By Sara Castellanos
Capital One Financial Corp. wants to be both a technology company and a bank.
Since 2011, Chief Information Officer Rob Alexander has expanded his technology staff to 9,000 from 2,500, hiring hundreds of software engineers, developers and artificial-intelligence experts to develop products such as a recently launched digital chatbot. He also helped spearhead an effort over the past three years to move the bank's back-end software development tools and infrastructure to the public cloud. That way, he says, the bank's technical staff can get software products to market faster, more easily and with top-notch security.
The rate at which Capital One is relying on the cloud, embracing technology and competing for AI talent is "surprising and remarkable," says Lex Sokolin, global director of fintech strategy for financial-company research firm Autonomous Research.
Mr. Alexander spoke with The Wall Street Journal about the development of the bank's chatbot, Eno, and the challenges and advantages of operating like a technology company. Edited excerpts follow.
WSJ: Did you run into any challenges creating the Eno chatbot?
MR. ALEXANDER: In 2016 we experimented with and piloted a version of Eno that was built on a commercial third-party natural language processing AI engine. We quickly came to the conclusion that while the performance was good, it really has to be great for customers to use it.
The off-the-shelf natural language processing models weren't good enough, and we had to start from scratch and build our own because we care about accuracy in two ways. One, Eno needed the ability to understand the intent of a request, and then the functionality to execute on that intent. The first version of Eno didn't understand the intent well enough.
For example, we calculated 2,500 ways a customer might request his or her balance, including misspellings and things that autocorrect does. The model is constantly learning what those variations are and what the customer's intent was.
WSJ: What have you changed about Eno over time based on customer feedback?
MR. ALEXANDER: If the chatbot can't help customers with things that are important to them, or they get generic responses, they won't use it.
A lot of customers ask Eno to remind them before their bill is due on their card. We have trained Eno to understand that request, and then automate the process of setting up reminders for customers in advance of when their bill is due.
Other examples of that proactivity would be if we see fraudulent or suspicious activity on your account, Eno can automatically reach out to you through the mobile app, SMS text or email. If we see a double charge for the same merchant right away or an unusually big tip on a restaurant bill out of pattern, Eno will send a notification or email or SMS text to a customer that says, "Hey, this looks suspicious, did you mean to do that?"
WSJ: You said there are 2,500 ways a customer might request their balance. What common customer phrasings did it not recognize before that it does now?
MR. ALEXANDER: Some examples of this include "How much money is in my account?" This is an example of an inquiry that doesn't directly use the word "balance," but Eno understands the intent behind the question.
"How much have I racked up on my card?" This is an example where a customer is speaking in natural language, in a way they would think about balance, despite the context being a bit unclear.
WSJ: Can you provide some examples of things that Eno wasn't doing well that you changed, based on feedback?
MR. ALEXANDER: As we started in our web application, we noticed customers typing in very lengthy messages. This would confuse Eno, due to multiple intents, and extraneous information. So we did a design change: Up front, Eno says that it does better with shorter questions. This reduced customer friction, and Eno has responded better. Typically, we prefer to adapt the tech to the human behavior, but in this case we saw an opportunity to influence human behavior in a way that set both Eno and the customer up for success while the technology continues to evolve.
WSJ: What makes Eno different as a result of it being created by an in-house team of technologists?
MR. ALEXANDER: One important point to call out is that Eno wasn't just built by a team of in-house technologists. It was built by an integrated team of product, tech and design professionals, all contributing to the project's success from start to finish.
Eno's personality is a great example of the outcomes from that operating model. When we set out to design the intelligent assistant, we recognized that character development was a capability we didn't have -- so we hired filmmakers, anthropologists, journalists and designers with deep experience in character development to help define and shape how our intelligent assistant would connect with our customers. Today, we have an incredibly talented in-house AI design team that is deeply focused on creating experiences that lead to contextually relevant, meaningful conversations and relationships rooted in trust, empathy and understanding.
For example, our decision to design Eno as a gender-neutral intelligent assistant was made by that internal team.
WSJ: What's an example of a challenge you've had to overcome in the transition toward operating more like a technology company?
MR. ALEXANDER: It's really about how to attract and retain the type of talent that you need, and in quantities that are required to fulfill on that mission of operating like a technology company. Today we hire software engineers, data engineers, machine learning and cloud infrastructure and cyber engineering. That's the toughest challenge in this whole thing, because you're trying to attract a kind of talent that's in high demand in the marketplace.
WSJ: What are you doing to attract that talent?
MR. ALEXANDER: We have to offer competitive compensation and benefits, and we benchmark ourselves against tech companies and their compensation. Our founder is still the CEO, and there's that ethos that you get with a founder-led organization. It feels much more like a startup. We're also very focused on providing an environment that allows people to learn and build their skills.
WSJ: You oversee 9,000 technology staff. As a bank, why do you need that many technologists?
MR. ALEXANDER: We recognized that customers ultimately are going to choose financial institutions that offer the same kind of experiences they can get in other domains in their digital lives. We need to operate like technology companies that are delivering those experiences in other dimensions of their lives, which means we need to be building our own software solutions, and doing that requires a highly technical workforce. Just like Amazon, Netflix or Uber need to have great engineering technology staffs to build their products, we need to have the same thing for banking.
WSJ: What are some areas of opportunity that you've identified for using machine learning to serve customers better?
MR. ALEXANDER: Fraud is a really interesting problem. The hard part from a technology perspective is you have to make decisions in real time. We've got a customer standing there at point-of-sale and there is this 50-millisecond requirement to return an answer on "do we approve this transaction?"
You've got to have really sophisticated technology to deal with the volume of data and deal with it in real time, and you want to apply complex algorithms to distinguish those transactions that are abnormal and out of pattern and look like fraud. But you've got to tune it well because you don't want to decline a real transaction. We've been investing heavily in this.
Another machine-learning example is a feature we've developed as part of our auto-finance business. You take an image of a car that is either a photo that you take -- or you can hold up your phone with your camera in real time and point it at a car that you're interested in -- and we can automatically in real time tell you what your monthly payment would be on that vehicle.
Ms. Castellanos is a Wall Street Journal reporter in New York. Email her at email@example.com.