Chatbots have come a long way since the original Eliza, but they still have a long way to go based on the underwhelming performance of contemporary virtual assistants like Amazon’s Alexa and Apple’s Siri.
AI-based tech firm Kasisto believes it’s doing better by taking the lead in conversational intelligence built on technology decades in the making. It was spun out of SRI International in 2014 and is backed by firms like Oak HC/FT, Propel, Two Sigma, and Mastercard.
In H2 2017 alone, major banks such as Canada’s TD Bank and the multi-national UK bank, Standard Chartered, joined Kasisto’s roster of customers that already included Singapore’s DBS Bank ($320 billion+ in assets), Wells Fargo, and Mastercard.
Dror Oren, Kasisto’s co-founder and chief product officer, talks to MarketBrains about the AI that makes their bots and assistants converse in natural language, and the future of Kasisto’s conversational AI platform for finance, KAI Banking.
MarketBrains: We hear a lot about the poor performance of chatbots, how do you avoid the major pitfalls?
Dror Oren: The way that our platform, KAI was really built addresses a lot of the shortcomings of the existing solutions out there, and that means you can build bots and virtual assistants that understand granular intents so you can scale the complexity of what the bot can do both exponentially and easily.
Other platforms will support 100 questions, but once you get to 2,000 questions and 15 different banking intents, it’s very hard to do, it’s very hard to build context. What the bot can understand and do begins to fall apart pretty quickly when you try to scale the complexity.
We keep adding in more and more intents to our platform so it continues to get more fluent about banking. For example, out-of-the-box it can handle anything around accounts, transactions, money movement etc. From: how much money do I have? what is my available credit for my account that ends with XXXX? how much have I spent on a certain item? how about this year and last year? what was my largest expense in 2016, how much did I spend last weekend?
We architected our platform with the finance industry in mind and the requirements of an enterprise-ready solution are foremost. From the bank’s perspective, KAI is virtually a white box because we can integrate with any existing API in the bank and really add a conversational layer to it.
MB: Can you tell me more about some of your customers and investors?
DO: Mastercard, in addition to being an investor, also rolled out a pilot of what they call Mastercard KAI: it’s a conversational financial bot that has everything you can think about in terms of questions about the features and benefits of their credit card, but it also presents contextual offers.
DBS Bank also made an investment, and over the course of the past 18 months, the bank has deployed KAI in three markets (India, Singapore, Indonesia), three channels (mobile, web, Facebook Messenger) and two languages. The results are measurable – the KAI-powered bot handles 82% of all customer questions and inquiries and DBS created digibank with one-fifth the resources of a traditional bank.
MB: How are the other banks integrating it?
TD announced that they will be integrating KAI Banking into their mobile app, and people will be able to ask questions about banking. Standard Chartered will integrate KAI Banking into the website and their mobile apps, and they’re going to start in one of their markets, Hong Kong. Wells Fargo is running a large, production pilot on Facebook Messenger with a KAI-powered bot.
MB: What can you tell me about KAI’s underlying AI technology that allows for more natural conversations?
DO: There’s a bunch of AI that’s taking place. Our architecture is open and API driven.
I’ll begin with language: the ability to have very good level of understanding in context requires really advanced natural language understanding capabilities. For example, when I ask: “how much have I spent on Uber in 2017?” and then follow up with “how about last year?” or “how about Lyft?,” most bots will fail to understand the follow up question because there is simply not enough context for it in the phrase – they cannot relate it to the first question.
Most of the chatbots out there were built with really basic Natural Language Processing (NLP) engines, that are trying to understand what you just said, but have no context, no map, no notion of what their previous response, or what the user is trying to accomplish – their goal. . I challenge you to go and look at other bots out there, and try to ask follow-up questions.
The way our platform understand the user, maintains its context and the user’s goal, is solved by our hybrid approach to NLP, where we are running multiple classifiers, and trying to understand which one of them understands best, and then the winner takes over.
“…a bank cannot just say I became a neo-Nazi fascist bank bot because that’s how you talk to me…”
There are two sides to this: there’s the side of the machine learning that is happening in terms of the training of the data, and then there’s runtime AI that makes decisions every time the system interacts with you.
KAI is using machine learning, but when it comes to very serious and regulated industries like banking, KAI on its own cannot decide to change an answer, it has been trained to respond with without taking the bank’s rules and regulations into consideration.
Supervised machine learning and deployment is important for improving the experience and accuracy and there’s also regulations that go into play here. If you search Google or ask Siri a question and get a different answer every day, it’s not a big deal. If you get an answer from a bank that is not very accurate, it is.
MB: What are some of the regulatory issues you run up against?
DO: For a regulated environment, KAI comes with self-service tools that we supply to the banks. We actually do a lot of unsupervised learning, but we never do unsupervised deployment.
One example of unsupervised learning that went haywire was Tay, the Microsoft Twitter bot. These systems are very prone to skewed datasets: what you are feeding the systems is where they end up getting skewed towards.
However, a bank cannot just say I became a neo-Nazi fascist bank bot because that’s how you talk to me, that’s not a good approach.
So, banks and, in general, regulated enterprise industries, have to take different approaches when they train and deploy their bots. And if you’re an AI platform that caters for these kinds of industries you need to have the right tools to collect, federate and train the data and the right tools to manage the AI training process after it’s deployed in production.
“All customer conversations with the KAI-powered assistant are digitally captured, cleansed, and anonymized for data mining and audits to support compliance requirements of the bank.” Source: DBS digibank
MB: How are banks using all that data from all those conversations?
DO: Banks use our customer portal for its business dashboard, AI training and analysis, and content management system. They can go and see both business metrics, performance metrics and the content, see how many people are using it, the performance across segments, where they are coming from, what performance they are getting.
We also have a Reporting API so they can integrate the data into their existing analytics infrastructure.
They are using this data to constantly improve their customers’ experience – whether that’s improving the accuracy of the bot, teaching the bot about new bank products and services, or extending the platform to new channels like Facebook Messenger or Alexa.
MB: Any major surprises coming out of the data analysis?
DO: We have surprises all the time – conversational AI is an entirely new customer experience and there is nothing like production deployments and real customers when it comes to learning.
Here’s one example in the US, we had KAI-powered bots and assistants running during the Zelle announcement, the new payment infrastructure. Suddenly a large number of banks’ customers started asking: “Are you supporting Zelle?” That’s something the bank had not trained the system to understand, but through these customer interactions with the bot, they very quickly saw that their customers are interested in Zelle.
For another customer, their ATM network was down and people started asking the bot about it before the bank was even aware of it.
“We are looking very carefully at the specific use cases for wealth management and insurance.”
In these examples, even if the bot is not trained to answer these customers’ questions, the platform was able to create reports in realtime to show that something is going on, such as a spike in new questions never seen before. and quickly add new content to handle the situation.
As you can imagine KAI helps uncover a lot of operational issues because with a conversational AI platform, basically for the first time you are putting an open-ended text box in front of a customer, and saying just tell me whatever you want.
MB: Are you planning on branching out past retail, into the wealth management space?
DO: Yes, absolutely – we’ll be extending KAI beyond retail and commercial banking. We are looking very carefully at the specific use cases for wealth management and insurance. All of the banks that we work with have wealth management businesses and many also have insurance businesses so these are definitely on our radar. We will make an announcement later in the year.
This interview is edited and condensed