How Bank of America’s Erica does the work of 11,000 people

Bank of America’s virtual assistant, Erica, is doing the work of 11,000 people, company leaders said during a recent investor day.

Processing Content

“We’ve been at the game with Erica for seven years, now with 20 million regular users and two million interactions with Erica a day,” said Holly O’Neill, president of the bank’s consumer, retail and preferred lines of business. The goal is to have 80% of the bank’s clients actively engaging with Erica, she said.

Erica is currently used by 42 million consumers and 40,000 business customers, who use a version built into the bank’s CashPro digital banking platform. Erica for Employees is used by 95% of the bank’s 213,000 employees to answer IT and HR questions. Call-center representatives use Erica Assist to help answer customer queries, and Merrill Lynch financial advisors a version of Erica called Ask Merrill to help prepare for client calls.

In recent interviews with American Banker, seven Bank of America executives provided a closer look at Erica and how it’s being used by the bank’s customers and employees. The interviews and demos showed how the bank’s approach to virtual assistants has diverged from the strategies of other banks, and where it’s achieving results.

With its pervasive use of home-grown Erica outside and inside the company, Bank of America is zigging where others are zagging.

Large-bank rivals like JPMorganChase and Morgan Stanley have been providing generative AI models to help employees with their work. Erica, by contrast, relies on more traditional forms of AI, like natural language understanding and predictive analytics. Erica draws upon hundreds of carefully curated answers to questions.

Using a large language model like ChatGPT to answer customer questions would be like using an elephant gun to capture a mouse, according to Hari Gopalkrishnan, chief technology and information officer at Bank of America, who leads a team of 60,000 technologists.

“I could use it, but it’d be overkill,” Gopalkrishnan told American Banker. “We’re making sure we’re using the right tools for the right purpose.”

Large language models are geared toward generating large tracts of content, like research papers, rather than clear, specific answers, he said.

“We intentionally did not go large, because we would have spent a ton of money on training and using inferences when the reality is the problem statement isn’t about a large language model,” Gopalkrishnan said. “The problem statement is: Understand short bursts of text — no one’s going to type in an essay on that chatbot screen —map it to a set of clear intents and go execute them.”

(“Intents” are Bank of America’s term for curated, rule‑based answers or workflows that tell Erica what to do when a client asks a specific question.)

Erica has been programmed to map customer requests to intents, which trigger a cascade of actions.

If a customer types, “I want to send Jorge $20 from my checking account for the lunch we had last week,” Erica recognizes the customer is most likely looking to make a Zelle payment from a checking account, finds Jorge in the user’s contact list, pulls the contact information and the amount of $20, puts “Lunch last week” in the memo field and presents the results to the user to confirm the details.

Another reason the bank’s tech leaders choose not to use LLMs as virtual assistants is for accuracy’s sake.

“When I go to ChatGPT and I’m looking to aggregate the best soccer programs for my kid in town, precision and accuracy matter a lot less,” Nikki Katz, head of digital at Bank of America, told American Banker. In banking, she said, “there’s no room for error. So it’s not as free-form as what people are used to. But that’s by design, because of the application.”

And even as the most popular LLMs have improved and are hitting 98% non-hallucination rates, that’s not high enough for a bank, Gopalkrishnan said.

“How often can we be wrong?” he said. “Can I be wrong 2% in telling a customer something? The answer is no.”

U.S. banks that do use large language models as virtual assistants don’t offer them to customers yet, due to this concern about the risk of errors and hallucinations.

Bank of America, like other banks, does use large language models internally for other purposes, such as generating client meeting prep documents.

The bank’s strategy of sticking with traditional AI rather than turning to LLMs for Erica makes sense because of all the data integration work Bank of America has done, according to Emmett Higdon, digital banking director at Javelin Strategy and Research.

“They probably spent hundreds of millions of dollars on the data integration behind Erica,” he said. “With all of that data being structured in a correct way for Erica to access it, that’s not throwaway work.”

Still, the bank could eventually upgrade Erica with a large language model, he said. “The investment they’ve done over the past 10 years really puts them in a great place to now start tapping those same data sources, but with a more modern model.”

Steven Ramirez, CEO of the consultancy Beyond the Arc, also sees merit in using more traditional AI rather than LLMs for this purpose.

“For a system like this that is built on a highly explainable infrastructure, the types of models that they are using for this kind of analysis can be validated,” Ramirez told American Banker. “You can track drift over time. You can understand aspects of bias. You can have a great degree of repeatability — almost assured repeatability.”

Erica is a stable, mature and well-governed system, he said.

By contrast, using a large language model, if you “ask the same question to an AI bot five times, you get five different answers,” Ramirez said.

Bank of America also benefits from having billions of interactions in its collective database, he pointed out.

“Once you have that kind of rich experience and data and it all works, it’s hard to give that up for something that is inherently untested,” Ramirez said. “I’m a huge advocate of what’s possible with LLMs. But, if you really have a very stable system that’s doing the job, you could see the hesitancy to upset the apple cart.”

LLMs do have an advantage in that they are able to make inferences about situations that they have not seen before, he said.

