United Bank in Georgia has two agents by the name of Sam working in its customer service department. One is a real live agent, Samantha, who is helping to test and train an AI-driven virtual assistant to serve customers. The other is the AI, which is known only internally as Sam.
The objective at United Bank is to train the virtual assistant Sam to be as effective at customer service as Samantha, one of her real trainers. And it’s well on its way.
Based in Zebulon, Ga., United Bank still treats its customers every year to happy birthday calls with best wishes for the year ahead. It had worked hard to maintain its deep connections to the community by keeping 18 branches open across central Georgia and maintaining a well-staffed call center to allow its clients to speak with a banker when they have questions or concerns.
So, when management began considering the adoption of an artificial intelligence-enabled “virtual banker” to handle some of the 45,000 to 55,000 calls the call center receives every month, they understood they would have to be extremely intentional about how the product rolled out.
“We have been banking generations of families over the years,” United Bank senior vice president Lauri B. Irvin tells The Financial Brand. “So, we have a large segment of our customers who prefer more traditional methods of banking and really want to talk to somebody. We have customers who go into our bank offices and sit down with their favorite banker and have that person help balance their checkbook.”
Keeping Up with Customer Needs
Over the years, she says, United Bank customers had been effectively “trained” to call when they had a question about their account and to expect a live person to pick up the phone. Management realized that rolling out a virtual banker would have to be done with extraordinary sensitivity to the expectations of its customer base.
Complexity of the rollout aside, the business case for automating responses to some of the call volume was clear and compelling. The bank was growing and so was the pressure on its call center. Management wanted to expand the call center’s service hours, but was running into the same reality as other employers in a tight labor market: staffing a 25- to 30-person call center was a constant challenge.
Additionally, fully 75% of the calls coming into the call center were routine requests for basic information — balance requests, inquiries about branch hours, questions about interest rates — that could be automated relatively simply.
Automate the Automate-able:
The number of phone calls coming through United Bank's call center that could quickly and easily be automated quickly:75%
Taking even some of those out of the queue would reduce wait times substantially, and allow call center employees to spend more time with customers who needed personal attention.
The trick would be finding a solution that could reduce the burden on the call center while maintaining the bank’s commitment to personalized service.
Encountering Banking-Specialized AI
It was a chance meeting at an industry conference last year that set United Bank on the road to adopting its virtual banker. Executive vice president Jaye W. Eubanks, who oversees IT for the bank, happened to be seated at a table with representatives of Directlink, which has developed what it calls a “conversational banking platform” specifically for community banks and credit unions.
The system is integrated into a bank’s existing call center architecture and is set up to answer calls in a way that gives the customer the opportunity to get an immediate answer to common queries — rather than being placed into a queue until a customer service representative is available.
Directlink’s product uses the same methods of authenticating customers’ identity as live tellers and customers are always given the ability to immediately navigate away from the virtual system and be connected with a bank employee. Certain key signals, such as a customer suggesting that they are calling to report fraudulent activity, results in an automatic offer to redirect the call to a live call center agent.
Ben Nichols, co-founder and director of product for Directlink, tells The Financial Brand the company’s offering has some of the features common to large language models that have become familiar to so many in recent years, like ChatGPT, but with important limits.
“We are evolving to achieve the level of conversational style of ChatGPT, but offering it in a very secure and audited way, so that you end up with finite call paths.”
— Ben Nichols, Directlink
The system has been trained to recognize dozens of common queries from banking customers, allowing it to engage conversationally with callers. On the back end, a connection to the bank’s internal systems gives it access to the information required to answer those questions and conduct a limited number of transactions.
However, where ChatGPT has access to a vast array of uncurated knowledge, and a tendency to “hallucinate” information that isn’t true, the Directlink system is closely constrained to a specific knowledge base that can be tailored to suit each bank’s needs.
“We are evolving to achieve the level of conversational style of ChatGPT, but offering it in a very secure and audited way, so that you end up with finite call paths,” Nichols says.
What that means in practice, he adds: “You always know what it’s going to say.”
