AI:
How Artificial Intelligence can Speak to Customers

Michael Krigsman: What is conversational AI anyway? It's a big topic, an important topic and, today, on CxOTalk, we are exploring this subject with somebody who is running a large program at a major financial institution.

I'm Michael Krigsman. I'm an industry analyst and the host of CxOTalk. I want to say a grateful thank you to IPsoft. We are in New York City in their AI experience lab. It's companies like IPsoft that make CxOTalk possible.

We're speaking right now with Ken Dodelin, who is the vice president of conversational AI products at Capital One. Hey, Ken, how are you?

Ken Dodelin: Great. How are you, Michael?

Michael Krigsman: I'm good. Tell us about Capital One. I think we know the name but tell us what you guys do and your size.

Ken Dodelin: Sure. Sure, so Capital One is a large bank in the U.S.--credit cards, banks, auto loans, all those sorts of things--and has been leaning heavy into the digital space and, in particular, into the intelligent assistant space, which is what we're here to talk about today.

Michael Krigsman: Now, you're the VP of conversational AI products.

Ken Dodelin: That's right.

Michael Krigsman: Tell us what that means and what you do. In practice, what does that involve?

Ken Dodelin: Sure. We build digital products. In my case, they're in the conversational AI space, which means that they have artificial intelligence behind them and that they interact with customers in a conversation. Very often, customers can speak or text in natural language and interface with Capital One through an AI. Sometimes it's our own, which we call Eno. We'll talk about that in a bit. Sometimes it's through a parent AI, like an Alexa, Cortona, or something like that.

Michael Krigsman: Now, Ken, as the VP of conversational AI products, what's the scope of your responsibilities?

Ken Dodelin: Sure. The products I just mentioned are all live today. We were the first bank to release a skill on the Amazon Alexa platform that enabled customers to ask Alexa about their accounts and get information, even pay their bill all through voice. We extended that skill over to the Cortona platform run by Microsoft.

Last year, we turned out attention a little bit towards the text space, and so we were the first bank to release a natural language chatbot or intelligent assistant through the SMS channel. Customers can text Eno, our intelligent assistant, and ask questions about their accounts, when is their bill due, pay their bill, and so forth, all through the familiarity of what we all do every day, which is texting.

Michael Krigsman: Why is conversational AI so important to Capital One?

Ken Dodelin: Yeah, it's a few things. First of all, Capital One is a very tech-forward company. We're the first bank to go to the cloud. We're the first, or one of the first, banks to have an API-based infrastructure that powers all of our digital experiences. We're always looking for ways to use technology to improve the customer experience.

In conversational AI, it used to be that you could either come to our graphical user interface, which would be an app or a website, where we, in some ways, try to guess what it is you're looking for, right? You have a certain finite number of buttons and links to click on or tap on. In the conversational interface, it's whatever is on your mind. You can just text in natural language, and then we're able to start with that rather than starting with all the guesswork. It's a great complement to GUI experiences, and it's also a great way to interact with customers without having to pull them into our uber website or app.

Michael Krigsman: Is it mostly a function of user experience? Is that your goal, or are there other goals as well?

Ken Dodelin: Well, there are fantastic business and customer benefits to conversational AI. There are three things going on here. The first one is, we have a recent emergence in natural language processing technology. Not too many years ago, we just couldn't understand customers, what the intent of their spoken or typed utterances were in the way that we can today. That advancement has enabled these experiences to reach a threshold where they're actually useful.

The second thing is the availability of data. Our ability to get answers and connect into our API infrastructure, the same one that powers our websites and apps, is at a place now where we can use real-time context to adapt the conversation, so it is more conversational and not just a predetermined conversation that wouldn't feel very human.

Then the third thing, in addition to the NLP and the data, is the proliferation of the Internet of Things devices. Whereas in the not too distant past it was pretty much a website or an app that you were going to interact with Capital One through, or else you were going to call the call center. Now we have things like Alexa that use voice-enabled touchpoints with the customer within their car, in their living room, or wherever they put those connected devices that enable interaction.

