In October, JPMorgan Chase said it would not allow outside AI agents to access customer accounts and act on customers’ behalf. An agent can still read a balance through an approved data connection, but it cannot move money or pay a bill. Citi has taken a similar line.
Wells Fargo is preparing to go the other way. Its updated online access agreement, effective November 24, sets out what customers are responsible for “if and when” the bank permits AI agents. In short, customers accept the risk, and transactions made through their agents are treated as authorized unless the law says otherwise.
It would be easy to read these as narrow decisions about passwords and liability. I think they are the first visible moves in a bigger contest over who the customer deals with first when money is involved. For most of banking’s history the answer was the bank, first through the branch, then the website, then the app. AI agents that can compare, decide and pay on a customer’s behalf could change that answer.
I have watched a version of this happen before.
The last time banks lost the front end
UPI, India’s real-time payment system, was built as open infrastructure. Approved third-party apps could plug into it, and banks could build their own front ends too. Most banks did, and most customers went elsewhere.
Today PhonePe and Google Pay together handle close to 80 percent of UPI’s more than 20 billion transactions a month. The money still moves between bank accounts. The customer simply opens someone else’s app to move it.
I spent four years at Paytm while this shift was happening, and the lesson I took from it was simple. Whoever owns the daily habit gets the first chance at every product that follows. The payment apps used that habit to sell loans, insurance and investments, with banks and other lenders often supplying the credit. The lenders carried the risk on their balance sheets while the apps kept the relationship.
AI agents could take this pattern further, and to every market. A payment app sits on top of a payment. An agent can sit on top of a customer’s whole financial life, including the bank’s own app, deciding which card to use, where to keep savings and when to pay a bill.
Agents are arriving faster than the rules
Customers are already bringing money questions to AI. In a NerdWallet survey conducted by The Harris Poll in June, 26 percent of US adults said they had used an AI chatbot to answer personal finance questions. The step from asking to acting is being built now. Visa and Mastercard both run programs that let AI agents pay with tokenized cards, and Visa has predicted that millions of consumers will use AI agents to complete purchases by this holiday season.
This is why blocking agents from online banking does not, on its own, keep a bank in the conversation. An agent does not need a bank login to spend money. When it pays with a tokenized card, the issuing bank still approves the transaction, but it may never hear from the customer at all. That is the UPI pattern again, where the bank moves the money and someone else keeps the customer.
The rules have not caught up either. A Consumer Bankers Association white paper prepared with the law firm Davis Wright Tremaine flags the problem. It notes that the Electronic Fund Transfer Act’s general limit on consumer liability for unauthorized transactions “may not apply when agents are involved.” Wells Fargo has put its own answer in writing. Most banks have not given one yet.
The contest will be decided on the bad days
Most of the talk about AI in banking assumes the prize is convenience, meaning who answers fastest and who has the smartest assistant. Convenience matters, but it is easy to copy and easy to switch. In my work on customer operations, the moments that shape how people feel about a financial company are rarely the easy ones. They are the card declined at the pharmacy, the charge nobody recognizes, the transfer that has not arrived on the day rent is due.
Agents will create new versions of those moments. Sooner or later an agent will pay a bill twice, or follow instructions planted by a fraudster who has learned to target software instead of people. When that happens, most customers will not call the AI company. They will call their bank.
The same CBA paper is blunt about it: customers “expect banks to make them whole for failed improper payment transactions.” A clause in an access agreement can move the legal liability. It does not move the phone call.
I have argued before that the AI race in banking will be won in the back office, in fraud, risk and operations. Agents make that true for the front end as well. The institution that resolves the bad day quickly and fairly will earn the right to be the one customers, and their agents, come back to.
What banks can do now
Blocking agents buys time, and given the risks, that is reasonable. Shifting liability to customers protects the bank on paper. Neither one is a plan for keeping the customer.
The stronger position is to become the place where handing money decisions to software is made safe. Banks already let customers freeze a card or set a spending alert, and the same idea can extend to agents. A customer could allow an agent to pay certain bills or spend up to a set amount, and switch that access off in one tap. The bank sees every transaction on the account, which makes it the natural home for those controls.
Operations will have to change as well. Dispute and fraud teams will need to tell an agent’s purchase from a customer’s, and an agent’s honest mistake from a fraudster’s manipulation. Mastercard has started testing a score that estimates how likely it is that a transaction was initiated by an AI agent. Tools like that will help, but the harder work is in the service model, with clear rules for agent-related disputes and people trained to fix what the software got wrong.
Banks should also assume that agents will compare them. An agent deciding where to keep a customer’s savings will read the rates and the terms closely, and it will not be swayed by a brand campaign. A product that only makes sense after a careful reading of the fine print will struggle in that comparison.
The banks behind UPI did not lose the front end for lack of technology. They lost it because someone else made paying easier and became part of the customer’s day. Banks may not win the everyday interface this time either. Agents will be good at the daily work of comparing, paying and reminding. What banks can win is the role that needs a balance sheet and a license: the institution that makes it safe to hand money decisions to software and puts things right when they go wrong. Those will be the banks customers tell their agents to use. The ones that only block or disclaim may find, as banks have before, that customers rarely wait for permission.
Manish Agrawal drives strategic initiatives at Chime®, a leading US financial technology company. He previously held senior strategy roles at Paytm, where he led major acquisitions and built new products, and worked in M&A at Goldman Sachs. He is a Global Advisor to Fintech Alliance Nepal and holds an MBA from Kellogg School of Management. The views expressed here are his own.
