AI agents that shop, pay and trade for their owners are already on the market in the US, but nobody has settled who covers the loss when one of them makes a wrong or fraudulent payment. That is the conclusion of an explainer published by banking software firm Backbase, based on an interview with Kelvin Chen, Head of Policy at the Consumer Bankers Association (CBA), and echoed by experts quoted by American Banker and The Financial Brand.
A rule built for humans. According to Backbase, Regulation E, which implements the 1978 Electronic Fund Transfer Act, caps consumer liability at $50 if fraud is reported within two days. But it has a hard exception: transactions the customer authorized. Chen's example: if you hand your cards to someone to buy a sandwich and they buy a Tesla, you cannot ask the bank to make you whole. Whether an AI agent counts as that kind of authorized delegate is still an open question.
The bank is left standing. Chen argues that consumer AI tools are not built for customer service or remediation, and many may carry arbitration clauses that block redress. That leaves the bank as the practical point of accountability, even for agents it never built. Backbase notes OpenAI launched ChatGPT Finances in May 2026, linking Pro subscribers' accounts through Plaid across more than 12,000 institutions.
Whether you want to be involved in agentic commerce or not, banks will be pulled in because at the end of the day, for almost all of the payment stacks, there will be a bank at the bottom.
Kelvin Chen, Consumer Bankers Association, quoted by Backbase
Not everyone agrees. American Banker reports that Robinhood, which launched agentic trading and an agentic credit card in May, takes the opposite approach: its terms say the user assumes "all risk for orders placed by your AI agent". Its accounts isolate funds, allow hard spending limits and include a one-tap kill switch. Backbase also warns that stablecoin rails offer no chargeback rights like card networks do.
Why it matters
Fraud at machine speed. Chris Ward of Truist told an American Banker panel he built working agents in about ten minutes and doubted his bank would flag them as anything but him. The Financial Brand cites Feedzai data showing more than half of fraud now involves AI. Meanwhile, according to Backbase, the Fed, OCC and FDIC left generative and agentic AI out of their April 2026 model risk guidance, SR 26-2. Ward expects rules to arrive through "a new set of accidents".
- Check the terms of any AI agent you connect to your bank: some put all the risk on you.
- Use spending limits and a way to switch the agent off instantly.
- Avoid letting agents pay on rails without chargeback rights, such as stablecoins.
- Review agent activity often and report anything odd to your bank quickly.
Know your agent. Cognizant's Meena Athinathan told American Banker that banks must confirm four things: the right human, the right agent, that the agent has permission, and that the customer meant it to act. The Financial Brand adds that agents can drift or be hijacked, so checks must be continuous.
Sources: Backbase, The Financial Brand, American Banker
Leave a Reply