Financial institutions have spent decades learning how to identify customers. Hong Kong is starting to test how they might identify the artificial intelligence (AI) agents acting for them.
Hong Kong’s four principal financial regulators on Thursday (Aug. 27) announced the first group of projects admitted to the expanded GenA.I. Sandbox++, a testing program aimed at moving agentic AI from policy discussions into financial workflows. The initiative will examine customer onboarding, payments, insurance claims and customer interactions.
The regulators selected 36 use cases involving 30 financial institutions and 27 technology partners from nearly 100 proposals. Technical trials are expected to begin later this year. The program will also continue “AI versus AI” experiments in which one AI system dynamically monitors the decisions and actions of another.
One of the most consequential projects for payments comes from HKT Payment, the financial services arm of Hong Kong telecommunications company HKT. The company said it will work with Red Date Technology to test an “Agentic ID” framework for registering and verifying AI agents that initiate payments, wallet top-ups, peer-to-peer transfers and transactions between financial institutions.
The framework would use decentralized identifiers and verifiable credentials to bind each agent to a verified person or business. Its aim is to reduce impersonation and unauthorized activity while creating a clearer record of what the agent did and whose authority it carried.
That addresses a growing gap in financial identity controls.
A bank may already know that Jane Smith is Jane Smith. When software attempts a transaction for her, however, the institution must also determine whether that particular agent belongs to Smith, whether she authorized it to take that action and whether its permission remains valid at the moment of payment.
The chain could look like this: customer, agent identity, authority and transaction. Each link would need to hold before money moves.
For banks, wallets and payment providers, that structure could turn an AI agent into a recognizable software identity with its own operating limits. A customer or business might allow an agent to buy only from approved merchants, access a designated account, stay below a transaction ceiling or operate within certain locations and dates.
Those controls would give financial institutions more precision than a simple choice between granting broad account access and blocking the agent altogether. If a credential were compromised, a provider could revoke the agent’s authority without closing the customer’s account or replacing every payment credential tied to it.
Agent identification could also produce better evidence when transactions are disputed. Today, an institution typically asks whether the customer or device was authenticated. An agentic payment record could add three useful facts: which agent acted, which principal it represented and what instructions or limits governed the transaction at that time.
The Hong Kong trials won’t settle how agent identity should work across banks, wallets and borders. Standards would still need to align, credentials would need protection and institutions would need rules for liability when an authorized agent exceeds its instructions.
Still, the sandbox moves the issue closer to the payment itself. As AI agents gain the ability to spend, transfer and manage money, knowing the customer will cover only half the identity job. Financial institutions will also need to know the agent.
