Getting computers to talk finance with banking customers by voice or text is a big challenge, as anyone who has encountered a bank chatbot knows. Called conversational artificial intelligence (CAI) it requires skills outside the usual developer’s portfolio and changes to the typical IT organization, said Don McInnes, who has been working in the field since 2011.
He led the Conversational Design team at a major bank for about a year before concluding the bank would lag behind because it wouldn’t take full advantage of commercially available software.
“I was frustrated because I couldn’t seem to help people understand how much value there is in learning the hard lessons of Conversational Design by using existing products. Things in this space are moving so quickly that unless you are a software company, specialized in - and dedicated to - refining CAI capabilities, you’re going to fall even further behind with every passing day,” he said.
The bank did not respond to a request for comment.
Many organizations seem to believe that the only way they can get a solution to address even the simple questions in banking, like finding an account number, checking a balance or disputing a transaction, requires building it themselves.
“To make them work you still have to analyze your own data. You have to decide which use cases that you’re going to address, McInnes said. "For customer service bots and FAQ-type bots, you have to figure out what answers you’re going to give to different questions, sometimes a surprisingly challenging task. You have to decide what the conversational interaction is going to look like in terms of tone and personality. Then you need to train the AI which requires even more data, special knowledge and skills. The particular platform that you’re going to use has little influence on all of this work.”
From late 2014 through 2017 he worked with IBM’s Watson team at Autodesk to develop a Virtual Assistant for customer service.
“Our small Autodesk team became sponsor-users of Watson - which meant that IBM’s UX team met with us to learn about what we were doing with the tool. I also interacted heavily with their product managers -- to help them evolve the platform. Many of the features in Watson today are there as a result of our early work building AVA.”
When McInnes joined the bank in 2017 he was one of the few on the team who had experience actually building and deploying CAI solutions. He found much of the discussion in planning covered ground that was familiar to him but was new to most of the others on the team.
The bank eventually interviewed CAI vendors. McInnes, who was part of the evaluation team, was delighted to hear Clinc present the concept of conversational design, because it showed that they didn’t treat it as just a technology problem.
“I think that Clinc is unique in that they understand that technology alone won’t make a successful virtual assistant.”
The bank licensed Clinc, but continued to stress in-house development so McInnes left to join Oracle as a senior product manager working on digital assistants, and his lead developer left to join Clinc. Since McInnes’ departure, the bank has increased its engagement with Clinc. The bank is now building on Clinc’s self-service AI platform to accelerate their capabilities while still developing in house too.
Jason Mars, co-founder and CEO of Clinc, said conversational AI is just at the frontier of its potential. The company has 30 million people using its technology, including six million customers of Isbank in Turkey who ask about balances and transfer funds with Clinc.
Clinc’s AI is also being leveraged by one of the largest banks in the UK. The bank is powering its Facebook Messenger bot with Clinc’s AI and successfully handles over 25,000 queries per week at a 92 to 96 percent completion rate, which, if handled in the call center, would cost an average of $10 per query. This amounts to an estimated $10 million dollars in savings for the bank since implementing Clinc’s AI.
Voice can also be very useful in fraud cases, Mars said. A user who has spotted a weird transaction or fears their password has been compromised can talk to a virtual assistant to report it. The virtual assistant could ask what happened, what the user did, and then ask whether the user wants to turn off a card or suspend an account.
“Chat can say to send us the email and we will investigate, or if you did give info, it will ask what info you gave. It can work through many types of fraud without a human in the loop. In one week, we had 100 percent containment with Facebook Messenger.”
In-house development has failed across the board, Mars added.
“This is a tech which is extremely hard; Google has teams of Ph.D.s in computer science working on it. Executives say they have hundreds of developers and they can build apps." But that doesn't mean they can build high quality, versatile virtual assistants.
The vast majority of banks try it in-house and quickly learn that they must explore vendor offerings, he added.
Conversational AI is an emerging market. Mars expects 2019 and 2020 will see massive activity because companies are spending seven figures to get conversational AI for activities like food ordering in the quick service restaurant business. (See the Clinc website for an example of a customer at a fast food drive-through who epitomizes the kind of customer you don’t want to get stuck behind.)
Firms may need to modify their organizations to achieve effective conversational AI, said McInnes.
“A bank may have a group devoted to AI, another to customer experience and another to the website — but moving to conversational AI requires bridging or eliminating those silos. Some of the biggest challenges that I’ve run into are at organizations that look at conversational AI as purely technology driven. Then the customer experience side gets overwhelmed by the tech leadership. Of course tech is a key stakeholder and part of the project because most meaningful use cases will require some back end integration. But they can’t dominate the whole project.”
If tech experts can be a problem in designing conversational AI, so can subject matter experts who want to address hard problems, like, at Autodesk, moving bearing walls.
“Early on in our experimentation with AI at Autodesk, there was hope that we could use it for really complicated issues. But what we quickly found is that the most common customer service calls came from users who couldn’t get their program to work because they were trying to use a module that they hadn’t bought a license for — something a virtual assistant could understand and address.” Just resolving that issue could reduce 30% of the call burden on people in customer service, but it didn’t much interest experts because it was so easy.
“How people interact with your virtual agent is often very different from what the internal experts predicted.”
(For another example of CAI, see Thomas Friedman on how an Indian call center has moved from people talking to using text and managing conversational AI.)
