Chatbots to humans: Move aside, I got this

While several U.S. banks are developing or piloting virtual assistants based on artificial intelligence, Singapore’s largest bank has been using an AI-powered chatbot to run its digital-only bank in India for about a year.

The results so far are impressive: Eighty-two percent of all customer support for digibank by DBS is handled through the bot, helping the company run the unit at about a fifth of the cost of running a traditional bank.

A look at how this bot is working in other parts of the world gives a sense of how the technology is likely to play out in the U.S., as banks like Bank of America, Capital One and Wells Fargo experiment in the space.

Indeed, according to Sandeep Lal, group head of digital bank at DBS, the company’s goal in pioneering the chatbot is to shape the future of banking.

The bank is using Kasisto’s virtual assistant, which started life in 2010 as a project of SRI (the research institute where Siri was developed) and BBVA called Lola, where it was trained to answer questions the way a human customer service representative would. In 2014, SRI spun out a new company, Kasisto, to offer the technology to companies. The technology was renamed “Kai.” Executives at Kasisto refer to Kai as “conversational AI” and digibank uses it to communicate with customers via voice and text.

“Our customers often do not realize that they are talking to a bot and not a person as the interactions are so natural,” Lal said.

Kai can do more for a customer than a live agent in a customer care center can, according to Lal.

“If you ask an agent, ‘How much did I spend on groceries last month?’ they might take some time to pull up that data,” he said. Kai can answer instantly. The virtual assistant could also handle complicated service questions faster and more accurately than a human, he said.

Kai can give customers their account balance, handle funds transfers and pay people. It can tell users how much they owe on their credit or debit card or when the next payment is due. It has a transaction locator, so it can answer questions like “Show me my last 10 transactions” or “How much did I spend on food in the first week of January?”

DBS headquarters
Illuminated signage is displayed on the DBS Group Holdings Ltd. bank building at night in the central business district of Singapore, on Friday, Feb. 10, 2017. DBS is scheduled to release earnings results on Feb. 14. Photographer: Nicky Loh/Bloomberg

Digibank launched in India in April 2016 and now has a million customers. It offers a savings account with a 7% interest rate and a debit card.

All customer onboarding for digibank happens on the mobile app, so a lot of the questions Kai gets are around the onboarding process, such as “What is DBS?”; “Who is eligible to open an account?”; “What are the documents I need?”; and “How do I verify my identity?”

“We wanted to have an interface that would be as natural as texting a friend, where customers can get instant and accurate replies to [more than a thousand] common questions, as well as hold natural language conversations on a range of subjects,” Lal said.

Kai has been trained to answer 1,178 questions specific to digibank products, services and customers. It can understand the same question asked in different ways. “We expect accuracy and performance to improve over time with sophisticated machine learning,” Lal said. “This is due to the fact that our virtual assistant has thousands of banking conversations built into the platform.”

Less than 18% of DBS digibank customers’ interactions require live chat sessions; the rest are “contained” within the virtual assistant. The session is transferred to a human operator when the virtual assistant is unable to answer or understand certain questions, there is a problem that requires urgent attention or a particular sales opportunity is identified.

“We think the bots are good but not perfect, so having this handoff capability allows us to learn from what people are asking us the bot is not able to handle,” said Dror Oren, co-founder of Kasisto.

Kai is also trained to look for signs of frustration, such as the same question being asked repeatedly or profanity. As soon as it detects such a signal, it will send that customer, along with the history of the interaction, to a live representative. A wealthier customer who is shouting profanities or asking the same question twice might be handed over to a live agent quicker than an average customer exhibiting the same behavior.

Sometimes the bot will say something like, ‘Hey, looks like we started off on the wrong foot, is there anything I can help with or do you want to be handed off to a live agent?' ” Oren said.

The bot is constantly being trained to learn to answer new questions that come up.

“There are questions people are asking that the bot is not trained to answer because DBS never thought about this question,” Oren said. “By analyzing the interactions, we’re able to give them a report on an ongoing basis about what people are asking about.”

For instance, if 20% of the people who have been handed over to a live agent are asking about something that the bot doesn’t know, the bank might want to add new questions and answers about it into the bot’s body of knowledge.

Phased rollout

DBS deployed the virtual assistant first within the digibank mobile app in India. It took eight months to get from proof of concept to the April 2016 rollout. In August of last year, Kai was integrated with digibank India’s website.

In December 2016, DBS put the virtual assistant in Facebook Messenger for DBS Singapore. There, less than 1% of DBS customers who have used the banking bot have requested callback from a customer service rep, Lal said.

DBS has plans to extend Kai-powered assistants to other messaging apps like WhatsApp and WeChat in the future, given many consumers are already spending significant time in their chat apps.

Also in December, DBS launched the virtual assistant in Indonesia on the digibank app there. It’s been localized to converse in Bahasa Indonesia.

“This is a bankwide, multichannel virtual assistant that lets you do your banking across touchpoints,” Oren noted.

Typical bot chats

In a typical virtual assistant “conversation,” there are six exchanges — the customer sends at least three messages to the bot and the virtual assistant responds to each one.

Customers often seek information about product features, fees, current offers, how to open an account, and the locations of ATMs, branches and self-service banking machines.

They also use the virtual assistant to check their account balance and transactions, and to filter and find transactions in their account — for instance, “How much did I spend last week?” or “Show me the bills I paid last month.”

Another popular request is to pay a person. Customers enter a natural-language command, like “Pay John 100 rupees.”

Although customers occasionally think Kai is a human, the chatbot is designed to not fool people into thinking they’re talking to a person.

“We at Kasisto believe that the line should not blur, and it should be clear to people when they are talking to a bot and when they’re talking to a person,” Oren said.

For one thing, he said, the expectations of a conversation with a bot and a conversation with a person are different.

“When you’re interacting with a bot, you expect immediate answers,” Oren said. “You’re not going to even accept a second delay because it’s a bot. When you’re interacting with a person, a second where they write the answer is acceptable.”

On the other hand, Kai might occasionally say something like, “I’m a bot, I’m only learning, I don’t know the answer to that specific question yet.” Oren said people would tolerate that kind of response from a bot.

When Kai hands a conversation over to a live agent, it tells the user they’re going to talk to a person. If the agent hands the user back to the bot, they’re clear about that.

“We think that creates trust with the user because we’re not trying to make false claims or blur the lines for who you’re talking to,” Oren said. “We’re pretty strict about that.”

Editor at Large Penny Crosman welcomes feedback at penny.crosman@sourcemedia.com.

For reprint and licensing requests for this article, click here.
Mobile banking Digital banking Artificial intelligence Virtual assistants Bank technology BankAI Conference
MORE FROM AMERICAN BANKER