Also, since we don't have to have a GUI come with us wherever we want to interact with customers, since we can do it simply through natural language, we can go to places like text messaging, messaging apps, and other similar things where the interaction is all just natural language. Even emojis, right? Who would have thought you'd be paying your bill with a thumbs-up emoji, but that's what we've enabled, and customers have gravitated to?

Michael Krigsman: Ken, I certainly can see very quickly the user experience value here, but can you elaborate on the business value as well?

Ken Dodelin: Sure. Every company that has a large customer service need is looking for ways to get customers to A) not have problems and B) be able to self-service in digital. There are business benefits to that. The other one is, there are lots of micro examples where this is helpful. You can imagine people coming to our websites and, instead of just having to look around and not finding what they need, they can actually ask and get to what they need to do, which leads to a lot of desired business behaviors. They're actually paying their bill; they're actually enrolling in paperless statements; whatever it is that is a desired customer behavior.

I'll give you another one that's kind of interesting. Like every bank, we have a fraud alert that goes out. That fraud alert is very rigid just because of when it was created where it will say, "Hey, we noticed a suspicious charge on your account. Text us back "confirm" or "deny" whether or not it was yours. If you text back "confirm" or "deny," everything is great.

If you fat finger it, you make a typo, or you try to say something like "yes" or, as we've seen in our logs, "Yes, that's the pair of shoes I bought when I was visiting my sister in Philadelphia last weekend," we don't understand any of it." You can imagine the business benefit of being able to understand these things that the customer is saying so that, if there is a fraud event, we can turn off the card and take corrective actions more quickly so that the fraud stops.

Michael Krigsman: Now, maybe this is a very basic question, but what's the difference between the type of AI that you're talking about that has access to data and has built-in intelligence? How is that different from the simplistic chatbot that every website has got these days?

Ken Dodelin: Well, a few things. One is, we're very niche, right? We are building an intelligent financial assistant. We found when we were building Eno, we used some off-the-shelf natural language processing solutions that were okay. We gravitated to, "Hey, if we train it on banking, we'll find that it's better at banking than the off-the-shelf ones." It's terrible at everything else. It doesn't know sports and things like that. But, we really home in on the types of things that our customers are looking to do, and we have very short conversations about prioritization because we know exactly what customers want to do with Eno by what they text in, right? You're always guessing on a website and such, but we have great data on that.

The other side of how it's different is that Eno will reach out to you. We like to say the call center rarely calls you as a customer, but Eno will reach out to you in the moments that matter in the channel that works for you because Eno is always minding your money and thinking about it even when you're not. The same things you would want from a human personal financial assistant will be accomplished through our virtual one.

Michael Krigsman: Your development goal then was to create a very narrowly focused AI that had a great deal of depth inside that narrow focus about which you care.

Ken Dodelin: That's right. That's right because, look, I have great empathy for the folks who are at the big five tech companies working on Siri, Alexa, Google Assistant and some others where they're trying to answer any question about anything, which is a really hard problem to solve. We have a more narrow focus and can achieve, I think, some higher levels of success in our niche because we're so focused.

Michael Krigsman: Now, you mentioned that there's a combination of products that you've developed in-house, as well as external products. Is that correct?

Ken Dodelin: We've worked with various vendors along the way and come up with a mix of homegrown technology and some other things. Look, we generally build things ourselves at Capital One. We leverage the fact that we're in the cloud, we have this API-based infrastructure, and sort of a passion for open source software. We tend to leverage the great talent we have on the team to build things and let the customer experience guide us.

Michael Krigsman: Your enterprise, your underlying enterprise architecture facilitates what you're doing with AI.

Ken Dodelin: Absolutely. The same technology that powers our websites and apps and authentication is leveraged in the conversational AI products that we release. The part that's new is that the inbound activity by the customer comes in natural language and the response comes in either voice or text or a mix of maybe text and GUI based upon where you're interacting. We have our Cortana skill, our Alexa skill, and our Eno SMS chatbot all running on the same platform right now, which is great.

Michael Krigsman: Did you hire people to do things like natural language processing?

Ken Dodelin: We haven't gone public with how we've built each component of the Eno experience. But, again, we went out and looked at what we were trying to do from a vision standpoint and said, "Hey, there are some places where we feel like we have the expertise and other places we want to bring it in."

