Inside Bank of America’s ‘build once’ AI strategy

BofA went on to tap the underlying architecture of Erica for various use cases across the business, including both for customers and employees. Now seven years—and an uncountable number of AI breakthroughs—later, Bank of America is still leveraging that initial AI build-out. It’s all part of its “build once, reuse” AI strategy. 

This doesn’t mean the company is operating with decades-old AI; the platform underlying Erica was designed to be adaptable to future AI models and breakthroughs. At the same time, there was no denying that the generative AI explosion was an inflection point. Bank of America experimented with other approaches and offerings, but ultimately decided to use what it had, double down, and yet again build once for the years to come. In 2025, BofA reinvested in the Erica architecture and built a second generation platform for the new AI era.

“We’ve learned, in a way the hard way, and over time, that by investing in a foundation and these capabilities, the third and fourth and fifth build is much faster,” said Hari Gopalkrishnan, chief technology and information officer at Bank of America. 

The evolution of Erica

Erica was originally built to meet customers’ desire for an easier way to navigate the bank’s applications, according to Gopalkrishnan, who worked on it as CIO of BofA’s consumer banking division, which was his role at the time. When word started getting out that other areas of the business had the same need, especially among the bank’s own associates, such as financial advisors, BofA sought to reuse what it had built with Erica. 

In 2020, Bank of America launched Erica for Employees, a virtual assistant for basic tech support, and Ask Merrill, a tool designed to help the bank’s advisors efficiently curate information for clients. Around this time, Gopalkrishnan said, the bank decided to lean into reusing the Erica architecture as a specific strategy. More features and tools built on the Erica architecture followed, such as CashPro Chat for clients, Capital Markets Insights, and a tool called Ask Private Banking.

The key to making this work was that Erica was a platform designed to be model-agnostic. The build started with open-source natural language models like BERT and OpenNLP, but allowed the bank to bring in new models better suited for particular tasks.

“We always knew we were going to replace models as we go along,” Gopalkrishnan said. “So it wasn’t that Erica is a model. Erica is a platform that can actually leverage multiple models.” 

He pointed to how, with foundation models, you typically have to “pave the path” each time. You have to set up a vector database, connect to it, and so on. The bank’s platform, on the other hand, enables engineers to think about things like performance, scalability, security, and guardrails in a way that’s shared and leveraged.

“It lets us, in some ways, get to market faster, because now each team doesn’t have to build the pipes and plumbing themselves,” he said.

Gen AI puts ‘build once’ to the test

Bank of America’s “build once” strategy took it this far, but would it hold up amid the recent explosion of more advanced AI? 

The proliferation of large language models sparked a reassessment. Along with the bank’s head of strategy, Gopalkrishnan started an AI council to find out, bringing all the business divisions together and asking them about their needs. They delivered around 15 proofs of concept to try to solve some of those problems, specifically letting people do their own thing—and not use the Erica platform—to see how it turned out.  

“As the results came back, it was clear we were reinventing the wheel a lot,” he said.

This cemented the decision to reinvest in the Erica architecture; once again build a singular platform to provide the core capabilities; and then build use cases on top of that. While AI has come a long way, today’s leading generative and large language models use the same Transformer architecture that underpinned the open-source technologies the bank originally used for Erica. So Gopalkrishnan and team took those technical underpinnings and many lessons from the first generation of Erica, and in 2025 built an updated iteration for the new AI era. 

“We just figured out what the architecture was to assemble them and repackage them in a way that’s appropriate for the new models,” he said.

Discipline and patience required

Bank of America’s approach isn’t always a smooth process, however. The bank has 60,000 people working on its technologies, according to Gopalkrishnan, and sometimes people want to go rogue. 

“You’re basically creating a funnel through which all the ideas have to bloom through,” he said. “The 10,000 ideas people come up with—everybody thinks they can just whip it up tomorrow, and off to the races they go because some vendor will pitch them something.”

He said there’s always an ongoing tension in the system that the more you take a “build once” strategy, the more you run the risk of slowing the organization down. But he believes it takes discipline to not just go with the fastest route, because that can come back to haunt you.

“The biggest challenge,” Gopalkrishnan said, “is to have institutional patience and institutional support to recognize the value you get by slowing down a bit upfront accelerates you a lot going forward.”