You gave an AI agent your bank password. Is what it does “unauthorised”?

When a bank tells a customer that an AI agent's actions are the customer's own, it is making a legal claim about a single word. In US consumer banking, almost everything about who pays for a bad transfer depends on whether the transfer was authorised. AI agents that sign in with a customer's password sit right on the line.

This explainer walks through the rules as they are written, the official interpretation, and the regulator's guidance on fraud. It ends with the questions no one has answered yet. It is a reading of public sources, not legal advice.

The rule: Regulation E and unauthorised transfers

Regulation E implements the Electronic Fund Transfer Act. Its central protection is a cap on what a consumer pays for an unauthorised electronic transfer. Under 12 CFR 1005.6, a consumer who reports promptly is liable for at most $50; liability rises to as much as $500 for later reporting, and can be unlimited for transfers made more than 60 days after a statement showing an unauthorised transfer that the consumer never reported.

Everything therefore depends on the definition. Under 12 CFR 1005.2(m), an unauthorised electronic fund transfer is one “initiated by a person other than the consumer without actual authority to initiate the transfer and from which the consumer receives no benefit.” Three kinds of transfer are excluded. The first is the one that matters here:

A transfer initiated “by a person who was furnished the access device to the consumer's account by the consumer, unless the consumer has notified the financial institution that transfers by that person are no longer authorized.”

An “access device” is a card, code or other means of access that can be used to start transfers. The official interpretation lists PINs and telephone transfer codes as examples; an online banking username and password is the obvious modern equivalent.

When the person you trusted goes further than you said

The official interpretation has a comment, headed “Authority”, written long before AI agents existed. If a consumer furnishes an access device and grants authority to someone who then exceeds it, “the consumer is fully liable for the transfers unless the consumer has notified the financial institution” that transfers by that person are no longer authorised (comment 2(m)-2).

Read naively onto agents, that is a hard result for the customer. You gave the agent your password and told it to pay one bill. It paid three. Under this comment, the extra two look like transfers by someone you furnished the access device to, who exceeded their authority, and they are yours until you tell the bank otherwise. This is also the shape of Wells Fargo's new terms, which make the customer “solely responsible” for “the AI Agent exceeding the authority granted by you” (section 15(e)).

When someone else tricks you, or your agent

The CFPB has drawn a clear line for one related case. In its Electronic Fund Transfers FAQs, the bureau says that when a consumer is “fraudulently induced into sharing account access information with a third party,” a transfer that third party makes is “an unauthorized EFT under Regulation E.” Such a consumer “has not furnished an access device,” so the exclusion does not apply. And “negligence by the consumer cannot be used as the basis for imposing greater liability” than the regulation permits.

Now combine the two. A customer willingly gives their password to a legitimate agent. A stranger then plants instructions in a web page or email, the agent reads them, and it sends money to the stranger. Is that the agent exceeding authority (customer liable), or a third party fraudulently obtaining the use of the access device (bank liable)? The FAQ was written for people being tricked, not software. Nobody has ruled on software yet.

The map, as of October 2026

The same line, seen from the UK

The UK has spent years on the difference between authorised and unauthorised fraud, because the money involved is large on both sides. UK Finance's 2026 fraud report counted almost £1.3bn of losses in 2025 and £576.4m in authorised push payment fraud across 248,070 cases, where victims are manipulated into sending money themselves. Unauthorised fraud, where the customer did not approve the payment, accounted for £703.4m.

Agents add a third category the existing split was not built for: payments the customer authorised the agent to make in general, but did not approve in particular. Every jurisdiction will have to decide which side of the line those fall on.

Why the word matters to banks, not just customers

It would be easy to read bank terms like Wells Fargo's as shifting risk away from banks. In practice they shift the argument to evidence. If a customer disputes a transfer, the bank will have to show what happened: was the customer in control of the session when the money moved, or was software acting on its own? Bank of America's Brian Moynihan has said that when a customer gives an agent their credentials, “it is you authorizing the payment”, as reported by USA TODAY. That holds only as long as the bank can show the payment came from inside the authority the customer gave.

The practical lesson is the same for both sides. The most valuable fact in a dispute about an agent is whether a person approved this specific action, at the moment it happened. Most banking systems were never built to record that fact. That is the gap HEIP is built to close: a check, at the moment of a sensitive action, of whether a person is in control, with a record of the decision and its reasons.