You Won’t Find GenAI Under The Hood of BofA’s Erica… Yet
Even as customer interactions with Erica tip towards outbound advisories and queries, at present bank officials hesitate to risk adopting generative and agentic AI.
By Steve Cocheo, Senior Executive Editor at The Financial Brand
Down in the Q&A section of Bank of America’s webpage for its virtual financial assistant, Erica, under the question, "How does Erica work?" you’ll find an interesting sentence:
"It’s important to note that Erica does not use generative AI or large language models, which means responses are not based on vast datasets of external information."
Intriguing in a day when many large financial institutions are falling over themselves to stick a GenAI pin their map.
What’s more, Bank of America’s increasingly popular virtual assistant benefits from the ongoing input of a "brain trust."
That’s literal. Half a dozen people — six human brains — play a constant role in determining how this AI-driven virtual assistant functions, the connections it will make, and the actual answers that it will give or the recommendations it will make when communicating with BofA customers.
Most Americans who aren’t living in caves have at least a rudimentary knowledge of GenAI, following an impressive run-up in awareness over two years.
"Last year, GenAI was all the rage. This year agentic AI is all the rage," says Jorge Camargo, managing director, mobile app, online banking and Erica AI executive. Erica, now about eight years old, was built on natural language processing and machine learning, with the human element a key and ongoing part. It accepts customer queries by voice or as text.
GenAI and agentic AI could someday become part of Erica, and are under evaluation for many roles at BofA. (At this point, Camargo says, there will always be a person between GenAI and the public.)
But for right now, if you look under Erica’s hood, you won’t find GenAI nor agentic AI.
"We have a team of data scientists that is constantly looking at how Erica is behaving, how it’s interacting with clients, how it’s responding to questions — and working to improve everything," says Camargo. A key reason for this is to ensure that Erica’s answers conform to BofA policies and practices, but there’s more to it. He adds that it’s essential to engendering trustworthy answers.
The bank has approximately 40 million mobile-active customers, and just under 50% of them — 19.7 million — use Erica to some degree, from occasionally to dozens of times a month. All told, the bank says, they interacted with the virtual assistant 676 million times in 2024. This reflects strong adoption in eight years, from about 5% in the early days. The bank has recently expanded availability to certain business customers.
Read more: How Bank of America Unified Five Apps Into One Experience
The Value of Human-AI Partnership
Now, here are some key reasons why BofA isn’t rushing towards the new AI technologies for Erica.
When Erica started out, it served as an engine for answering customer questions, growing in sophistication as it went forward. Today, that covers about 40% of the product’s interactions with enrolled BofA digital users, chiefly consumers.
The other 60% of interactions are outbound messages from Erica to the user, preemptive or predictive messages, queries and insights that are part of the bank having the customer’s back.
An example of the latter from Camargo’s own life: Using pattern recognition, Erica discovered a "coincidence" — a bunch of $4.99 charges were coming through. Chuckling, the banker says this is how he discovered that his kids had figured out how to buy apps on their iPads.
Camargo says that as digital staffers developed and refined Erica, they found that the range of topics and intents that most clients will ask about encompass only about 700 topics.
"So we don’t have an unlimited question-and-answer space as do some of the generative AI platforms that need to work with any potential question that might come up," says Camargo. Everything customers will see from Erica has been past the eyes and through the heads of the half dozen data scientists, working with automated tools to flag trends but applying human thinking.
Camargo labels this "supervised machine learning," a term of art in AI circles.
Case in point: Readying Erica to field customer questions where they’ve been caught up in natural disasters.
Camargo says that as Covid became a pandemic, the bank developed customer assistance program guidelines and benefits. The Erica team worked to get this timely information into the virtual assistant’s routines.
Subsequently, team members realized that much of what they learned could be applied in other troublesome situations.
"So, with the recent storms that hit North Carolina and the wildfires in Los Angeles, they were quickly able to associate search terms the public was using — we call these ‘utterances’ — and tie them back to existing topics," says Camargo. In this way, he says, Erica could be updated quickly to handle the latest problems.
Camargo believes the human input is a key differentiator for Erica and that continual investment in the "tuning function" is critical.
"Our team has made over 50,000 improvements in the background of Erica since it launched," he says. This can be as simple as working out "disambiguation" — a standard term in Wikipedia entries — meaning, in Erica’s context, working out routing where a word or term can be used more than one way in a banking conversation.
A key point about Erica that also relates to a different facet of the human element is what Camargo refers to as the "off-ramp." He explains that the idea is to not have dead ends — a common failing of even modern interactive phone response systems — but to always give the customer a way to segue to a call routed to a human banker or even to scheduling a live appointment or video chat session with one.
Read more:
Proactive Words from Erica to Newer Forms of AI?
An increasingly seen trend in usage for inbound queries from Erica users is directional. Rather than using menus and other manual paths in the bank’s unified mobile app, many people are asking Erica to bring them to the appropriate place to perform a given tasks. But, as noted earlier, a growing portion of interactions are outbound, from Erica to the user.
These arise from continuous monitoring in the background of all of a customer’s relationships with the bank. Flagging issues like growing levels of subscription spending is one example of Erica’s "awareness" of trends.
As these kinds of advisories grow in number, it begins to point to where expansion of AI capabilities in Erica could lead, says Camargo.
"When you do almost 700 million interactions annually with nearly 20 million clients, you want to make sure those answers are correct and consistent," says Camargo. Right now that’s not something GenAI can be counted on delivering, he explains.
He’s experimented with ChatGPT and the like personally and finds the technology fascinating. However, too often the results are definitively wrong, he says. Further, originality and innovation aren’t there.
"The technology obviously holds a lot of promise and we are absolutely looking at it [for Erica] and seeing where it’s going to make the most sense from the perspective of client experience," says Camargo. "But there can be no hallucinations, no kinds of errors that would be impactful to our clients."
Beyond flagging matters for attention, the next steps could be asking for authorization to take the next step and automatically fix things, such as disputing a suspicious charge with a merchant or, having spotted a fraudulent charge, locking down the account pending further investigation.
"Very exciting stuff," says Camargo. "But it’s a nascent field."
Camargo notes that GenAI could help BofA tackle a challenge that it’s faced since Erica debuted: Erica only communicates in English.
"Something we are researching for Erica is how we can leverage GenAI to provide multilingual capabilities," says Camargo. "Nothing firm to announce just yet, but it’s definitely one of the areas of research and focus right now."
