Blab Business
Bank of America Prioritizes AI Accuracy Over Efficiency, CEO Brian Moynihan Says
Bank of America’s chief executive, Brian Moynihan, has made it clear that the firm’s AI strategy is built around precision rather than speed. In a recent interview, Moynihan explained that the bank’s virtual assistant, Erica, must deliver flawless answers to customer queries, or the consequences could be catastrophic.
Erica is a small language model that has been integrated into more than 110 of the bank’s internal systems. It is designed to respond accurately to 700 distinct customer intents, ranging from balance inquiries to transaction details. According to the bank, 20 million customers use Erica, which is invoked about 200 million times each quarter.
Moynihan emphasized that the tolerance for error in a banking context is far lower than in other industries. "When a customer types a question into the bank’s AI assistant, the infrastructure to get that inquiry to us and back has to be instantaneous, or you’re mad at us," he said. "If you ask about a check transaction of $20 that month and we give you the wrong amount, that’s a problem. Accuracy is critical."
The focus on accuracy contrasts with the prevailing narrative among banking leaders, who often highlight AI’s ability to cut costs and boost productivity. Theodora Lau, co‑founder of Unconventional Ventures, noted that Moynihan’s stance “stands out because most CEOs—at banks and other companies—are still talking about AI from the perspective of operational efficiency, productivity gains and cost cutting.”
Moynihan’s comments were echoed by Máté Jendrolovics, CEO of Intuitech, who said in a podcast that an 80 percent accuracy rate would be “zero value for a bank.” The point is that banking decisions hinge on exact figures; a small margin of error can lead to regulatory violations, customer mistrust, or financial loss.
To achieve this level of precision, Bank of America partnered with researchers at Stanford to develop Erica. The model is deliberately constrained so that it can only provide answers that the bank has verified as correct. When a customer asks for a balance, for example, the assistant pulls the data from the bank’s own records rather than generating a response from a broad knowledge base.
Human oversight remains a key part of the system. Moynihan explained that for any query that requires real judgment, a human is kept in the loop. "We keep it constrained enough to give the right answers," he said. "That’s why we have a human in the loop for real judgment."
The bank’s emphasis on accuracy also reflects a broader industry concern about the risks of generative and agentic AI. While some institutions measure success by the number of calls deflected or the speed of response, Bank of America’s metrics focus on the correctness of the information delivered. The firm’s approach is designed to reduce complaints, lower compliance costs, and maintain customer trust.
Looking ahead, Bank of America is expected to invest further in data infrastructure and model verification. The bank’s strategy signals that other financial institutions may follow suit, especially as regulators increasingly scrutinize AI systems that can impact financial decisions.
At present, Erica continues to serve 20 million customers and handle 200 million interactions a quarter. The bank has not announced a timetable for expanding the system’s capabilities beyond the 700 intents it currently covers. However, the emphasis on accuracy sets a benchmark for how large banks can deploy AI responsibly while safeguarding customers and the broader financial system.
The next few months will likely see additional disclosures from Bank of America regarding its AI governance framework, potential regulatory filings, and any planned upgrades to Erica’s underlying models. Investors and regulators will be watching closely to see whether the bank’s accuracy‑first approach translates into measurable risk mitigation and operational resilience.
Erica is a small language model that has been integrated into more than 110 of the bank’s internal systems. It is designed to respond accurately to 700 distinct customer intents, ranging from balance inquiries to transaction details. According to the bank, 20 million customers use Erica, which is invoked about 200 million times each quarter.
Moynihan emphasized that the tolerance for error in a banking context is far lower than in other industries. "When a customer types a question into the bank’s AI assistant, the infrastructure to get that inquiry to us and back has to be instantaneous, or you’re mad at us," he said. "If you ask about a check transaction of $20 that month and we give you the wrong amount, that’s a problem. Accuracy is critical."
The focus on accuracy contrasts with the prevailing narrative among banking leaders, who often highlight AI’s ability to cut costs and boost productivity. Theodora Lau, co‑founder of Unconventional Ventures, noted that Moynihan’s stance “stands out because most CEOs—at banks and other companies—are still talking about AI from the perspective of operational efficiency, productivity gains and cost cutting.”
Moynihan’s comments were echoed by Máté Jendrolovics, CEO of Intuitech, who said in a podcast that an 80 percent accuracy rate would be “zero value for a bank.” The point is that banking decisions hinge on exact figures; a small margin of error can lead to regulatory violations, customer mistrust, or financial loss.
To achieve this level of precision, Bank of America partnered with researchers at Stanford to develop Erica. The model is deliberately constrained so that it can only provide answers that the bank has verified as correct. When a customer asks for a balance, for example, the assistant pulls the data from the bank’s own records rather than generating a response from a broad knowledge base.
Human oversight remains a key part of the system. Moynihan explained that for any query that requires real judgment, a human is kept in the loop. "We keep it constrained enough to give the right answers," he said. "That’s why we have a human in the loop for real judgment."
The bank’s emphasis on accuracy also reflects a broader industry concern about the risks of generative and agentic AI. While some institutions measure success by the number of calls deflected or the speed of response, Bank of America’s metrics focus on the correctness of the information delivered. The firm’s approach is designed to reduce complaints, lower compliance costs, and maintain customer trust.
Looking ahead, Bank of America is expected to invest further in data infrastructure and model verification. The bank’s strategy signals that other financial institutions may follow suit, especially as regulators increasingly scrutinize AI systems that can impact financial decisions.
At present, Erica continues to serve 20 million customers and handle 200 million interactions a quarter. The bank has not announced a timetable for expanding the system’s capabilities beyond the 700 intents it currently covers. However, the emphasis on accuracy sets a benchmark for how large banks can deploy AI responsibly while safeguarding customers and the broader financial system.
The next few months will likely see additional disclosures from Bank of America regarding its AI governance framework, potential regulatory filings, and any planned upgrades to Erica’s underlying models. Investors and regulators will be watching closely to see whether the bank’s accuracy‑first approach translates into measurable risk mitigation and operational resilience.