“A more rigid model is not able to be as flexible, particularly about situations that it’s never seen before,” Ramirez said. “In the dynamic environment that is banking and consumers and small businesses, you can see why you might not want to just present answers that are largely rule-based.”

Creating Erica

Ten years ago, Bank of America’s tech leaders saw customers’ use of its mobile banking app start to plateau.

“Even though the mobile app was awesome, it was popular and people were using it, people were still calling the call center a lot, and people were still coming into the financial centers,” Gopalkrishnan said. “And we’d ask, why did you just call us to order a checkbook? Why did you just call us to look at what this transaction was about? Why did you call us to wire funds? You could have done this yourself. And the typical answer was, ‘Well, it’s so hard to find stuff on a five-inch screen when it’s buried in three levels of navigation.'”

So the first goal was to make it easier for people to find features of the app by letting them ask questions in plain English, rather than forcing them to understand the bank app’s menu structure.

Another realization that Bank of America’s tech leaders had at that time was that it would be helpful to provide customers with insights into what was happening with their money.

Ten years ago, there was no commercially available software that could do these things, Gopalkrishnan said.

“Big vendors were touting products that would do all this stuff, and it’d be amazing, and they could deliver in a year,” Gopalkrishnan said. “And it was clear after the first two conversations, they couldn’t.”

At that time, none of the vendors could support giving customers the ability to communicate with the bank by talking, tapping or texting, either, he said.

Bank of America’s tech team started with open-source natural language processing engines, such as one developed at Stanford University, that were not specific to financial information.

The bank hired PhDs in linguistics to work with software developers and designers. Legal, risk and compliance people were added to the team “because you don’t want to get this wrong,” Gopalkrishnan said.

Erica combines licensed, open-source and internally developed solutions, including natural language processing, to understand client requests; machine learning to analyze and match more than 700 intents; and predictive analytics. It’s been updated 75,000 times.

“We’re constantly tuning and updating,” Katz said. “We keep also looking at conversational design. What are the millions of ways you might ask the same question? What are the next steps? How do we work this into the next intent, or the next question you might ask, and what are we okay with Erica doing next? What recommendations can we make? Over time, it becomes more and more personalized.”

Erica recently told Gopalkrishnan that his account was charged three times for his son’s gym membership, something he would never have noticed otherwise until the end of the month.

Gopalkrishnan declined to say how much the bank spends on Erica. Bank of America’s overall tech budget is $13 billion a year; $4 billion of that spending is on new technology, of which Erica is considered a part.

How consumers use Erica

Roughly 40% of the time, a customer starts using Erica by typing a question into the prompt field, gets presented a list of commonly asked questions to choose from and picks the one that seems most relevant. If none of the questions seem like a fit, Erica suggests related topics or offers the option to connect with a human associate via live chat, phone or an in-person appointment.

The most common questions consumers ask are about account information (recent transactions, account and routing numbers, credit scores, viewing statements), money movement (ACH transfers, bill payments, Zelle, wire transfers), card management (activating a card, locking/unlocking a card, replacing a card), recurring charges and subscriptions and navigation (help me find…).

The other 60% of the time, customers go to the Erica home page, which offers insights and notifications. Erica might tell a user that her spending for the month has increased, for instance.

Erica can also alert people to changes in their recurring payments.

“I didn’t read my email from Netflix, so Erica told me my subscription was going up,” Katz explained. “It’s the information you need in the place you need it.”

A summary of subscriptions and other recurring charges is pushed out to users monthly. (Customers can’t cancel subscriptions directly from Erica. The bot provides information about the merchant, and the user contacts that company directly.)

Erica can also tell jokes. One is, “Why did the student eat his homework? Because the teacher told him it was a piece of cake!” Another is, “What kind of tree fits in your hand? A palm tree.”

When a customer needs to talk to a human, Erica can direct the person to the right associate for a live chat or phone conversation. The authentication process the customer went through on Erica is valid for all other channels, so the person doesn’t have to re-authenticate themselves.

The transcript of the conversation the customer had with Erica gets forwarded to the human agent, “so by the time it gets to the other side, the person can actually pick up and say, ‘Hey, Penny, looks like you were having problems canceling this transaction with a merchant. Let me tell you how you can best do that,'” Gopalkrishnan said.

On the consumer side, the next steps for Erica include more personalization, according to Katz.

“You’re going to see Erica continue to get smarter about operating across all of your relationships with the bank,” Katz said. “You’re going to see Erica continue to get smarter in terms of more personalized insights. You’re going to see Erica get smarter in terms of when’s the right moment to tell you things. And then the tuning and the work we’re doing, based on the data that we have, Erica is going to continue to feel more and more intuitive and more and more integrated into financial lives.”

How corporate clients use it

In 2023, Bank of America embedded a version of Erica into the CashPro online banking system that corporate customers use to initiate and approve payments, send wires, reset passwords, retrieve documents and manage their accounts. The first thing that CashPro Chat does — in response to a request from the bank’s lawyers — is disclose it’s not a live person.

The service is used by 65% of the bank’s corporate clients, and it handles more than 40% of these clients’ interactions, the bank says.

