Maybe it’s chutzpah, maybe it’s a gentler regulatory climate, but Israeli banks seem quicker to launch artificial-intelligence-based virtual assistants than their American counterparts are.
With a few exceptions such as Ally Bank, U.S. banks have been cautious about launching virtual assistants that interact with customers, interpret their speech and texts and hazard answers. Several are testing the technology carefully over many months, and in some cases years, to reduce the chance it will commit a faux pas.
But Bank Leumi began offering an
“I don’t know if it’s our willingness to take risks, but we were certain that this was the way we needed to go,” said Arik Frishman, head of digital at Israel Discount Bank.
The bank partnered with two fintechs to create Didi.
Nanorep provided natural language processing in Hebrew. “That was a challenge for us,” Frishman said. “Most of the providers out there don’t really know how to tackle Hebrew yet.”
And Personetics provided Didi’s “brain” and predictive analytics.
Though she does chat, if you refer to Didi as a chatbot, Frishman will be quick to correct you. He sees Didi as a hybrid between a chatbot and a virtual assistant. Virtual assistants are smarter, more predictive, and can be used for multiple things, in his view.
At Discount Bank, one group worked on the natural language processing challenge with Nanorep, and another worked with Personetics in vetting the AI and predictive analytics.
Two months into development, the Nanorep group launched a public chatbot on the bank’s Facebook page.
“It was very limited: We could only answer inquiries about public things, like where’s your branch, what’s your exchange rate and so forth,” Frishman said. “But it was a great beta site for us to improve our [natural language processing] to see how Israeli customers approach such a bot, what words they use, how long their text is, how often they make mistakes while writing and so forth.”
The group that worked with Personetics went through the 150 ready-built “insights” Personetics offers that it has gleaned from work with other banks.
“We had a team working on that list and seeing which ones are most suitable for Israeli customers and where the data is reachable,” Frishman said. The bank made a short list of 30 insights and connected those with the necessary data sources.
One insight, for example, is drawn from an analysis of credit card transactions and expenses. An alert will tell the customer if there is a suspected double charge on her account, if her expenses are lower or higher than in previous months, if there’s an unusually high charge from a merchant, or if there’s a charge from a new merchant. Another is more forward-looking: With its knowledge of upcoming bills and deposits, Didi can warn customers that they are in danger of a low balance in a few days.
Six months later, the bank combined these two streams of work — the chatbot that understands Hebrew voice and text and the 30 premade insights — and ran a test with 100 customers. A month later, it invited 6,000 customers to try it. Two months after that, it rolled out the product to all customers.
“It was all a very quick, interesting journey,” Frishman said. “We’re happy with the process.”
In the two months customers have been using Didi — and more than 220,000 bank customers have — the bank has already learned a few useful things about how customers interact with these digital helpers. There may be differences between the preferences and mores of Israeli consumers and Americans, but some things may apply here as well.
1) Customers ask deeper questions of a virtual assistant than you’d think.
“If a customer has a loan, he will go into specific questions about refinancing loans,” Frishman said. “We covered loans [in Didi’s answer base], but we didn’t think they would go so deep.”
Customers will try to drill down into a customer service incident with Didi, saying things like, “I called an agent or my branch, and they didn’t call me back — why is that?”
“That’s a field that she can’t really answer,” Frishman said. “So we’re learning every day.”
Discount is rolling out a new customer relationship management system, to which Didi is connected. So if Didi can’t answer a question, the customer will be forwarded to a human agent. The agent receives a record of the conversation with the bot. This lessens the chance that the customer will need to answer a question more than once.
In a recent survey conducted by eGain, 59% of respondents (62% in the U.S. and 55% in the U.K.) found that having to repeat information and context to a human agent after handoffs from chatbots was the biggest hassle by far in using virtual assistants. This is a result of chatbot deployments that are disconnected from agent assistance. The second biggest deterrent was chatbots "getting stuck and not knowing what to do next" at 32% (29% U.S. and 37% U.K.).
2) Customers prefer texting over any other mode of communication.
“When we speak with our customers, we understand that they prefer texting instead of speaking,” Frishman said. At the end of the day, this will probably help the bank decrease costs.
3) Customers would rather ask for something than look for it.
“That’s the most common answer that we get, and it’s pretty understood,” Frishman said. “Those are expectations we can meet with Didi; that’s one thing she gives us.”
Asked to rate their experiences with Didi at the end of each interaction, 81% of the time customers give the bot a high rating.
Discount Bank’s road map for Didi includes adding more prefabricated insights and providing the technology to customer service reps, so they can get quick answers to customers’ questions from the system.
Overall, Discount Bank’s hope is that Didi will help increase customer engagement, but also minimize customer service inquiries.
“If I can tell you two days in advance what’s going to happen, you probably won’t call at all,” Frishman said. “Didi’s predictive capabilities and forward-looking things we hope will prevent inquiries irrelevant of who’s going to take care of them, a human or a bot.”
Editor at Large Penny Crosman welcomes feedback at
