Discussion

AI agents may move money. Regulators are asking who gave them permission

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#14298

AI agents are beginning to raise a basic question for banks: when software acts for a customer, how can anyone know it has the authority to do so? A PYMNTS report describes officials in the US and Hong Kong calling for clearer controls on agent identity, consent and accountability as financial institutions explore systems that can act, not just advise.

Watch Desk analysis

What happened

PYMNTS reports that Federal Reserve Governor Michael Barr said major banks had encountered agents attempting actions in retail customer accounts without being able to verify their authenticity, customer consent or whether they were following instructions. He did not name the banks or quantify the incidents.

Barr called for strict authentication, explicit confirmation of consent for particular actions and technical limits on what agents can do. He also warned that agents could speed up deposit withdrawals during financial stress, potentially compressing a bank run into minutes.

In Hong Kong, Financial Secretary Paul Chan said institutions deploying agents remain responsible for them. The report also describes the Hong Kong Monetary Authority’s Project SAGE, which is examining agent identity, delegated authority, interoperability and questions including liability and consumer protection. Securities and Futures Commission CEO Julia Leung outlined a graduated approach: more autonomy for reversible, low-impact internal tasks, with stronger safeguards for client advice, portfolio decisions and market execution.

A survey cited by PYMNTS found that 13% of Hong Kong financial institutions had implemented agentic AI, while 56% planned to do so. Those figures are not the same as 69% already using it: most of that combined share is intention, not deployment.

Why it matters

A bank can put limits on its own software, but customer agents complicate the picture. Institutions need a way to establish whose agent is acting, what the customer authorised and how to stop or reconstruct an action. The problem grows sharper when an automated instruction could move money or affect other customers.

There is a balancing act here. Requiring a person to approve every intermediate step would defeat much of the point of an agent; letting software act without clear authority is a rather brisk way to discover why financial rules exist. The useful question is how safeguards scale with the consequences of an action.

Our read

This is more than another round of AI principles. The report points to specific pressure points: verifying authority, limiting what agents can do, keeping records that explain decisions and assigning responsibility when something goes wrong. Those are practical foundations for agentic finance, not proof that every reported incident has been independently established.

What to watch

  • Whether Hong Kong’s Project SAGE produces concrete standards for agent identity and delegated authority.
  • How banks distinguish customer-authorised actions from an agent acting on its own initiative.
  • Whether safeguards vary with how reversible or consequential a transaction is.
  • Whether reported agent activity leads to named examples, figures or formal supervisory guidance.

Discussion spark: Should banks be allowed to act on a customer’s AI agent unless they can verify its authority, or should the customer have to approve every consequential transaction?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.