Several of the biggest US banks have started shutting out AI agents that sign in with a customer's own username and password. Most want agents to read account data and nothing more, and many cannot tell an agent working for its owner from one working on its own.
A password cannot say what an agent may do
A password is all or nothing. When an agent signs in with it, the bank sees the customer. It cannot tell the agent from the person, cannot limit the agent to viewing, and cannot cancel the agent without locking the customer out too. That is why the bank's real objection is not to the agent but to the password.
A key the bank issues itself fixes all three. It can name the accounts the agent may see, say whether it may view or pay, set a cap, and be cancelled on its own. Banks reached this answer once already. After the 2015 fight over money apps, they built exactly that kind of connection, and Wells Fargo had nearly stopped the password method by 2020.
The ending of that story also holds a warning about price. In 2025, JPMorgan began charging Plaid for each data pull on its customers, after years of free access. Agents will probably meet the same thing: bank-approved access that works, with a fee somewhere in the chain.
What it changes for the work an agent does
Take a shop owner who asks an agent to check Friday's supplier invoices against her balance and pay the ones that match. With her password, this works until the bank spots the agent and shuts it out, and then she has no one to complain to, because she broke the bank's rules.
With a key from the bank, the same job splits in two. Looking at balances and invoices comes first, and it is probably available soon. Paying comes later and carries limits: only suppliers she has approved, up to an amount she sets. That is a better arrangement than the password, because a mistaken or hijacked agent can only do so much damage.
The first version already exists for ordinary customers. ChatGPT's finance tools connect through Plaid, let a customer see accounts and cannot move money, and are in preview for Pro subscribers in the US. We wrote earlier that banks would decide which assistants can move money. This news shows the method they will use: not a ban, but a key they hand out.
Who loses when the password route goes wrong
Under the US rules on electronic transfers, if you give a family member or colleague your login and they go beyond what you allowed, you are fully liable until you tell the bank to stop them. The rule does not mention AI agents, but a bank has every reason to treat one the same way, since the customer handed over the login.
So the password route is a worse deal than it first looks. If the bank blocks the agent, the work stops. If the agent makes a mistake, the loss is probably yours. For now, give an agent only a connection the bank or a connection service has set up, and if you have already given one your banking password, change it.