Banks Put New Limits on Third-Party AI Agents

Several US banks have updated their terms or warned customers about third-party AI agents that use account credentials and act on their behalf. The policies put responsibility for an agent’s mistakes in sharper focus, even as banks develop AI tools of their own.

WTF Index NEUTRAL
◄ Terminator 2 Idiocracy 2 ►

Banks are limiting third-party agents over account access and disputed mistakes, a modest concern about autonomous actions and human reliance without a clear dominant lean.

Banks Put New Limits on Third-Party AI Agents

AI agents can shop and carry out online tasks by clicking through websites for a user. But when those actions touch a bank account, financial institutions are making clear that customers may bear the risk. Several US banks have recently revised their terms or issued warnings about outside agents.

Banks are drawing a line around account access

Wells Fargo is set to enact an updated agreement with millions of customers, specifically addressing AI agents. Its new terms warn: “Your use of any AI Agent is AT YOUR OWN RISK.”

Other institutions have taken similar steps. JPMorgan Chase changed its terms in late October to ban third-party agents that log in using a customer’s credentials. On the same day, Bank of America warned customers not to share login information with agents. Citibank and PNC have also added warnings to their terms.

The concern is not limited to banks. Amazon blocked Meta’s Muse from making purchases on its platform, and United announced that personal bots would be blocked from purchasing flights on a customer’s behalf. The pushback reflects a broader dispute over whether outside software should be allowed to take action through people’s accounts.

Automation makes responsibility harder to pin down

An agent may be asked to achieve a goal, then make a series of decisions and clicks to carry it out. Kacy Kim, an associate professor of marketing at Bryant University who studies consumer behavior and emerging tech, summarizes the gap: “Consumers authorize goals, not every click.”

If an automated purchase goes wrong, responsibility can be disputed among the customer, the retailer, the agent’s developer and the bank. Luca Cian, a professor at the University of Virginia who studies consumer psychology and AI, offers a hypothetical example: an agent accidentally orders 10 roasted chickens.

The retailer could say it fulfilled an order the customer placed. The AI company could point to its software and disclaimers. Meanwhile, the bank may argue that using the outside agent violated its terms or exposed the account to risks it does not accept. The example shows how a simple mistake can leave the customer sorting out who, if anyone, will reverse the charge.

Trust depends on what the agent is buying

People may be willing to let an agent handle routine purchases while keeping closer control over choices that are expensive or require judgment. Cian says, “People are very comfortable with AI buying toilet paper.” He describes consumers as more open to AI shopping advice than to handing over control of larger, more dynamic purchases.

That distinction helps explain why a broad promise of convenience may not be enough to make agent-led shopping feel safe. A customer might welcome help comparing products or tracking price changes, but still want to review the final order personally. The more consequential the purchase, the more important that review can feel.

Some card providers are exploring a limited backstop. American Express announced that certain agent mistakes, such as ordering too much of an item, may be eligible for reimbursement in the future if the customer cannot recover the money from the retailer. That possibility does not remove the uncertainty around who is responsible for an automated transaction.

Banks are cautious about outside tools and building their own

The warnings about third-party agents do not mean banks are avoiding AI altogether. Wells Fargo has a generative AI tool that answers questions and has been used more than a billion times. Bank of America revamped its EricaAssist agent this year to help employees, while Citigroup launched an internal development platform centered on agents.

Sucharita Kodali, a Forrester analyst focused on ecommerce and consumer behavior, says most people do not know what agents are, let alone use them. She views banks as risk-averse and sees a reason to set policies early, even if agent-related problems remain uncommon.

Kodali also questions whether shopping agents solve a real problem for many consumers. Buying from Amazon, Walmart or elsewhere is already straightforward, she argues, so an agent may not be necessary for many transactions. For now, customers considering one should pay attention to what account access they grant and what their bank’s terms say: the convenience may come with responsibility for what the bot does.