
Episode 247 | November 02, 2026
AI is changing banking. Can UX design keep up?
Citi's Meg Roebling explains where AI in UX research falls short, why humans belong at every gate, and how to design AI experiences people trust.
AI is changing banking. Can UX design keep up?
Research has never been easier to run. It has also never been easier to get wrong.
That's the tension at the center of AI in UX research right now. Tools test and summarize in a fraction of the time they once took. Developers build prototypes. Designers crunch numbers.
Meg Roebling, a director at Citi who leads user experience for consumer credit cards, has more than 20 years in digital design, much of it in financial services. Her first job out of college was at Citi. She left Wall Street in the dot-com boom, hand-coded HTML at a multimedia startup, and eventually came back.
In a conversation with Lija Hogan, principal experience research consultant at Auros, on Insights Unlocked, Meg made a case that should reassure researchers. AI speeds the work up. It doesn't replace the judgment that makes the work worth doing.
Research won the argument but not the budget
Most organizations now accept that research makes better products, Meg said. They've gotten the memo. That doesn't mean research happens.
The reason is the one she ran into decades ago: competing priorities. The person funding a project may care most about modernizing a tech stack or bringing in new business.
So Meg doesn't argue for customer experience for its own sake. She shows how a well-researched, well-designed product serves the funder's goal. In a bank, one frame tends to land.
"Maybe it's mitigating some sort of risk. That's usually a winning argument," she said.
Other battles persist. Stakeholders mistake UAT for usability testing and assume passing UAT is good enough. And research still arrives late, with no time left for discovery.
BENCHMARK
AI Relationship Quality
The UserTesting AI Relationship Quality (ARQ™) benchmark gives you a clear measure of how well your AI experience is working for customers
Small wins still build research buy-in
Meg's method for research buy-in hasn't changed much, and she doesn't think it needs to.
Start small. Answer a simple question quickly. Do people prefer the drop-down or the multi-select box? This headline or that one? Show the result, and stakeholders come back again and again for validation.
The other half is restraint. Once stakeholders get a taste, some want to test everything. Meg's answer is to pull up previous research when it already covers the question.
As she put it, "if you can demonstrate a small win and become that sort of go to person, then you're going to start changing… how people see you and how that function is seen within the organization."
Research storytelling helps. Meg learned at agencies that people respond to the hero's journey: a problem, a process, a clear answer. Add a verbatim clip of a real person struggling, and it clicks for stakeholders.
Where AI in UX research falls short
Now everyone has the tools, and roles are getting fuzzy. Designers build HTML prototypes. Developers design screens.
The results can look great. That's the problem.
When she gets a design prototype from a developer, sometimes it's really good, she said. "But sometimes at first glance it's like very impressive. But then when you start digging into the details, you're like, oh, that doesn't make any sense."
The same gap shows up in AI research synthesis. AI is good at surfacing themes and weaker on nuance. It misses subtext and the participant who contradicts themselves. It doesn't see body language. It may toss the outlier that matters most, or flatten results until everything sounds the same.
Then there's the obvious one. "And then of course there's hallucinations, right. Like they're still making stuff up," Meg said.
Think of AI as a metal detector on a beach. It tells you where to dig. It can't tell you whether you've found a ring or a bottle cap.
None of this makes Meg anti-AI. It makes her specific about where humans still belong: in the rigor of the analysis.
A chatbot is not an information architecture
Asked where AI covers up a design problem instead of fixing it, Meg didn't hesitate. The biggest example is the chatbot used to get around poor information architecture. Can't find something? Ask the chat agent and skip the navigation.
She still thinks screens matter, especially in financial services UX. Applying for a card or a mortgage is a multi-step process heavy with information, and people need visual support to process it.
Her model is a pairing. The chat assistant helps. The screen confirms: the payment went through, the card was approved.
Insight+
Get on-demand access to some of UserTesting's most popular events, including from Crafted Seattle and Crafted London. Register for Insight+ and start watching today.
Designing trustworthy AI experiences starts with transparency
Lija pushed further. Usability asks whether people can complete a task. AI asks them to cede control to a system acting on their behalf. Does that change what a good experience is?
Yes, Meg said, around trust. She also admitted something surprising for a UX leader: she has played around with AI agents but isn't giving them her banking information.
That gap is the opening for banks. "People trust their banks to keep their money safe. So we need to enable experiences that make people feel like we're looking out for them," she said.
For Meg, designing trustworthy AI experiences comes down to a few principles. Put enhanced security on transactions. Clearly mark anything an agent initiated. Have a way to stop an agent that misreads a command and goes on a spending spree. And practice AI transparency: say when an experience is AI-enabled, how customer data is used and which models are giving advice.
Most customers, she added, don't want a lecture on why they got a recommendation. The bargain is simple: you have my data, so personalize it. A light nudge that mentions previous shopping is usually enough.
Agents are becoming users, too. In demos she's seen, they come back with screenshots of the booking experience. "I think we definitely are going to have to start designing for agents now," Meg said.
Humans at every gate
Meg described the model she sees emerging: an end-to-end, AI-driven product development lifecycle. One agent analyzes insights. Another generates requirements. Others handle concepts and user testing.
The safeguard is the gate. Controls at every stage confirm the work hasn't gone off the rails and uses the right design system. Then, she said, "you still ideally have humans in, in the loop to, to check the work and monitor the work before it passes to the next gate."
Lija raised a counterargument from engineering: let AI run the evals, because human review can't keep up. Neither was persuaded. At Citi, Meg said, caution wins: gates, a central agent registry, and agents treated almost like employees held to the same codes of ethics.
Her advice for leaders depends on maturity. With no research practice, start small again. With more, invest in analytics to see where people drop off, then dig into why. Invest in rapid research: a platform such as UserTesting, with a couple of key people overseeing studies, rather than a big team that looks too slow and too costly.
That's the real promise of AI in UX research. Speed where it helps, and trained humans watching every gate.
Meg put the stakes plainly:
"I hope for all of our sake that we still choose to play a role in designing the things that are for us."







