U.S. banks restrict AI agents’ access to customer accounts

Read-only is the compromise

The banks’ preferred boundary is clear: an agent may see account information, but it should not be able to move money or make other transactions. The article does not say how banks would enforce that distinction when an agent enters an account through a browser.

That leaves banks with a problem beyond setting permissions inside their own systems. Customers can connect agents using browser tools, while banks may struggle to stop that access on their side.

Permission is the unresolved question

Banks say customers want to use AI agents, but consider the security risks too high for now. They also say they often cannot tell whether an agent has been instructed by the account holder or is accessing the account on its own.

I think that uncertainty is more fundamental than the read-versus-transact distinction. A read-only agent still sees sensitive financial information; limiting what it can do does not establish who authorized it.

OpenAI’s U.S. financial service offers one example of a narrower scope: it can review financial data, but not execute transactions. What the announcement leaves unclear is how banks would verify a customer’s authorization across the browser-based tools people can already use. Until that is clear, restricting agents may slow adoption without resolving the question of trust.