“CashPro has a lot of bells and whistles, and clients may not know exactly where to go,” said Abbey Novak, CashPro service product executive in global payments solutions at Bank of America. “Maybe they haven’t had time to do the training that comes along with it.”

The “containment rate,” or percentage of inquiries that can be answered by CashPro Chat without going to a live human being, is about 43%, Novak said. “The goal there is to continue to improve that and chip away at that rate as you expand the knowledge of CashPro Chat and be less reliant on the live service team to swoop in and answer things.”

All of this saves corporate customers time, according to Tom Durkin, global product head for GPS CashPro.

“Treasury teams today are not hiring,” he said. “We’ve heard that consistently from our CashPro customers. So finding ways that drive convenience, ease of use — those things resonate when they’re rushing through their day.”

Erica for Employees

All Bank of America employees have a version of Erica on their PC or laptop, in the application tray at the bottom of their screens. It can also be accessed through an internal portal.

The tool pops up with the question, “What would you like to do?” and a prompt box. A staff member might type in, “Set up WiFi.” Erica confirms that the person is asking to set up WiFi, asks several follow-up questions to pinpoint what the person needs, then provides instructions and ends with, “Is there anything else you need?”

This process avoids a call to tech support, according to Richard Knafelz, managing director and employee experience technology executive at Bank of America.

The process of retrofitting Erica to provide IT and HR help to the bank’s employees started during the pandemic.

“Everybody was obviously at home, and there were a lot of calls to the help desk for certain things, like setting up a video collaboration tool like WebEx,” Knafelz told American Banker. “We started to automate functions that we were getting calls for, and we used the help desk and then inputs and outputs of Erica for employees to learn what we needed to do next.” His team has created about 614 Erica workflows, mostly for the help desk.

Knafelz and his team are creating proofs of concept of using agentic AI to round out the questions that Erica can answer. They’re also thinking a lot about how to put safeguards in place that verify an answer is correct before an employee acts on it, Knafelz said.

“So, no hallucinations here,” he said.

Employees’ most common questions for Erica are about accessing their corporate devices, dealing with authentication challenges and connecting to Wi-Fi and virtual meetings.

“It knows what devices you have, and it asks you which one you’re locked out of, and then, based on that, it’ll remotely unlock for you,” Knafelz said. “It knows the employee that it’s dealing with, so it doesn’t have to ask all kinds of upfront questions. It understands the context and who you are, and it can take it from there. It’s doing some pretty complex integration behind the scenes, calling to the different systems to understand what devices you have, and then doing the unlock.”

About 97% of employees have used Erica for employees at least once; 94% are repeat users. The number of calls to the help desk has dropped 55%, which the bank attributes directly to the bot.

Ask Merrill

Merill Lynch representatives use a version of Erica called Ask Merrill, and private bankers have a similar version called Ask Private Bank.

Ask Merrill is used by about 23,000 representatives a month. It has a library of about 10,000 unique answers generated by content creators and from bank policies and procedures. The answers are kept up to date through an annual refresh procedure and through more frequent updates, according to Salvatore Schiano, managing director and head of wealth management client care and fraud services at Bank of America.

Ask Merrill ties into other applications across the enterprise. So for instance, if a Merrill representative wants to know what happened to a client’s wire transfer, they can ask that question with the account number, and Ask Merrill will pull the status of that transaction from the right place.

Schiano’s team is testing the use of generative AI to create answers. For instance, processing a deceased person’s account can get very complicated, he said. “We have a published six-page PowerPoint that the user would have to potentially read through,” Schiano said. “We’re letting AI generate an answer based off of the questions.” There’s still going to be a human in the loop reviewing and approving the content.

The gen AI model (Schiano declined to say which one) will save the users time by providing a short summary rather than a long PowerPoint or text document.

Call volumes at the contact center that supports this user base have dropped about 38% over the last six years, Schiano said. And usage of Ask Merrill has gone up 39% over that same period. “You can see that more users are trusting the application, using it, and generating less calls into the call center to gain the information,” he said.

Erica Assist

In Bank of America’s call centers, customer-service representatives use Erica Assist to provide more personalized and efficient support, according to Ashley Ross, head of client experience.

The first thing Erica Assist does is pop up and tell the agent the reason the customer is calling — for instance, to ask about a declined card transaction. It also pulls information from several different systems into one desktop tool to help guide the conversation with the client.

And it helps with loyalty recognition, for example, guiding an agent to acknowledge how many years the customer has been with the bank and thanking them for their loyalty.

Erica might note that the customer has not enrolled in digital banking and offer to help them sign up. If the customer has a merchant dispute, It will provide procedures for dealing with the matter. It shows past interactions with the customer. It displays any fraud alerts the customer has received.

Gopalkrishnan’s goal for Erica is not to have the most advanced technology, but to provide the best combination of accurate answers and a shortcut to talking with a human when needed.

“There are times when a client just needs help,” he said. “They don’t want to have the bot anymore. The seamlessness with which you transition from a high tech to a high touch and get your problem resolved, and you’re in and out the door — those are all things that we focus a lot of customer experience efforts on, beyond just sort of saying, what’s the new shiny model that’s out there?”