Capital One: What we learned about what our chatbot learned

Capital One: What we learned about what our chatbot learned

Ken Dodelin, vice president and head of conversational AI products at Capital One reflects on the learnings from a six-month pilot trial of the company's new conversational chatbot Eno, as the interface is rolled out commercially to all customers.

In March at SXSW 2017, Capital One launched an invite-only pilot of Eno, the first natural language SMS chatbot from a U.S. bank. Today, we’re thrilled to announce that Eno has emerged from the pilot and is now available to text with millions of Capital One U.S. credit card and bank customers.

So, what has Eno learned during the past seven months? Well, three things, mostly…

1. Eno learned about the diversity of how people text about their money
To date, more than 100,000 invite-only pilot customers have chatted with Eno about their credit card and bank accounts. They text at all times of day and in all kinds of ways. Some are very to the point and all about abbreviations (“bal”). Others are chatty and speak in natural language (”Hi Eno! How much money do I have today?”). Some fumble every other word (“val” is a common one since “b” is adjacent to “v” on the keyboard), while others prefer emojis. In fact, we’ve seen customers use a thumbs up emoji to confirm their payment more than 50% of the time, and Eno understands.

It’s great to see our customers chat about their money in and on their own terms! Some of the more creative texts we’ve seen include:

• I’ll talk to you later. Make sure you’re keeping my money safe!!

• Hey Eno. It’s fantastic to meet another sentient. I will enjoy your help and your company. Can I get my bal?

• Wow, I’m impressed with your abilities Eno! I am glad to have you in my contacts!

2. Eno learned to understand what customers mean
Eno learns from every conversation through supervised machine learning. Full disclosure: Eno is a lifetime learner and can’t yet do everything that customers throw at it. While we’ve expanded some capabilities during the pilot, our main focus has been in making Eno great at the basics: high-frequency use cases like balance, available credit, recent transactions and paying a credit card bill where the conversational interface performs particularly well (…oh, and the fastest way to find that elusive routing number around.) For example, we’ve trained Eno on more than 2,200 different ways that customers have asked for their balance. Thanks to stellar work by our amazing Conversational AI Design and Tech teams, customers increasingly find that Eno understands them and they are delighted when Eno does. This is a significant step away from banks training customers to speak in monolithic “bankspeak”, and a meaningful contributor to the positive feedback we’ve received on Eno from customers to date.

3. Eno began learning about building relationships through conversations
This one is admittedly an odd claim. The Eno pilot has provided a fascinating view into the human-AI interaction model, and it’s decidedly different from our traditional ways of interacting with customers. One insight that speaks to this: the 3rd most frequent intent (an intent is something like “account balance” or “recent transactions” derived from a customer’s text) that Eno receives from customers is “Thank you.” There is no purely functional reason to thank Eno, yet the conversational interface and personal interaction allow people to make some form of emotional connection. Today, it’s over basic Q&A conversations, but we see plenty of opportunity to grow trusted and valued relationships between Eno and our customers through more in-depth and personal conversations over time.

We come at this space with a great deal of humility and, like Eno, we continue to learn every day. Today begins our next step in Eno’s journey, and we’re excited about bringing Eno to millions of Capital One bank and US card customers.

Comments: (3)

A Finextra member
A Finextra member 25 October, 2017, 10:33Be the first to give this comment the thumbs up 0 likes

I wonder how ENO responds if one types something like "Eno, please transfer the bank's cash reserves to the following numbered account in the Gugu islands"?

A Finextra member
A Finextra member 25 October, 2017, 14:35Be the first to give this comment the thumbs up 0 likes Amazing results! Congrats! Thank you for sharing.
Ketharaman Swaminathan
Ketharaman Swaminathan - GTM360 Marketing Solutions - Pune 25 October, 2017, 19:04Be the first to give this comment the thumbs up 0 likes

If everyone that deploys chatbots takes the trouble that CapitalOne has taken to learn what their chatbots have learned, chatbots will replace humans even faster than I'd imagined in Can Chatbots Replace Humans?

Coincidentally, on my adjacent browser tab, I have a press release from Persado, the leading provider of AI-generated marketing language for top global brands. Persado found that there were 250 ways of writing "Call now". That's beaten hollow by Eno's learning that there 2200 ways that customers have asked for balances!