Citi has begun deploying a virtual agent in its call centers serving commercial card holders. The move is part of the bank's overall campaign to digitize in every area.
“Our intent is to try to handle the more routine calls that cardholders and our corporate clients are telling us they want serviced as quickly as possible so that we could let the live agents handle the more complex questions,” said Gonca Latif-Schmitt, the bank’s global head of commercial cards.
During the pilot phase, that amounted to about 40% of the calls that came in. The bots' load will probably grow over time, Latif-Schmitt said. The centers get roughly 10 million calls each year.
Though banks have been cautious about implementing virtual assistants or chatbots, adoption is expected to accelerate. Juniper Research recently estimated that in the next five years, use of chatbots in banking will grow 3,150% and that the technology will save banks $7.3 billion in 2023.
For this project, Citi partnered with the virtual agent technology company Interactions, which works with large companies like MetLife and Hyatt Hotels.
“It was really important to us as a business to find a partner that has done this for a long time,” Latif-Schmitt said.
She also wanted a system that would seem natural.
“We spent a lot of time trying to find the right kind of voice and even accent,” she said. “Because we're a global organization, we need to make sure it's a neutral, global accent. The last thing we wanted was something robotic. That is superfrustrating to a cardholder.”
What the technology does
Like all virtual agents or “conversational AI” programs — Bank of America’s Erica, Google’s Duplex, Amazon’s Alexa, Apple’s Siri, Kasisto’s Kai, Clinc, Finn AI and Abe AI — Interactions uses natural language understanding to convert a customer’s speech into computer readable text and machine learning to get better at it over time.
Jim Freeze, chief marketing officer at the Boston company, refers to most virtual assistants as “gussied up IVR," or interactive voice response units. He singles out Google’s demo of Duplex reserving a table at a restaurant a couple of years ago as software that has been trained on a narrow set of questions.
One unique aspect of Interactions’ agent software is that it is guided by a network of so-called professional listeners employed by business process outsourcing firms, onshore and offshore. Whenever the AI is not sure it is understanding what someone is saying, it sends a snippet of audio and the question that preceded it for context to a person who sits at the ready, waiting to offer an instant translation.
“Once they determine what the intent is, that intent then is tagged and fed in near real time to our machine learning loop in the artificial intelligence model,” Freeze said. “So it gets better and better over time.”
The use of human listeners naturally raises privacy questions: What if a customer shares personally identifiable information during a call?
Freeze said his company only routes short snippets of voice interactions to human intent analysts in as little as a second.
“There's no single intent analyst who listens to an entire conversation,” he said. If there are three different times during a call that a customer is hard to understand, three different snippets will be captured and most likely they will be sent to three different people, he said.
“There isn't anybody who's listening to the entire call recording,” he said.
The software is especially good at interpreting speech when there is a lot of background noise or a person has a heavy accent.
“There are hundreds of different things that can happen that make it difficult for technology to understand human intent,” Freeze said.
Freeze said that Interactions’ technology encourages natural dialogue, for instance, by starting with an open-ended question like, how may I help you?
The Interactions agent handled 300 million calls for one customer last year, he said.
The virtual agent can handle spikes in volume better than humans can, Freeze pointed out.
“One of the advantages of the technology is it doesn't force people to wait,” he said.
Other virtual agents, such as Kasisto and Clinc, feed information about a bank’s products and services into the bot, so that it understands those terms and can answer basic questions about them.
Interactions has its technology learn as it goes, with the help of its human analysts.
In action at Citi
One way Latif-Schmitt and the operations leadership team tested the software before investing in it was by sitting in a room where the virtual agent was taking calls for other Interactions clients and listening in.
They wanted to be sure the system could handle calls from customers who are traveling, where there might be a bad connection, lots of background noise and a multinational caller might have a heavy accent.
“We listened in on live calls from very loud and noisy areas,” Latif-Schmitt said. “We wanted to hear, if there is an incredibly loud noise, let's say a roller coaster in the background, how does the agent handle that?”
Latif-Schmitt also wanted to make sure the software could handle unhappy customers.
“I was really curious about, what if we have a cardholder who's really upset?” she said. “We don't want to have someone be frustrated. How does that get handled?”
She found the system immediately turned such calls over to a live agent. Customers can also ask to speak to a live agent any time.
The bank has been launching the software in North America, and it is now available to 80% of clients here. To start, the bank is having the virtual agent handle six inquiry types that lend themselves to a virtual agent, including updating contact information and reporting lost or stolen cards.
The system has been getting smarter over time, Latif-Schmitt said.
“It starts to build a profile of what a cardholder would likely be calling about, so that we can anticipate that as opposed to an IVR,” she said.
