Testing the limits of AI-powered customer service
TD Bank is going where just a handful of banks have been willing to tread — allowing a piece of software to be the first line of communication with customers.
In the hopes of adding a “conversational element” to mobile banking, TD is embedding its app with an artificial-intelligence-based virtual assistant that can answer customers’ spoken or typed questions.
It is joining USAA, Capital One Financial, Bank of America, Wells Fargo and a small number of other financial institutions that are piloting AI-powered chatbots.
The concept is promising but still risky: the software could misfire and say something inappropriate or off-putting to a client, or fail to answer the customer’s question at all.
Yet the reward could be an ability to respond more quickly — and in some cases more accurately — to customers who increasingly expect immediate answers, while saving the bank money along the way. A side benefit that USAA recently discovered with its Amazon Skill chatbot is that people seem to be more comfortable asking embarrassing financial questions of a bot.
Some banks have formed partnerships with tech firms. TD is using Kasisto’s Kai, USAA is working with Clinc, Wells Fargo has teamed with Personetics, and Capital One and Bank of America are relying on in-house developers. Kai has also been used for a year by digibank, a mobile-only bank launched in India by DBS Bank in Singapore.
Kasisto is a spinoff of the research lab SRI International, which also developed Apple's Siri. The first generation of Kai was a collaboration with BBVA. Thousands of calls at the bank were transcribed and fed into the artificial intelligence engine. Full-time writers at Kasisto produced additional dialogues for Kai and general banking knowledge, such as information about CDs, IRAs, and credit scores, was also fed into the system.
Through Kai, TD app users will be able to check account information, review transaction histories and monitor spending levels. They’ll be able to get instant answers about spending-related questions, including how much they spent on recent weekend getaways, what their largest transactions were last week, or what they spent on groceries or trips to the coffee shop last month.
Human vs. bot
The chief questions chatbot pioneers grapple with are, how do you make a chatbot seem like a human, and should you even try to? Umpqua Bank, for instance, recently made the choice to forego a chatbot and instead equip its branch staff to answer customers’ chats.
Zor Gorelov, CEO of Kasisto, often gets asked how virtual assistants compare to humans, and which ones are preferable in which situations.
“It’s a complicated question to answer,” he said. Banks like TD that have large call centers will make different decisions than companies like digibank in India, which has no call centers.
But Gorelov says there are two ways in which virtual assistants can do better than humans.
“One is the sheer breadth of knowledge,” he said. “AI systems are good enough today where they can answer thousands of questions. Humans can’t keep that much in their heads without going to manuals and looking things up. In an age where many consumers, especially millennials, demand instantaneous response, that’s where AI can shine.”
The other is the chance to answer questions a consumer would never think of asking a banker, Gorelov said.
Some customers have asked Kai, “I have too much debt, what do I do?” Others have asked about a gambling problem. Kai has been taught to answer such “social responsibility” questions.
TD will promote its chatbot under an overall “Bank Human” theme. According to the bank, this is not a contradiction.
“ ’Bank Human’ is about being customer-centric,” said Rizwan Khalfan, executive vice president and chief digital and payments officer at the bank. “Rather than trying to figure out what customers want, we can give them a plethora of experiences and use the data behind that to determine” the best way to interact with them. For some customers, such as the highly engaged mobile customer (the bank has 6.4 million of these), the ability to have a two-way conversation in the moment is critical, he said.
Kai draws a strong distinction between bot answers and human answers, so people understand when they’re talking to a person and when they’re talking to a machine. Other virtual assistants can do this, too.
“It’s very tempting to make AI look like human,” Gorelov said. “But AI is not human today, and people should know the difference.”
Independent AI experts have come to the same conclusion. In an opinion piece for The New York Times, Oren Etzioni, a professor of computer science at the University of Washington and CEO of the Allen Institute for Artificial Intelligence, proposed three rules for AI systems to ensure they do no harm. One of them is that an AI system must clearly disclose that it is not human.
“As we have seen in the case of bots — computer programs that can engage in increasingly sophisticated dialogue with real people — society needs assurances that AI systems are clearly labeled as such,” he wrote. He cited a case in which a bot known as Jill Watson, which served as a teaching assistant for an online course at Georgia Tech, fooled students into thinking it was human.
“My rule would ensure that people know when a bot is impersonating someone,” Etzioni said.
Kai can be configured to shift a conversation to a live chat or a call with a human being based on several conditions. One is, if Kai didn’t understand the user for some reason. Another is if the AI engine senses anger or frustration — for example, the use of curse words. A third is fraud.
“If you have fraud on your account and you’re agitated, you want to speak to a person,” Gorelov said.
A fourth trigger might be a potential upsell.
“If somebody is contacting the virtual assistant and says I’d like to buy a new, million-dollar home, maybe that’s the type of interaction that is handled by humans,” Gorelov said.
TD has configured Kai to trigger a transfer to a human before an interaction reaches the point of frustration.
“One of the things we liked about Kai is that it lets us put the rules in place to say this is the kind of transition that will occur,” Khalfan said. “This way we can deliver the experience that truly the customers want.”
Kai will tell customers they are being transferred to a human agent and provide background on the conversations.
To communicate with TD’s Canadian customers, Kai had to be trained to speak Canadian French.
It had to learn when to use the formal version of the French word for “you” (vous) and when to use the informal (tu).
“We had to make some interesting decisions in persona design and create a system that is designed for French Canadians,” Gorelov said.
And because TD is a cross-border bank, Kai had to be taught to handle queries from Canadian customers like, “I want to go shopping in New York.” This prompts Kai to set up banking services on both sides of the border.
Kai is the third fintech-partnership technology TD is integrating into its app, following Moven’s MySpend personal financial management software and FlyBit’s digital-concierge software, which sends real-time notifications about special offers and nearby events.
All the data gathered from these other components as well as basic banking transactions will help inform Kai’s interactions with customers, Khalfan suggested.
“Each fintech partnership is a building block,” Khalfan said. “Moven was driving engagements. Flybit was figuring out context and delivering contextualized services at the right time. We now have all this data and analytics that we don’t recreate in Kai, but drive data insights into Kai to have richer, more meaningful humanlike conversations with our customers.”
It’s part of a continuum of contacts with customers, Khalfan said.
“We have a lot of digital properties — we have a bot on Facebook Messenger and we also have click to chat and click to talk,” he said. “It’s really bringing the bank to the customer.”
TD simply hopes to drive engagement, Khalfan said.
“Customers want to engage with us, they want us to be a big part of their lives,” he said.
With Kai, the customer can have a two-way conversation with the bank at all times of the day, he said.
“You can render experiences based on the needs of the customer, how well you know them, the data you have and the interactions you’ve had in the past,” he said.
If a customer prefers texting over voice chat, Kai will make a note of that. Or if there is a type of question they ask most frequently, Kai will adapt.
Over time, Kai will help the bank maintain the context of conversations across channels.
“If you’re interacting in a connected home and then going to a connected car, you keep the context of that interaction as you move from one interaction to another,” Khalfan said. “That’s the world we are going into.”
Editor at Large Penny Crosman welcomes feedback at email@example.com.