As banks around the world struggle to communicate effectively with customers quarantined in their homes, Isbank in Turkey has seen almost six million of its mobile banking customers use Maxi, its virtual assistant with a personality.
The Turkish-speaking Maxi debuted in November 2018 to serve customers of Isbank, Turkey’s largest private bank, by answering questions about their bank accounts, facilitating bill payments and money transfers, and dispensing personal finance advice through talk or text.
Then, after the novel coronavirus spread to Turkey in March, Maxi started helping customers navigate their loan deferrals and learn about precautions the bank was taking.
The conversations between Maxi and Isbank’s customers, which take place over Isbank’s IsCep and Maximum Mobil mobile apps or on WhatsApp, jumped 70%, from 1.8 million in February to 3 million in April.
Quick coordination between the call center (where many questions were redirected after branches closed) and the conversational banking team meant Maxi could take on these new responsibilities with less than a week of training.
But this fast turnaround was also possible because Isbank spent many more months training Maxi when the assistant was first conceived, with a comprehensive strategy toward collecting data.
The same strategies — painstakingly collecting data from real people, quickly analyzing new customer concerns as they come up, and uniting data science, design and product employees into one dedicated conversational artificial intelligence team — can be replicated at U.S. banks as well. To date, virtual assistants are scarce in U.S. banks as they try to figure out how to make them work flawlessly and usefully. But all banks are trying to find the right tools to interact with customers at a time when person-to-person interactions are still not possible in most states.
Training a virtual assistant
Halim Memis, unit manager of digital banking at Isbank, likens conversational artificial intelligence to the “original way of interacting,” where bank customers would talk to a teller directly.
To replicate that experience with a virtual assistant, the bank started developing Maxi in early 2018 with Clinc, a Michigan-based conversational AI software provider that counts a top-five U.S. bank among its clients.
Clinc allows clients to either license a platform they can build on however they like, or — as Isbank did — take its fully built and trained chatbot, Finie, and customize it and integrate it into their apps or messaging services. Finie can handle matters related to balances, transactions, spending history, locating an ATM, making a transfer and more.
But the crux of Clinc’s product is an ability to analyze speech patterns, word structure and sentiment, so these virtual assistants can understand and respond to everyday human speech, even if a question is posed informally or phrased indirectly.
“We train our customers on the art of building good conversational AI,” said Johann Hauswald, chief customer officer and co-founder of Clinc.
To imbue Maxi with a unique personality, Isbank held 14 focus groups to gauge what types of traits and skills bank customers wanted in a virtual assistant. To nail the voice, Memis and his team scouted speakers from around the world before settling on a Turkish woman who was based out of the United Kingdom.
For more than a month she spent eight to nine hours a day in a studio reciting sentences related to banking tasks. Now her digitally altered voice is featured in
Isbank’s conversational banking team came up with these sentences by considering the way real people would phrase their needs. On Clinc’s recommendation, the team paid participants on crowdsourcing marketplaces such as Amazon Mechanical Turk to supply different ways they might express the same questions, such as a request to view their balances (“what is my balance,” “how much money do I have in my account,” “show me the cash in my account”) or pay a bill (“pay my bill,” “bill payments”).
“These models need lots of data that represent how people would talk to it,” said Hauswald. “Crowdsourcing gives you diversity and coverage as to how people could potentially ask about a topic. The better your data and models, the higher your understanding of what customers are asking.”
For the initial training before Maxi’s launch, Isbank collected more than 10,000 sentences and 220,000 words. A question like “how are you” may require only 100 related sentences, while more than 2,000 sentences were needed for money transfers.
The goal was to make conversations with Maxi as natural as possible, rather than sending users through a decision tree of options.
“When you are talking, there is no such thing as offering options like ‘press A or B to continue,’ ” Memis said.
Isbank also personalized its virtual assistant by integrating customer data so Maxi can interpret more specific questions and deliver insights into spending.
For example, if a customer asks, “How much did I spend last year in January in Barcelona?” Maxi will list the relevant transactions. Or if customers ask Maxi, “Can I spend $500 on dinner tonight?” Maxi will examine their monthly spending habits and decide if they have enough money to splurge.
To date, 5.7 million of Isbank’s 8.2 million mobile banking customers have used Maxi.
Enter COVID-19
When Isbank closed its branches as part of its pandemic response, customers flooded the call center with questions, many of which related to the pandemic.
Using logs from the call center, Isbank’s conversational banking team grouped the queries into four categories: protective economic measures the bank was taking (such as alleviating certain fees); loan debt postponement; credit card debt postponement; and coronavirus precautions, or what Isbank was doing as a precaution against the virus as part of Turkey’s larger efforts to halt the spread.
The team trained Maxi to list the service hours of a branch, explain how to apply for a loan deferral online, and more. The most common questions pertained to postponing loan payments, as well as Isbank’s increased spending limits for contactless payments, said Memis.
Hauswald noted that Isbank’s ability to react quickly to market shifts has helped them adapt to customer needs.
“We recommend that even before launching experiences to your end users, you should have a team at the ready to look at what customers are asking about and do analytics to group those topics,” he said. “You want to hit the wave of initial interest.”
