- Key insight: Bank of America's CEO is emphasizing accuracy of AI models, where others focus largely on efficiency and cost savings.
- What's at stake: If a large bank's AI model is allowed to make errors in code, operations or customer service, the result could be catastrophic.
- Forward look: Expect more banks to invest in data infrastructure and model verification.
When most bank CEOs talk about AI, they speak of efficiency and cost savings. But in an interview last week,
When a customer types a question into the bank's AI assistant, Erica, "You can be driving 60 miles an hour up Sixth Avenue, and the infrastructure to get that inquiry to us and back has to be instantaneous, or you're mad at us, so the data has to be perfect, and you ask about one check transaction of 20 that month in one of your accounts, and you have five, and we have to be exactly accurate," Moynihan said. "So it takes work."
The accuracy message is critical, according to Theodora Lau, co-founder of Unconventional Ventures, a consulting firm.
"It stands out because most CEOs — at banks and other companies — are still talking about AI from the perspective of operational efficiency, productivity gains and cost cutting," she said.
In a podcast Lau recently hosted, Máté Jendrolovics, CEO and founder of Intuitech, said, "Having a solution that does something with 80% accuracy means zero value for a bank. It doesn't mean 80% of the value. It's zero."
"When we asked the question, 'What's my balance?' It could give you a picture of a scale, it could give you a yoga class. So we said, 'We've got to take the language that people speak, the language that the bank speaks, and put them together and make sure we're accurate.' And so we went and said, 'Develop a predictive model.'"
Developers within the bank worked with researchers at Stanford to create Erica, which Moynihan referred to as a small language model. The system touches 110 systems and has been designed to answer 700 types of questions or "intents" accurately, he said.
Today, 20 million customers use it 200 million times a quarter, he said.
"Asking a bot questions around finance is vastly different than asking for a book or music recommendation," Lau pointed out. "The tolerance for inaccuracies when it comes to money matters is very low, and rightfully so. Precision matters. Which goes right back to — where you use generative AI and agentic AI matters. How you deploy it matters. Context matters."
Where banks put efficiency first, some use "calls deflected" as a metric, Lau said. "You can provide an inaccurate but quick answer with the bot. It will score well from the metrics perspective, hence it will show up well as a productivity gain. But the customer on the receiving end will not get the accurate answer that they need," she said.
Inside the bank, about 200,000
"So you can see this having the impacts we want, but you have to remember that for anything that takes real judgment, a human has to be in the loop, because the answer has to be perfect if you're in our business and trusted," Moynihan said. "That's why we keep it constrained enough to give the right answers."
Someday
"But right now it works, and that's a different element than a lot of people have," Moynihan said.
