Banks whipped up basic chatbots in the early pandemic to deflect customer queries. Years in, they've been learning from customer preferences and turning them into more fully formed
Virtual assistants from
Other banks are revamping the bots they have with more intents — that is, the purposes or goals customers have for querying a digital assistant — and training phrases to boost self-service. Additionally, they are tweaking them with convenient features such as seamless handoffs to live agents (along with estimated wait times) and the ability to pick up a dormant chat where it left off.
Expectations for sophisticated virtual assistants are rising, said Steven Poling, director of customer success at the conversational AI platform provider Clinc.
"ChatGPT has thrown a wrench in everything," he said. "There is a much higher expectation from customers and prospects about what the virtual assistant can and can't do."
It's important that a virtual assistant not overpromise and underdeliver, cautions Emmett Higdon, director of digital banking at Javelin Strategy & Research.
"It's fine to have a narrow chatbot where all it does is take your FAQs and search them for the customer," he said. "But don't set the expectation that you can ask my bot how much you spent last week at Home Depot if that's not the case. One of the worst possible things, even if you've got a great bot to start with, is to [have it say] 'ask me anything.'"
Enabling a seamless transition to live agents is also key for a positive customer experience.
"We hear a lot, 'what live chats are you pre-integrated with?'" said Lindsay Soergel, chief product and customer experience officer at conversational AI provider Kasisto. She also notes interest from banks in providing credit card or other offers in a chat session, as well as Spanish language support, partly to gauge demand for fully bilingual staff in the contact center.
Truist Financial in Charlotte, North Carolina, announced the launch of Truist Assist, a digital assistant that uses natural-language processing and natural-language understanding to interpret customer questions, in September 2022. Previously, the $555 billion-asset bank
"Speed to market was how we looked at bots in 2020," said Chad Elley, head of client enablement at Truist, in a
Fifth Third Bancorp in Cincinnati installed the "1.0" version of its virtual assistant, Jeanie, in 2020 to help handle an influx in phone calls. The bank has made significant changes to Jeanie "2.0" in 2023. Jeanie is built on a conversational AI platform from LivePerson.
"We took time over the past year to look at what we did and whether it was the right approach, or would we take another approach if we were not in that type of a crunch," said Michelle Grimm, senior director of conversational artificial intelligence at the $206 billion-asset Fifth Third. That meant analyzing customer queries and determining whether Jeanie was accurately capturing their intents, and building more dynamic conversation flows.
Jeanie is now trained on 156 customer intents, up from 35 in 2022. The number of phrases used in training has jumped from 3,000 to 30,000 between 2022 and 2023. Fifth Third made these leaps by analyzing thousands of customer utterances, and supplementing them by querying ChatGPT for different ways customers might ask for a specific action, such as requesting a debit card.
The initial version of Jeanie would deliver step-by-step instructions for, say, how to send money via Zelle.
"Quite honestly, oftentimes she was giving you a phone number to call us," said Grimm.
In the Zelle example, Jeanie 2.0 will now link the customer to the right spot to transact in the mobile app. When Fifth Third can't resolve a query so easily, it will connect the customer with an agent directly in the app.
The bank has woven in several conveniences for both customers and agents. The assistant will now estimate a wait time when fetching a live agent. When chats have gone idle for 24 minutes, customers are placed in a "virtual lobby" to avoid clogging an agent's queue, and given the choice of continuing the conversation with a new agent or ending it entirely. Moreover, when an agent receives a conversation that began with Jeanie, they will see a bullet-point summary. In the past, they had to skim the entire conversation.
Customers can now free-type their queries instead of choosing from a preset list, as they had to using Jeanie 1.0. For instance, if a customer writes "Potbelly charged me twice," Jeanie will respond "It looks like you want to chat about duplicate charges on your card. Is that right?" Grimm reports that these classifications are correct 90% of the time, up from 25% in 2022.
KeyCorp in Cleveland followed a similar arc. It launched two bots on public-facing Key.com in 2020, one to drive down the number of frequently asked questions flowing into the contact center and the other to address questions around the Key2PrePaid debit card from its commercial payments business.
"They served a purpose, but it was not a strategic direction with an intelligent virtual assistant," said Jordan Olack, senior vice president of intelligent automation and contact center delivery at the $193 billion-asset Key. "We were interested in more engagement with our client base."
That public-facing bot still exists. But MyKey, the bank's digital assistant, has been in the general market since the fall of 2022. It was custom-built by Key on Google Cloud's conversational AI Dialogflow platform.
KeyBank's first step was making navigation in its mobile app easier after a J.D. Power retail banking study from 2022 revealed this as a weak spot. Customers asking MyKey to pay a bill, to lock a debit card or to make a Zelle payment are directed to the right spot. Going forward, the bank wants MyKey to mature from a navigational assistant to one that can dispense advice, such as, "How do we proactively let you know that you have a payment due in three days for your credit card, and [ask] do you want to pay it now in a conversational way?" said Olack. As with Jeanie, customers can transition to a live agent within MyKey.
Call volumes have decreased by 10% since MyKey launched. The rate of customers returning to use MyKey has also steadily risen over time, from 55% in February to 59% in May to 69% today.
U.S. Bancorp and Truist picked literal names for their virtual assistants to avoid confusing customers, while Fifth Third and Regions created characters called "Jeanie" and "Reggie" to put clients at ease. Other companies are trying to split the difference.
Regions Financial in Birmingham, Alabama, has a different story: It installed its conversational messaging assistant, Reggie, in 2019, before the pandemic started. The initial version of Reggie addressed the top handful of questions that came into the contact center. Regions has since added more features and capabilities, including the ability to pass customers with a payment dispute question to a fraud specialist and provide information about its new
The $155 billion-asset bank has made enhancements similar to the others mentioned. Reggie will estimate a wait time when transferring customers to a live agent — something Regions found to be a popular feature when it tested the concept in a focus group — and swap the icon from the perky Reggie avatar to the agent's name when a human takes over the chat. Regions also introduced asynchronous messaging, or the ability to return to the same chat thread within the same calendar day that they exited the window; chat history is available in the mobile app for longer.
"These are little enhancements we made over the last year to make it feel more engaging than a standard bot," said Chris Brasher, modernization and transformation executive at Regions, referring to asynchronous messaging and the estimated wait times. In a crawl-walk-run scenario, "I'd say we were crawling in 2019 and we are now walking pretty good," he said. "We're not quite at a run yet, but we are getting there."
