Banks can't yet say who authorised an AI agent to touch a customer's account

Major US banks have detected third-party AI agents accessing customer accounts, and some of those agents went further than the customer had instructed, according to Axios. Several of the banks are now restricting agent access while they work out how to tell genuine customer activity from behaviour that should be stopped.

The timing is awkward. The Financial Times reported that the FCA plans to use AI to automate more of its supervisory and enforcement work, extending a programme that already covers documents, emails and interviews into authorisations, supervision and enforcement. Compliance firm Corlytics argues that banks must therefore separate legitimate customer activity from unauthorised actions, and show how controls, decisions and lines of accountability fit together.

A single blocked request shows the difficulty. An agent tries to reach an account and the bank stops it: was the agent a rogue actor, or was it doing what the customer asked? "The customer authorised it" is not a catch-all answer, Corlytics notes. Banks need to know what was authorised, for how long, across which accounts and for which actions, and whether the agent could pass instructions to other systems.

Blocking agents contains one risk but does not settle how banks will recognise, limit and monitor agents that customers expect to work. The real decisions are what an agent can see, what it can do, how permission is verified and what evidence is kept. With AI-powered supervision, firms may face faster and broader questions, and will need joined-up evidence rather than summaries written after the event.

The likely friction is mundane. The customer says, "I told it to do that." The bank says, "We didn't know it was doing that." The regulator says, "Show us how you knew." Banks that can trace a rule through to the action, the control and the accountable person will have a far stronger answer than "we have an AI policy".

What it means for Pakistan: For Pakistan, the lesson lands on the State Bank's framework for digital banks and on PVARA's licensing of virtual asset service providers: neither yet sets out how a bank should verify what a customer's AI agent is permitted to do, or what record it must keep. Pakistani banks rolling out app-based assistants, and fintechs building agent-driven payments, would be the first to face the question of who authorised an action if a customer's agent moves money without clear consent.