U.S. Bank Cut Research Time in Half With Synthetic Personas

By Priya Raman

U.S. Bank cut the time needed to reach key research and positioning answers by roughly half using synthetic personas, CMO Michael Lacorazza told Chief Marketer. The bank began experimenting with AI-generated customer profiles in 2024 while developing a new brand campaign, and it has since expanded the program to dozens of personas supporting product campaigns and broader business questions.

Lacorazza discussed the program at POSSIBLE earlier this year, then detailed in a follow-up conversation how the personas operate and where they sit alongside traditional research.

What do the personas actually do?

The bank initially built synthetic personas around its core target audiences to develop the brand campaign. The team used them to accelerate communication strategy, assess competitors and evaluate advertising concepts.

One objective was understanding how the brand should appear in the marketplace, including which messages and product proof points might resonate. Another was identifying competitive white space.

The personas draw on third-party data and social listening alongside the bank's own research: audience profiles, customer segments and answers to questions U.S. Bank has previously asked real people.

"We have third-party data sources like YouGov and social listening, data sources that are indicative of how people feel about things, attitudes," Lacorazza said. "Then the entire library of insights that we have from the proprietary first-party research we've done against audience profiles and segments, and whatever business questions we've been asking of actual people, is part of the dataset too."

The program now covers product-level campaigns involving deposits, small businesses and merchant acquiring, and it helps refine value propositions for new products.

What does this change for regulated marketers?

For a bank, the line between AI informing marketing and AI making decisions about customers is significant. U.S. Bank deliberately started with publicly available third-party information and research findings, not individual customer transaction data. The personas inform employee decisions; they do not drive credit offers or money movement.

"We're most careful about anything that involves first-party data about our own clients being ingested into a model, particularly a model that's doing decisioning on behalf of the customer," Lacorazza said. A system making decisions about credit offers or money movement would face a different level of evaluation.

Traceability serves as the safeguard. The model can show which data sources inform an answer, giving the team a way to assess reliability.

"If the data source it's using to give you the answer is really thin, that's probably not an answer you put a lot of trust in," he said.

Faster research, not a replacement

The time savings were largest in competitive analysis and synthesizing marketplace content. U.S. Bank still runs customer panels and custom customer experience research, though the personas have reduced the need for some traditional studies.

"If it's important enough, you do [traditional research]. If it's not, maybe you trust what you have here," Lacorazza said.

He cautioned against assuming human-origin research is automatically more reliable. Respondents may withhold honest opinions on personal topics, and poorly designed questions can miss the intended insight.

"Just because something has a human origin doesn't make it more accurate or superior," he said. "In both cases you have to manage it and examine it."

The bank reported improvements in creative performance, audience engagement, brand recall, consideration and awareness following the campaign. Lacorazza credited the creative itself: "At the end of the day, that's because the ideas were great and not because of the large language models we used to evaluate them."

Five recommendations from Lacorazza

• Start with the data. "The output of the model is only going to be as good as the quality of the data sources you have," he said. Inventory existing research and identify gaps third-party or public sources could fill.
• Set a narrow objective. "Have really clear goals for what you want your initial build to achieve. And do it in a way where you're not trying to solve for world peace in rev one."
• Build in transparency. Ensure the system shows which sources inform its answers so employees can judge reliability.
• Test what you know. Ask the model questions with established answers and compare responses to find where refinement is needed.
• Stay flexible on vendors. U.S. Bank uses multiple major AI models. "You're better served if you don't tie yourself to a particular vendor, because these things are always going to evolve."

Beyond efficiency

At U.S. Bank, AI now operates across four major areas: research and insights, process automation, personalization and content creation. The content work includes optimizing email headlines, body copy and image selection, building web pages faster, and broadcast production. In a commercial featuring Christian McCaffrey, the bank used AI to fill a stadium with approximately 50,000 people instead of hiring extras.

Lacorazza rejects the vision of fully automated marketing departments or dramatically smaller headcounts. "Personally, I think that's an unrealistic vision and not where we want to go," he said of expectations that AI will slash operating costs.

Instead, he frames efficiency as a resource shift: "Where we can find efficiencies, it enables us to free up resources for higher and better uses that can have more impact."

For practitioners, that is the operative lesson: the personas buy speed on synthesis, and the reclaimed capacity goes to judgment work the models cannot do.