Capital One is determined to turn its digital assistant Eno into an affable, personal finance watchdog to build loyalty between the brand and its customers. Two years after Eno’s launch, the bank is adding capabilities that go beyond simple balance inquiries to proactive insights on customers’ spending behavior.

Eno, which is named after “One” spelled backwards, began as an SMS chatbot through which customers could ask questions about their accounts, review transaction histories and pay bills. The product has since evolved to become a personal spending adviser, Ken Dodelin, Capital One’s vice president of conversational AI product development, told Bank Innovation. It’s a process that has involved training Eno — with the help of humans — to watch for irregular transactions. The bank also has moved from SMS to embedding Eno’s capabilities within the Capital One website or app. For now, the bank has no plans to add voice as a channel, but it doesn’t rule out that option in the future.

“The blinking cursor is a harrowing thing for a product leader because customers can say anything and you need an AI to  understand it all,” Dodelin said. “We anchor ourselves to what we would want a human assistant to do — if you had a human who was brilliant and did nothing but focus on your money 24/7, what would you want them to do? You would like them to be able to answer any question you had about your money.”

Capital One sought to embed capabilities beyond reactive inquiries about accounts. From the end of last year onwards, the bank enabled Eno to proactively reach out to customers. Examples include duplicate charges from the same merchant, if a recurring charge all of a sudden increases or if a customer gave an unusually large tip on a meal out.

While Eno gradually is taking on more personal finance advisory duties, Dodelin stressed that the product is still a work in progress. It needs to be trained to deal with the diverse types of inquiries and keyword queries, so humans are being called in to help bring Eno up to speed.

For example, customers have different ways of asking for their balances. Some type the word “balance”; some misspell the word; while others may use abbreviations. Eno needs to be trained to handle a range of situations, along with the escalation of a conversation to a human when the customer is frustrated.

The bank also has added natural-language processing to handle fraud alert questions, during which customers are often asked to confirm or deny transactions. Instead of “confirm” or “deny,” Capital One initially found that 15% of the time, customers offered a personalized reply to the question that the bank could not understand. With the help of natural-language processing technology, the bank was able to increase its comprehension of fraud inquiry responses to 99%.

“We began with hundreds of thousands of live chat transcripts between our human agents and our customers to provide a baseline understanding of how people talk about money [to Eno],” Dodelin said. “We use supervised machine learning.”

To relate to customers effectively, Dodelin said the bank intentionally took a gender-neutral approach to naming the chatbot, and it was careful to not infuse any judgement in its personal finance recommendations to customers. “We brought folks in from the outside to help us with character development,” he added. “Eno is never judgmental.”

Eno is a good example of how banks are experimenting with digital assistants to improve their value propositions to customers. Indeed, Capital One isn’t alone in rolling out these tools; other examples include Bank of America‘s Erica and RBC‘s NOMI.

Dror Oren, chief product officer and co-founder of conversational AI company Kasisto, which works with banks on the development of digital assistants, said banks still are early in their evolution of how they use digital assistants. However, early assessments show these tools have the potential to advance broader business objectives.

“There’s a lot of numbers and ROI that show that there’s real value in it, and banks are starting to realize that,” Oren said. “The way we look at the value that these kind of systems bring, there are three layers: servicing, engagement and acquisition.” While banks may start with the servicing layer, they can add opportunities to promote other products tailored to consumers’ needs.

By focusing on guarding the customer’s financial health through proactive alerts, Capital One is looking to build trust with customers as a basis to offer more tailored product suggestions, according to John Caruso, chief creative officer of digital marketing agency MCD Partners. “If you build trust, you’re more receptive to all the communications the brand is sending you,” he said. “When you trust the brand, it feels less like sales and more like [the bank] presenting you with opportunities as a member.”