One project Art Weston has completed in the past year is a major overhaul of Fifth Third Bank's chatbot, which is called Jeanie.
He has applied for patents on the technology his team created to make Jeanie better, and because of that is reluctant to offer in-the-weeds details. But much of the improvement is in the natural language understanding that's used to decipher customers' questions. The chatbot used to interpret questions accurately about 25% of the time; now it's right about 90% of the time.
"One of the things that all companies struggle with is a natural language understanding model that can identify what customers are requesting at the onset of a chat experience, such that companies can then initiate automation, so those requests don't have to go to human agents," said Weston, Fifth Third's director of AI transformation.
Some algorithms get confused when people ask something with too much nuance. Other algorithms classify questions too narrowly and can then misclassify requests.
"The challenge in the beginning was, how do you get that just-right perspective?" Weston said. "I and a few members of the team spent many months trying to figure out what that perfect point was. In the end, they determined that there is no such perfect point. They re-architected the way the natural language understanding algorithm operates such that this was no longer an issue." The natural language understanding algorithm is from Live Person, Fifth Third's chatbot provider.
In the future, advanced AI could leapfrog over this work, Weston acknowledged.
Weston also has applied for patents on technologies that address the risks associated with using generative AI in customer-facing scenarios, such as hallucinations and the possibility that a bad actor could come in and jailbreak the bank's AI and get it to perform actions or activities unintended during its design phase.
Jeanie can also now complete certain tasks. The bot can recognize 150 actions, and about half of those now have some form of automation; the other half will have some level of automation by the end of the year.
Some involve being able to relay information from an API, like a routing number. Others can dip into an application to retrieve, say, a customer's card order status.
"Under Regulation E, we need to get you to our disputes resolution department, which is human controlled at the moment," Weston said. "So we simply provide a phone number and a mechanism to help them get to the right parties."
Weston's team uses software from Lucid to create these automated processes. Customers always have a way to get back out to the main menu and to a human agent if need be.
"We have no doom loops in our systems, so no one gets trapped in our box," Weston said. "If someone wants an agent, we do try to get them to tell us why."
In other uses of AI, Fifth Third uses machine learning algorithms to understand which consumers to try to acquire and retain and which customer relationships could best be deepened. Weston's team is exploring a lot of generative AI use cases internally.
"We're trying to help make generative AI a productivity tool for the common employee," he said. "We're nearing our first deployment."
