UBank eyes potential voice interfaces for AI chatbot

How UBank put Watson to work

NAB’s UBank is experimenting with new interfaces for its online RoboChat service, with Facebook Messenger, Amazon’s Alexa and Google Home all considered potential front-ends for the chatbot, according to Jeremy Hubbard, the bank’s head of digital and technology.

UBank launched RoboChat in May this year. The service, which can help guide a customer through a home loan application, is built on IBM’s Watson cognitive computing platform.

UBank’s partnership with IBM for RoboChat kicked off with a hackathon in mid-2016, Hubbard yesterday told IBM’s Watson Summit in Sydney.

“In a really short period of time we came up with more than 20 applications that we whittled down to just three for the purposes of the hackathon.”

Those three potential AI-based applications were then prototyped during the hackathon.

“One of those three options was a chatbot for home loans, which is of course where the idea behind RoboChat was born,” Hubbard explained. “Why home loans? Well home loans are a key focus for UBank and it’s a fairly complex process compared to some of our other processes.”

The core value proposition of UBank is “simpler, better and smarter” and a chatbot could help make that a reality for home loan applications, he said.

“We could see that a chatbot could really help customers through this process,” Hubbard said.

The UBank site already featured live chat to help customers, and an analysis the data derived from that service revealed that around 80 per cent of the questions customers asked were concentrated in just 20 per cent of  topic areas.

For the chatbot project the UBank project team whittled those topics down to 44 key areas and thousands of associated questions.

A group of less than a dozen people across UBank and IBM teamed up for the six-week initiative.

“Six weeks may sound pretty short, but I’m here to say that it’s certainly possible with cloud technology and a really narrow focus,” Hubbard said.

A key first step was to train the whole project team in “UBank’s tone of voice,” he revealed.

“Our tone of voice is ‘simply helpful’, ‘disarmingly honest’ and doesn’t use any bank jargon — and we needed to make sure that RoboChat was no exception,” he said.

“We paired this work with a conversation flow expert that helped us construct natural, flowing conversations. Now this might sound trivial but there’s a real art to creating a conversation that flows well.” (“And I’m sure that anyone who’s ever sent a sarcastic text knows just how important it is to get that right,” he added.)

The next step was building up the bot’s knowledge base.

“Now of course RoboChat is powered by Watson’s artificial intelligence engine but it didn’t know anything about UBank or UBank’s products, so we had to teach it,” Hubbard said.

That process covered three key areas: The technical aspects of UBank’s home loan products; risk and compliance (including, for example, threats that might come through to the bank via the bot); and “chit chat” — “the sort of off-topic conversations that often happen with a chatbot”.

(With regards to the third of those areas, Hubbard revealed that after RoboChat went live two of the most common questions it received were “What is your favour colour?” and “Will you marry me?”; RoboChat’s response to a marriage proposal is “I’m definitely flattered, but sadly, only available online.”)

In parallel to that teaching process was the process of building the technical architecture for the system, Hubbard said. The team built a Node.js-based orchestration layer on IBM’s Bluemix cloud platform to sit between Watson and LivePerson’s cloud-based LiveEngage platform (which the bank uses for agent-based chat).

“It’s fair to say with the timelines that we had this was the area that I was most nervous about, but it went extremely smoothly,” Hubbard said.

Plugging Watson into LiveEngage meant that from a customer experience perspective the process of engaging with RoboChat was the same as standard live chat (although RoboChat openly acknowledges that it is a chatbot) and from an operational perspective UBank could leverage the chat platform’s features (such as rostering the bot on for certain hours and a variety of operational reporting).

Testing took place in multiple phases; initially with the project team and later with friends and family, Hubbard said.

“I can say that RoboChat improved an awful lot through this period,” he added. “Despite all of the prep work that we did, that testing and training was really important and that real customer engagement with friends and family was particularly key.”

“We got a bit of a false sense of security with that first phase of testing that we did and that’s because we tested predominantly with the UBank team and our project team,” he said. “And, of course, that team was all familiar with what we could call brand UBank or the Ubank lingo and they couldn’t unlearn that knowledge — and that is just so different to how a customer talks.”

Testing with real customers was key and RoboChat has continued to learn based on its customer interactions, he added.

Deciding when to launch an AI-based service can be tricky, Hubbard said, because they require real-world data in order to become more effective.

The team launched RoboChat once it was “good enough” and staged a one-week pilot. RoboChat had a “tough” first day, but it was retrained daily — “and by day three it was a completely different experience,” Hubbard said.

One of the next steps for UBank is likely to be leveraging IBM’s Tone Analyzer service, Hubbard said. IBM Watson Tone Analyzer attempts to understand the sentiment of a text-based conversation.

“A simple example would be looking to see if a customer is frustrated in the process and … transfer them to a live agent,” Hubbard said. (Currently a customer can request to be transferred to a live agent, and if RoboChat fails twice to answer a question it will offer to transfer the conversation.)

Another step for the bot that UBank is excited about would be having it potentially complete a home loan application on behalf of a customer, Hubbard said.

“I’d say for UBank this has been a really great first step in AI but recognise that we’ve barely scratched the surface of what’s possible,” Hubbard told the conference.

“At UBank we’ll continue to explore and innovate beyond the boundaries of banking and we see artificial intelligence of a really key way to help us do this,” he said.

Join the newsletter!


Sign up to gain exclusive access to email subscriptions, event invitations, competitions, giveaways, and much more.

Membership is free, and your security and privacy remain protected. View our privacy policy before signing up.

Error: Please check your email address.

Tags Case StudyCase studiesUBankIBM Watsonmachine learningcognitive computing

More about AmazonBluemixFacebookGoogleHomeIBMLivePersonMessengerNABToneUBank

Show Comments