Joe Fuqua
Intelligent Automation Architecture Strategy & Governance
Algorithm & Blues · Weekly
Charlotte, NC · Est. 1988
Algorithm & Blues

The Customer Has an Agent

Customers may ask AI assistants to compare bank accounts long before trusting them to move money. Research suggests those recommendations can change with the order in which an agent encounters the same information.

Issue #78
Published November 8, 2026
Series Weekly publication
Source Original

[On Tuesday, I’ll be speaking at the Banking Transformation Summit in Charlotte on “Time Capsule: US Banking in November 2027.” This week’s issue explores one development I’m considering ahead of that discussion.]

A human hand and a robotic hand hold cards showing the same three banks in different orders.
Conceptual editorial illustration.

As banks work out how to use AI to serve their customers, I’ve been thinking about the other side of that relationship, with customers using their own AI agents to evaluate what banks offer. Asking an assistant to compare accounts seems like a fairly natural place to start. It might save some reading and help narrow the choices, without requiring anyone to trust it with their money. A customer might rule out a bank based on an assistant’s recommendation without ever talking to anyone there.

A recent study from Wharton explored how those recommendations take shape. Researchers asked AI agents to choose a fitness watch from a selection of eight products. Across roughly 26,000 tests, they varied the information the agents received before making that choice, including screenshots of reviews and recommendations from outside sources.

When shown only the products, the models tended to choose consistent favorites. Introducing an outside recommendation sometimes changed their selections considerably. Some models also chose differently when the same sources were presented in a different order. In another experiment, choices changed depending on whether the sources arrived together or separately. So the agents were responding to how the information was presented as well as what it contained.

These were controlled experiments using screenshots, so they don’t establish how an assistant would compare bank accounts. They do, however, suggest a complication for businesses expecting agents to evaluate their products consistently. Publishing accurate information may not be enough to ensure that an assistant gives it the same consideration each time.

Consider a customer looking for a savings account. The bank’s website explains the rate and eligibility conditions, an independent comparison ranks the account favorably, and a customer review describes difficulty withdrawing money. Each source could contribute something useful. How much weight to give the review would depend partly on whether it describes an isolated experience or a problem likely to affect that customer.

An assistant would have to make that judgment while deciding which accounts to recommend. If the sensitivity observed in the watch experiments carries into this scenario, encountering the review first could produce a different recommendation from reading the product terms first. The customer’s needs and the available information could remain the same while the shortlist changes.

People can be influenced by the order in which they learn things, too. When an assistant presents a comparison, though, the customer may have no way to tell how it weighed the information. They might reasonably assume it would recommend the same account if it had read the same sources in a different order.

That also makes it harder for a bank to understand why a customer chose another account. Its offering may simply have been a poor fit for what the customer needed. But if the assistant might have recommended it after reading the same information in a different order, the bank could be trying to fix a problem that had little to do with the account itself.

I’m unsure how quickly people will trust agents to manage their finances, but asking one to compare accounts seems a much smaller step. As banks develop their own AI services, they may also need to understand how their products are being evaluated by customers’ assistants. Seeing the comparison a customer received would help, though it might still leave the bank wondering whether a different rate or better service would have changed the recommendation.

Algorithm & Blues publishes on Sundays. If this was forwarded to you, you can subscribe at joefuqua.blog.

Reference

Kumar and colleagues, Agentic Shopping is Complicated and Contingent, Wharton Generative AI Labs, August 2026.