How Truist, product of the decade’s biggest bank merger, built its first AI assistant

One summer day in 2021, Chad Elley joined a Zoom meeting from his home desk in Atlanta, with his basement bar—and 100+ bottles of bourbon—as his background.

Elley was helping to build a brand-new AI banking assistant for Truist, the new bank formed in 2019 by the BB&T-SunTrust merger. At $66 billion, it was the biggest bank merger in a decade—and in parallel with the merger integration, Elley was part of a team of about 50 in business, design, and technology that had been working to shore up the new institution’s customer-facing AI tools.

“We learned quite a bit over the merger and looking at different channels that our clients were wanting to operate in,” Elley, who is SVP head of client enablement at Truist, told us, adding, “As we look at the younger, more millennial age of our clients and how they want to engage, a lot of times they don’t want to necessarily pick up a phone. So we knew that there was a channel that had to be created for that.”

In the past, BB&T and Suntrust had only ever offered simple chatbots, which operated via multiple-choice inquiry options for customers and automated responses tied to each. Truist Assist is the new institution’s first-ever AI-enhanced virtual assistant, which uses natural language processing to answer 100+ potential customer questions ranging from account details to how to buy a home, and also passes off queries to one of Truist’s six contact centers.

Building the bot

In August, after months of beta tests and user feedback, Truist rolled out the assistant to all of its personal banking customers via the bank’s app and website.

The institution’s previous chatbots were used for basic transactions like scheduling bank appointments—relatively cut-and-dry, without any machine learning involved, Sherry Graziano, Truist’s head of digital banking and contact centers, told us.

“If you look at where we are now, it’s much more robust,” Graziano said. She added, “It’s much stronger in nature with natural language processing [and] understanding…[Clients can ask] questions like, ‘I want to learn about buying a home,’ ‘I want to manage my alerts,’ maybe, ‘I want to go and order checks,’ as an example.” She added, “But then you’ve got that layered in with much more comprehensive…abilities to search for and learn additional things.” That could be digital banking questions or financial education, she said, instead of just booking an appointment at a branch.

To pull this off, after months of trials with different chatbot builders Truist decided to build the assistant using Amazon Lex, AWS’s conversational AI platform, which also powers Alexa. But there was still plenty of personalization involved even when relying on an outside tool, according to Bjorn Austraat, SVP head of AI acceleration at Truist.

“Understanding that the vocabulary specific to an enterprise has to be taught—so that, for example, ‘Truist One’ doesn’t get misrecognized, misunderstood, as the number one, because it’s a product for us. Truist One is a checking-account style,” Austraat said “So it’s very, very important when considering vendors for any type of chatbot to not underestimate the amount of effort that will go into customizing. Even a very competent platform…you still need to tweak it and make it understand your vocabulary, and without that, it will not perform well.”

One way Amazon Lex helps companies tailor virtual assistants is by allowing a client to feed historical call transcripts into its platform, and then using language models and deep learning techniques to deliver a product. Truist declined to confirm whether it fed historical call transcripts into Amazon Lex.

“[Amazon Lex’s] automated chatbot designer can typically analyze 10,000 lines of transcripts within a couple of hours to identify intents such as ‘file a new claim’ or ‘check claim status,’” Swami Sivasubramanian, VP of Amazon AI, reportedly said at AWS’s annual event in 2021.

Looking ahead, Austraat said, there are specific, more complex customer inquiries that Truist Assist can’t yet help with—for instance, if a customer wanted to know if overseas political turmoil could affect their investments. But he said he feels confident that starting small, paying attention to what customers ask for, and prioritizing a “warm hand off” to contact centers is key to planning out an AI expansion roadmap.

“It’s best practice in agile AI to build a smallish chatbot that has a really good warm handoff,” Austraat said, adding that if a chatbot doesn’t know how to handle a query, it can pass things off to an agent. “I think customers have no problem with that, especially if then the agent can pick up and say, ‘Oh, hi, Mr. Smith, I see you have a question about options.’ Because immediately…you have a landmark. And you know that your effort wasn't wasted.”

For his part, Elley recalled that the toughest part of building the AI-powered, virtual assistant was also the most human part: Building an internal team dedicated to Truist Assist—especially a team that had never worked together as a unit before and had to build personal relationships remotely.