January 27, 2021

GM Financial’s AI chatbot proves saving grace

When the coronavirus pandemic swept the U.S. last March, shuttering auto factories and forcing most of the country into their homes, GM Financial’s chief experience officer Bob Beatty had his hands full.

More than 25 years in customer service hadn’t prepared him to convert 700 active customer service employees to home offices, training the majority on softphone technologies — web-based calling services — while addressing some of the highest call and message volumes the company had ever experienced.

“There was a problem around each corner,” Beatty said.

On the second night working from home, an IBM representative called Beatty to discuss enhancing the lender’s ongoing project: an artificial intelligence chatbot named Nanci.

Nanci — whose name is derived from “financial” in the lender’s title — was developed in partnership with IBM Services through its Watson Assistant program and has been functional on the company’s site for almost a year. It could manage 10 to 15 percent of the 230,000 online customer messages the lender received per month.

IBM offered to augment Nanci’s programming so it could handle more customer messages.

“He said, ‘We have a team for you. They can be working through the day and through the night to get Nanci smarter,’ ” Beatty said.

For days, the four-person GM Financial team worked long hours with IBM to increase Nanci’s abilities. The process involved programming intents to ensure the virtual agent recognizes a customer concern and has the training to resolve it.

Nanci proved to be a saving grace.

By mid-April, Nanci tackled half of the customer messages sent to GM Financial, resolving 90 percent of those queries without intervention from a human representative. GM Financial employees handled the remaining messages, and answered the average 350,000 phone calls per month the lender received. The next phase of the program is to give Nanci a voice so she can answer phone calls as well.

With each new intent added to the chatbot, “That was another 100 customers that didn’t have to go to a team member,” Beatty said. “I don’t know what we would have done without Nanci.”

The desire for more dynamic messaging capabilities stemmed from shortcomings with GM Financial’s previous messaging system. In order to communicate with the lender, customers needed to keep the application open and use it during designated business hours. If too much time lapsed before a customer typed a response into the messaging app, the chat would close, Beatty said.

“The joy of the asynchronous [messaging] is, send us a message and go about your day,” he said. “Then check your phone, check your device. With the AI the answer will be there nearly instantaneously.”

Nanci can answer customer questions at any time, and can even offer advice. Nanci informed customers requesting payment deferrals due to sudden job loss related to the pandemic about how that may help — or harm — their account in the long term, Beatty said.

Interest continues to accrue on auto loan accounts even if payments aren’t due, Beatty said. For that reason, customers with large balances and higher interest rates could find themselves in a worse situation following a two-month deferral.

“We used our bot to really inform customers when they typed in, ‘I need a deferment or payment extension on a retail account’ the bot would explain, ‘This is how deferment works,’ ” he said. “I think that helped a lot of customers make more sound decisions.”

GM Financial launched the program in late 2019— which likely gave the lender an edge when the pandemic drastically altered business operations. According to Glenn Finch, global managing partner of cognitive business decision support at IBM Services, it takes time and training for companies to achieve successful resolution rates — a range of 70 to 90 percent.

“Virtual agent technology is always going to perform better when you use it before you actually need it,” Finch said in an email.

Virtual agents use a combination of artificial intelligence technologies, including machine learning, natural language processing and generation, sentiment analysis and language translation, Finch said.

Not all programs that use this technology take the form of chatbots, and not all chatbots qualify as virtual agents. Machine learning and natural language processing allows a service to meet the definition, and integrated voice response systems are considered virtual agents only when they employ artificial intelligence for conversation. If the user is limited to a set of specific keywords, Finch said, it doesn’t count.

Nanci’s programming grows more sophisticated every day, though some situations require a human touch.

Beatty cited a recent example of a customer who, over the course of discussing the automotive account, revealed mental health concerns and mentioned suicide. A GM Financial employee took control of the situation, and eventually a transcript of the call reached Beatty.

“Whoever could anticipate that a customer would share that with a chatbot?” Beatty said. “But it also showed us, we have a responsibility to do something here.”

Working with IBM, GM Financial developed a process to ensure Nanci could identify concerning comments from customers, signify that their concerns were being addressed and point them in the right direction to receive help. The incident was an important lesson for Beatty.

“You can’t just deploy this AI and be happy with the results of the customer satisfaction,” he said. “It has to be monitored and it has to be trained and taken care of so that you’re leveraging it in a very responsible way. This thing has to be ready for [interactions] you can never even imagine it needs to be ready for.”