Read more:
A Phased Rollout
Last fall, United Bank began planning to introduce its virtual banker to customers. Callers to the bank were already used to hearing a woman’s voice on the existing phone tree, so the virtual banker was given a “female” voice, and right after Labor Day, the system was turned on for a limited trial.
The bank’s normal call center hours ended at 8 p.m., so the virtual banker was originally turned on then, to accept what the bank assumed would be relatively limited call volume after hours and on the weekends.
The first surprise the system delivered was that the volume of customers seeking assistance after hours was larger than expected. Some calls came in late at night. In other cases, customers who had to be at work at 6 a.m. were calling the bank at 4 a.m., often just looking for answers to simple questions.
“Prior to this we didn’t have a way to know who was [calling] after hours,” Irvin says. “So, it validated that there is a need for more extended hours of service in some capacity.”
Usage data also demonstrated that many customers were able to get the information they needed, or conclude basic transactions using the virtual system.
By early October, Irvin says, the new virtual banker had been performing well, and the bank was convinced that they had a “good, stable product” in place. The decision was made to bring the system online during part of normal business hours, beginning at 2 p.m. each afternoon.
A Moment of Concern
In the United Bank call center, employees are highly attuned to a call board that provides real time information about how many representatives are engaged with customers, how many customers are waiting on hold and how long the average wait times are.
Not long after the virtual banker was brought online during regular business hours, Irvin says, there was a moment of panic in the call center. The board, which normally indicates that most call center agents are actively engaged with United Bank clients, instead indicated that 75% of the agents in the room were idle.
The immediate assumption was that something had gone terribly wrong — the phone system was down, or some other problem was preventing customers from calling in. But managers soon came to realize that the “problem” was no problem at all. So many customers were opting to use the virtual banker that demand for live agents had temporarily plummeted.
Over time, Irvin said, the United Bank virtual banker began to capture about 35% of the call volume on any given day, providing itself especially useful on days where volume is predictably high, such as the first of the month, when customers call in to see if their paychecks have cleared.
The successful rollout of the virtual banker at United Bank has not been without its hiccups. The system cannot always understand callers who speak with particularly heavy accents and has trouble with certain situations. For example, it may struggle to make sense of a request from a caller who has a television playing loudly in the background.
Dig deeper:
- Call Center Staff Can’t Help Customers If They Can’t Find Key Data
- Customer Service ChatGPT: Don’t Take AI Too Far in the Call Center
And while the system has been trained to recognize and understand a large number of queries, it can’t anticipate everything. As Irvin says, “There are a zillion different ways that customers ask for their balance.”
Every day, Irvin has the opportunity to review words and phrases that the virtual banker was unable to recognize, and can create a new “training phrase” that helps the system get it right the next time around.
One benefit of the Directlink system is that just as Irvin identifies problems and helps the system learn from them, so do other managers in the different banks that also use the product, generating a constant flow of information about how it can be fine-tuned and improved.
“The more customers and the more partnerships that we form, the better the system becomes, because it’s learning from those experiences,” says Directlink’s Nichols. “It learns and grows with every new client and learns and grows with every new partner.”
Providing the Bank To Build ‘Bench Strength’
In the United Bank call center, the impact of bringing the virtual banker online has been significant. Call center workers — always hyper-sensitive to the duration of wait times and the number of callers on hold — have a little more leeway to spend extra time with customers who have complex problems, without worrying that they are inconveniencing other customers.
It has also helped in other ways.
“One of our longer-term objectives is to be able to develop ‘bench strength’ — to help our entry level agents see their career paths,” Irvin says.
In the past, with some agents handling as many as two hundred calls a day, burnout was a real concern.The need to keep up with call volume made it a challenge to mentor young bankers.
“But if the virtual banker can handle 35% of those calls, then I have time,” Irvin says. “I can take you off the phone, take you into the training room, and teach you the skills and what you need to know to be that next-level agent.”
She added, “Freeing up time so that we can teach, and coach and develop our employees has been a win for everybody. A win for our customers, and a win for our employees too.”