Hong Kong’s Project SAGE targets identity and liability before AI agents move money

Before a software agent can be trusted to transfer money or act on a treasury instruction, a financial institution needs to know who authorised it and what it is allowed to do. The Hong Kong Monetary Authority launched Project SAGE on 2 November to address that governance problem. The programme’s name stands for Securing Agentic Finance with Governance and Enablement, and the HKMA has convened an expert group from financial institutions, technology providers, professional advisers and academia. It is a policy and industry-development initiative, not a rule that already licenses autonomous trading agents.[1][2]

The authority set out four workstreams: practical use-case exploration; agent identity and delegated authority; interoperability standards; and legal and regulatory responsibilities. In application terms it envisages consumer payments, business procurement and treasury management, and machine-to-machine payments for services. These are not all equal in risk. An assistant paying a small software invoice and an agent reallocating corporate liquidity across countries should not receive the same authority merely because both interact through a payment API.[1][2]

The missing identity layer in automated financial services

The identity workstream introduces the concept of Know Your Agent, or KYA. Traditional know-your-customer checks establish the identity and relevant characteristics of a natural or legal person. KYA would instead need to establish the system acting for that person, the scope of the delegation, its validity and potentially how authority can be revoked. The HKMA is exploring approaches rather than announcing a final universal KYA standard. That distinction is critical for brokers considering automated portfolio services or AI-generated trading instructions.[1][2]

Interoperability is another operational risk. A payment request may be initiated by one agent, authorised by a customer at another institution, passed through a third-party provider and then settled at a bank. Each transition can lose information about who consented, which limits apply and how an error should be investigated. Standards could help preserve those instructions. Equally, strong identity alone cannot determine who bears the loss if an agent makes a legitimate but economically damaging decision; the legal and liability workstream has to address that separate question.[1][2]

The announcement sits alongside the HKMA’s Project Ensemble tokenised-settlement programme but tackles a different layer of the system. Ensemble concerns how eligible tokenised obligations can be settled in money. SAGE asks how AI systems can be trusted to request financial actions in the first place. Reporting them as one completed automated banking platform would misrepresent both. The meaningful future tests will include revocable mandates, audit trails, approval boundaries, dispute handling and the ability to interrupt an autonomous workflow safely. The project’s current result is a governance agenda and expert group, not proven operational autonomy.[1][2]

Report an error or submit an update