The use of a chatbot in financial services is high risk and high reward.

Having a chatbot answer questions a human normally would handle can save the bank money and make it productive. The bank could also win points with customers looking for an easier way to communicate.

But if the nascent technology misfires, it could become a public relations nightmare. Microsoft saw this when it launched a chatbot, Tay, that within a day began spouting racist and offensive comments, such as one saying the Holocaust never happened as its machine-learning engine picked up the language from the people with whom it interacted.

So, it is not surprising Bank of America is taking its time to roll out erica, its high-profile voice and text bot that it demoed at Money2020 last year. The bank is acutely aware that if it’s going to put something in the hands of its 65 million customers (and specifically its 22 million mobile app users), it had better work well from the get-go. The chatbot’s designers are giving erica’s artificial intelligence engine the time to properly learn the nuances of conversing with humans and to be equipped to provide useful, rather than puzzling, off-base or off-putting responses. erica is expected to debut in the fourth quarter.

“In an app world, you get one shot with a client,” said Aditya Bhasin, consumer and wealth management technology executive at Bank of America, during the bank’s inaugural NYC Fintech Summit last week.

Consumers will instantly decide whether erica is cool or useless, he said.

“We're trying to get to that, ‘Hey, that's cool,’ and then as the capability grows, clients can learn to do more with it and see it’s very useful,” Bhasin said.

The bank is also looking to roll out erica with robust functionality, said Hari Gopalkrishnan, client-facing platforms technology executive at Bank of America.

“One thing we want to avoid is getting out there with parlor tricks. You could go out with a launch that answers three questions and the rest of it is all a joke,” Gopalkrishnan said. “It’s got to be a collection of things the customer wants that gets packaged up so they're like, ‘Oh, it's worth talking to this thing, because it doesn't just do three things, it does 46 things.’ ”

It’s worth noting that erica is referred to as “it” rather than “she” because, despite the female-sounding name, the bot is meant to be gender neutral. The name is a play on the last five letters of the Bank of America, hence the lowercase “e.”

The starting goal for erica is to be a digital assistant that helps consumers navigate the bank’s app.

“Whether you use your voice or text, [erica] becomes another interface in the app that allows you to navigate the app better, which hopefully is more client-driven than a series of menus,” Bhasin said.

Learning nuances of language and conversation

A lot of the work going into erica right now is enabling the chatbot to understand the nuances of the English language and the intent and context of a conversation, whether spoken or texted.

For instance, Gopalkrishnan pointed out that people almost never say, “Check my balance.” Instead, they say things like “What have I got?” or “What's left in my account?”

The bank is collecting the many ways customers have asked for their balance in its interactive voice response systems over the years, and feeding them to erica.

Aditya Bhasin, consumer and wealth management technology executive at Bank of America.
First impressions
“In an app world, you get one shot with a client,” says Aditya Bhasin, consumer and wealth management technology executive at Bank of America. Consumers will instantly decide whether the bank’s forthcoming chatbot is cool or useless.

In the hopes of avoiding the fate of Microsoft’s Tay, the bank has built a profanity filter for erica.

“You have to think about all these things as you think about building something that's going to talk back to you,” Gopalkrishnan said. “You have to make sure the customer experience is truly a delighter.”

The digital assistant is being trained to obey commands to do basic tasks, like transfer money from one account to another, pay a credit card or reorder checks.

It’s also being designed to be conversational, to take some back and forth. A customer might ask when a credit card payment is due. erica might reply by asking which card, if the customer has more than one.

Then erica might ask which account the payment should be made from.
Over time, it will learn the customer’s habits and preferences, and just have the customer confirm its choices.

Insights and advice

Besides assistance, erica is also being trained to find insights for customers and make recommendations. If a customer’s FICO score dropped, for instance, erica might suggest better money habits, drawing on a partnership with Khan Academy, a provider of educational tools. The chatbot may identify ways the customers could save more or pay down debt. erica could help customers avoid mistakes like missing a mortgage payment.

Gopalkrishnan said such insights will improve over time, as the chatbot learns more about the customer.

“This is the evolution of a new platform that's a post-mobile experience,” he said. “It starts with conversation but eventually leads into a lot more dynamic insight and being able to engage the customer.”

erica is being further programmed with employees’ ideas, including 15 to 20 that came out of a recent internal innovation challenge. Employees are also testing erica, to teach the chatbot about human interactions.

“We always start with an associate launch,” Gopalkrishnan said. “We start with a small tech team and business team that tests it, then we get it to all our associates, who get to kick the tires. We’ll give ourselves feedback and fix any gaps.”

The bank has also partnered with fintech firms, such as voice recognition and artificial intelligence providers, throughout erica’s development. It would not say which ones.

Alexa, what’s my balance?

In the future, Bank of America may integrate the technology behind erica with mainstream virtual assistants like Alexa, Siri and Google Home. It already has working proofs of concept with those platforms.

“It's not that we don't know if the technology can work — we have it working as a prototype,” Gopalkrishnan said. “It's all the other factors that go into determining do we want to go to market: Is this the right time? Is there customer demand for it?”

Bhasin said the bank is completely agnostic to the input mechanism, whether it’s the bank’s app or a third party.

But such integrations raise critical privacy and data ownership considerations.

“If somebody says into any one of these devices, tell me the balance on the account ending in these four digits, who owns that information?” Bhasin said. “While they're in our app, we're taking the voice, converting it and we can protect and secure it in the way our clients expect. If it's passing through someone else, we have to be a lot more careful about what we're doing with client information. You don't want someone yelling across your kitchen table, ‘Alexa, what's my balance?’ ”

Editor at Large Penny Crosman welcomes feedback at