Hari Gopalkrishnan leads Bank of America’s Consumer, Business and Wealth Management Technology team. On a Tearsheet podcast episode published on 1 April 2025, he told the story of Erica, the bank’s virtual assistant, from the problem that prompted it in 2017 to the bank’s current work with generative AI. Tearsheet’s introduction notes that clients used Erica 676 million times in 2024, bringing the total to 2.5 billion since its launch in 2018. The conversation covered why the bank built Erica itself, how the team was organised, what customer testing changed, how the assistant spread to new groups of users, and how the bank governs AI pilots.
The problem in 2017
Gopalkrishnan said his team noticed something odd as the mobile app gained features. Customers still went to branches and call centres first, before trying to serve themselves. The reason, as he saw it, was the small screen: once an app grows past ten to fifteen key features, finding your way around it becomes hard.
That led to the founding idea. Customers should be able to deal with the bank in the way they choose, instead of the way the bank chooses for them, and a conversational assistant offered a way to do that.
Building it in-house
Buying came first, and it did not work out. He described a volatile vendor market in which start-ups disappeared or were acquired within weeks of the bank starting to evaluate them. Vendors with some natural language understanding lacked knowledge of how financial services concepts fit together. After assessing third parties, the bank concluded that none would do, and adopted well-regarded open-source language engines instead. It hired a team that included people with PhDs in linguistics to work alongside its existing engineers, then worked with business teams to define the experience.
The bank also pushed its agile model further than usual. Regional teams sat with engineering, design and legal colleagues. Designers worked inside each sprint instead of arriving weeks later, and legal staff reviewed issues day to day. He said it was sometimes hard to tell the designers from the engineers, which he counted as the approach’s main strength.
What testing changed
The team had expected customers to split their use roughly evenly between voice and text, and planned to invest equally in both. Testing with a small group of customers overturned that plan.
“people were just typing in text 90% of the time”
Voice was rarely used, and the team changed its priorities before launch.
From customers to employees and business clients
Erica’s scope has since grown beyond consumer banking to several lines of business, to individual and corporate clients, and across the bank’s global operations. For employee use cases, the bank uses pre-trained models that are fit for the purpose, such as ones that can already connect to HR and help-desk systems. He said the bank avoids rebuilding what others have already built and builds its own only where nothing suitable exists. Its experience in language understanding and user design carried across well.
Each new area needed its own decision on whether to build natively or integrate an existing model. Data sources had to be clean and the right APIs had to exist. Many older services had been designed for one specific app, so the bank rebuilt them so that a chatbot could call them, sometimes with several rounds of clarifying questions. That work, he noted, also strengthened core platforms for later use.
Generative AI under a council
He expects generative AI to improve how Erica recognises what a customer wants and to let people speak to it more naturally. The bank had about 25 proofs of concept using large language models in some form, many of them close to production, though he said Bank of America had been quieter about generative AI than most of its peers.
Pilots go through an AI council of senior business, strategy and risk leaders, which decides which experiments to run and why. Each project is checked against a 16-point risk framework covering areas such as bias, intellectual property, transparency and explainability. The council looks at whether human oversight is adequate, how performance will be measured and which guardrails apply. He acknowledged that the process can slow things down and said the bank accepts that cost.
On money, he warned that generative AI is expensive and that many firms announce large spending without a clear view of the return. The bank’s approach is to map how work gets done, by activity, job and task, put a cost on each task, and then measure the return on any solution against those costs.
What it means for leaders
- Customer behaviour should settle design arguments early. One round of testing stopped Bank of America from investing equally in a voice channel its customers barely used.
- The bank built Erica because vendors at the time lacked banking knowledge, and it now buys where suitable models exist. Our reading is that the same test still applies: build only where the market has nothing that fits.
- Putting legal and design inside delivery teams from the start removes the late reviews that stall many AI projects.
- A written risk framework with a fixed set of checks gives a council something consistent to apply, and the delay it adds is a trade-off worth accepting openly.
- Costing tasks before choosing tools is the step that most AI business cases skip.