I'll give you a couple of examples. To work on natural language processing, we brought in one of the lead engineers who helped build the first AI to win on Jeopardy. When we were thinking about the character of Eno, we kind of looked around and said, "Well, we're a bank. We don't really build characters," and so we went out and hired a woman with a lot of experience in working with characters from Pixar to come in and give us a backstory for Eno, some character traits, and some consistent personality so, as you interact with Eno, you don't have one personality in one place and one personality in another.

Michael Krigsman: That's really interesting.

Ken Dodelin: Yeah, it's been a fun journey. We certainly had to stretch our skillsets to succeed in this space.

Michael Krigsman: It sounds, as well, like this is a strategic mandate of the organization to develop these types of technology products. What are the kind of metrics related to AI that you track?

Ken Dodelin: There are a lot of similarities between what you would track on mobile apps, for example, but one of them we look at is a customer success rate. We want to know, customers who start interacting with Eno, whether it's through a proactive outreach by Eno or a customer's inbound outreach, did they achieve their goal? We look at--whether it's a direct answer from Eno or Eno steering them to a certain place--how can we continue to improve the ability for Eno to provide the customer with what they're looking for?

Michael Krigsman: That's a function both of the technology as well as the data and the way that you're using the data?

Ken Dodelin: Yeah. Everything is data driven. There is some subjectivity sometimes around the edges but, like I said, we have short conversations about what the customer is looking for now because there's just such a great insight into what the customer is thinking at any given time.

Michael Krigsman: Now, what about your call centers? You mentioned that this is a complement to your existing contact centers. The contact centers remain unchanged? These are just additional channels that customers can talk with you?

Ken Dodelin: That's right. I think the call center folks are actually excited about the deployment of machine learning to enhance what they're doing. Certainly, anything that can predict what types of issues a customer might be calling about to help route them more efficiently to the agent who can give them their answer to cue up the right systems that get them to the information they need all is value-add and allows them to spend more of their time doing the higher value human side of servicing customers versus trying to do that simple navigation stuff.

Michael Krigsman: The level of investment that you've placed here obviously is very significant, which means that you must have ways of evaluating your progress and evaluating how do we know if we're being successful here or not.

Ken Dodelin: Mm-hmm.

Michael Krigsman: Could you share with us some of those mechanisms for evaluating what you're doing?

Ken Dodelin: Sure. Not unlike most of our products, our digital products, we want to know, do people use it; do they love it?

Michael Krigsman: Adoption being one?

Ken Dodelin: Sure. What's interesting is that, when we launched Eno as an SMS intelligent assistant, it's different than launching an app in that, when you launch an app, it comes with all the infrastructure of, well, you're in an app store now and you have ratings and reviews and places where people can leave those ratings and reviews. We didn't have that, but we wanted it, so we actually stood up our own.

We solicit customer feedback, and we post it up on a webpage. The feedback so far has been surprisingly positive for a product that's in a nascent stage. I think it has north of 4.5 out of 5 stars, and we see people enjoying the fact that it's another way for them to get access to information very quickly, get their questions answered, and they can ask, particularly if it's something like, "Well, this might be on the website. I'm not sure. Let me just ask, and I can find out if it's there or not before I invest the time it takes to make a phone call or what have you."

Michael Krigsman: Are there specific technologies that you see coming down the pike that are important?

Ken Dodelin: Sure. I think there's the natural language processing, which is very good at deriving a customer's intent, where you can then go and give them a response. There are a lot of advances starting to take place along dialog management and natural language generation. These types of things are still in the formative stage, but they allow conversations to not just be natural from a question and answer standpoint, but more of a conversation that can go in many directions, which is more of how human conversations go.

We're excited about those. We're monitoring them. We do some of our own research on them. I think they'll be a driving force in the years to come.

Michael Krigsman: We've been talking with Ken Dodelin, who is the vice president of conversational AI products at Capital One. Ken, thanks so much for taking time with us today.

Ken Dodelin: Thanks for having me, Michael.

Michael Krigsman: Everybody, go to our website, don't forget to subscribe on YouTube, and I want you right now to tell all of your friends and your family that they should watch, and they should subscribe too.