and safeguarding financial stability. That balance is essential. Agentic systems could reduce friction in payments and corporate treasury operations, but wider adoption also creates questions about authority, evidence and intervention that conventional software testing cannot fully answer.
Why Know Your Agent Is Different
The identity pillar is the most novel part of HKMA Project SAGE. Know Your Customer and Know Your Business controls were designed for natural persons and registered legal entities. A person may have a passport, biometrics and an address. A company has registration records, ownership information and regulatory permissions. A software agent has none of those attributes in the ordinary legal sense. It can also be copied, updated or withdrawn by changing software rather than replacing a document or transferring a company.
The discussion paper asks what a practical KYA standard could require. It discusses cryptographically signed identity certificates that attest to an agent’s authorised scope, disclosures of the capabilities it possesses, and a centralised registry of material AI agents maintained by the authority. For each listed agent, the proposed registry could capture its purpose and deployment status, business, technical and control owners, capabilities, data leaving the system, infrastructure dependencies, stop mechanisms, log locations, and change and suspension history.
These are consultation proposals, not settled requirements. The Hong Kong Monetary Authority has expressly not decided what a KYA standard must contain. It has not announced a mandatory agent identity standard, a licensing regime or a binding registry requirement. The paper creates no new obligations and does not modify existing ones. First-round written responses are due to the Banking Supervision Department before February 1, 2027.
Authority Needs More Than a Human Approval Step
One of the paper’s central warnings is that formal approval by a person does not automatically amount to meaningful human oversight. A manager might click through many automated actions without understanding the underlying chain, particularly where an agent can negotiate terms, call external tools or instruct other agents. The paper therefore considers human-in-the-loop, human-on-the-loop and human-in-command models, with the degree of intervention matched to the risk.
The proposed intervention mechanisms extend beyond a simple emergency shutdown. The Hong Kong Monetary Authority describes stopping task execution, revoking access or credentials, disabling connections to external tools, pausing outbound communications, activating a fallback process, and preserving logs and evidence. A useful control therefore combines prevention, containment and forensic capability. It must be able to act before a faulty instruction travels through an automated chain, while leaving a record that can support investigation.
Project SAGE also develops the idea of intent-anchored accountability. An agent would operate only to achieve an approved business purpose and remain within a defined delegation of authority. Responsibility would stay with named individuals and organisational functions rather than being transferred to the software itself. That approach attempts to preserve a clear line of accountability in systems where one agent may have received instructions from another agent.
Evidence From the Sandbox and Existing Market Work
The programme follows the authority’s practical experimentation with the GenAI Sandbox++. In the previous GenAI Sandbox round, more than 60 percent of use cases involved agentic AI rather than basic generative AI. The expanded programme received nearly 100 proposals and selected 36 use cases involving 30 financial institutions and 27 technology partners. Some pilots also examine using one AI system to monitor and assess the actions of another, reflecting the operational reality of multi-agent systems.
Private companies are already exploring technical components of agent identity. Baselayer raised 35 million dollars in 2026 for work on credentialing systems, while Skyfire has engaged with KYA as a compliance category. FATF working groups have also examined the issue. These efforts do not amount to regulatory consensus, but they show that banks and technology providers are beginning to treat agent identity as an operational requirement rather than a purely theoretical question.
Project SAGE also builds on the Accountability, Resilience and Trust framework announced by the Hong Kong Monetary Authority on October 15. It further connects with the authority’s Fintech 2030 DART strategy, covering data and payment, artificial intelligence, resilience and tokenization. That lineage suggests the programme is part of a broader supervisory direction rather than a standalone technology pilot.
Consumer Acceptance May Shape Adoption
Industry readiness does not mean customers are ready to hand control to software. Research published by Signicat found that 79 percent of financial services providers were prepared to handle applications initiated by AI agents, while only 14 percent of consumers were comfortable with AI opening a financial product on their behalf. The study was conducted by Sapio Research in August 2026 and surveyed 4,050 consumers and 450 financial services professionals across nine European countries. Signicat has a commercial interest in digital identity and fraud prevention, so its findings should be considered in that context.
The same research found a divide between assistance and delegation. About 65 percent of consumers said they would welcome AI helping with applications, while 11 percent wanted AI to complete an application for them. Nineteen percent would allow AI to make smaller purchases, compared with 8 percent for high-value purchases. Just 15 percent would trust AI to use their digital identity for verification, and 14 percent would trust an agent to sign contracts when opening an account.
These figures suggest that governance cannot rely only on technical controls. Customers may need clear disclosure about whether a human or an agent initiated an action, what authority the agent holds, how they can withdraw that authority and where records can be reviewed. The findings also strengthen the case for proportionate delegation and meaningful intervention, rather than presenting every agentic system as if it performs the same function.
Questions Still Open
Several hard questions remain. A registry could improve visibility, but it would need reliable ownership data, a workable definition of a material agent and rules for agents that span multiple jurisdictions. Identity credentials could reveal authorised scope, but they may not prove that an agent followed its instructions. Interoperability may widen adoption while making delegated chains harder to reconstruct.
Liability is no less difficult. If a bank deploys an agent, the bank cannot transfer responsibility to a third-party developer merely because the software selected an action. Conversely, developers, model providers, infrastructure operators and business customers may each control different parts of the environment. The consultation will test whether common principles can reduce that fragmentation without fixing technical architecture prematurely. The HKMA Project SAGE programme is a consultation rather than a rulebook, and no deadline binds anyone yet. Even so, the HKMA Project SAGE consultation has already done something unusual: it has made the identity and authority of an autonomous banking agent a supervisory topic in its own right, before the first enforcement question arrives.
HKMA Project SAGE therefore matters less as a finished regulatory regime than as an early attempt to shape one. Its immediate task is consultation, experimentation and shared vocabulary, not compulsory rules. For banks, the message is practical: identity, delegated authority, intervention and evidence should be designed before autonomous finance reaches the market at scale. For consumers, the larger question is whether that governance can earn trust without making useful services too slow or rigid.