Third-party AI agents are knocking on the doors of customer bank accounts just as regulators begin turning AI on the firms they oversee. For banks, the challenge is no longer simply spotting suspicious activity. It is proving where authority began, where it ended and who was accountable in between.
According to Axios, several major US banks have detected outside AI agents accessing customer accounts, with some reportedly trying to go beyond what users had instructed. A number of lenders are now restricting access while they work out how to separate genuine customer activity from actions that should be stopped, detailed Corlytics.
At the same time, the Financial Times reported that the FCA intends to automate more of its supervisory and enforcement work using AI. The regulator already applies the technology to documents, emails and interviews, and plans to extend it across authorisations, supervision and enforcement.
Aretha Franklin’s “Who’s Zoomin’ Who?” captures the dynamic neatly. The customer’s agent watches the bank, the bank watches the agent, and the regulator watches the bank. Yet observation is not the same as control.
An agent may well be acting on a customer’s behalf, but “the customer authorised it” offers little protection on its own. Banks need clarity on the scope of that permission: which accounts, which actions, for how long, and whether the agent can delegate tasks to other systems. No login screen can answer those questions.
Blocking an agent may also create problems. The bank could be stopping a malicious actor, a tool the customer genuinely wanted, or an agent cleared only to read data rather than move funds. Rather than a binary choice between access and denial, banks must define what agents can see and do, how permissions are verified, and what evidence is retained when things go wrong.
That evidence is becoming critical. If supervisors can analyse far more information with AI, firms may face faster and broader scrutiny. AI-driven oversight will not be flawless, nor will it replace human judgement, but it raises the bar for banks to justify decisions with records rather than retrospective summaries.
Crucially, those records must connect. What did the customer request? What did the agent attempt? Which control applied, and who reviewed the outcome? When the answers sit across disconnected systems and teams, a bank may hold every record yet still fail to explain what happened.
The realistic risk is less dramatic than a rogue agent draining accounts. It is the everyday scenario in which the customer says, “I told it to do that.” The bank replies, “We didn’t know it was doing that.” And the regulator responds, “Show us how you knew.”
AI agents do not create these gaps, but they make them harder to hide. Banks able to trace a rule through to the action, the control and the accountable individual will have a far stronger answer than “we have an AI policy”.
Read the full Corlytics post here.
Copyright © 2026 RegTech Analyst
Copyright © 2026 RegTech Analyst
