Bank of America's Erica virtual assistant serves 42 million customers with 2 million daily interactions

Bank of America faced a digital engagement challenge common to large retail banking institutions: despite strong mobile app adoption, routine customer interactions were not shifting away from expensive call center channels. Tasks like ordering checkbooks, reviewing transaction history, and initiating wire transfers required navigating menus three levels deep on small mobile screens, creating friction and driving unnecessary call volume. No commercially available vendor offered the combination of multimodal input — voice, tap, and text — and deep integration with Bank of America's core banking systems the project required. The status quo meant sustained high call center costs and a widening gap between digital capability and actual customer behavior.

Unable to source a vendor meeting its requirements, Bank of America built Erica entirely in-house over a decade. The system combines open-source NLP engines with machine learning trained to recognize more than 700 curated customer intents. Rather than relying on large language models, Erica maps natural-language queries directly to specific actions — parsing intent, retrieving account data, and executing transactions in real time against core banking platforms. It handles card management, money movement, and account inquiries across voice, typed, and touch inputs. The system has undergone more than 75,000 updates since launch, adding proactive capabilities that flag subscription price increases and duplicate charges before customers ask. A parallel deployment, Erica for Employees, extends the same NLP infrastructure to IT and HR support for the bank's 213,000-person workforce.

Erica's scale is captured in a single benchmark: the system performs the equivalent work of 11,000 people. Today, 42 million consumers and 40,000 business customers use Erica, with 20 million regular users generating 2 million interactions per day. The bank targets 80% of all clients actively engaging with the assistant. Internally, the employee-facing version has achieved 95% adoption across the bank's workforce, handling IT and HR queries alongside external customer support — demonstrating that the same NLP platform can scale across fundamentally different use cases without being rebuilt.

• 42M total consumer users
• 2M daily interactions
• 20M regular users
• 95% employee adoption (213,000 staff)

• Traditional NLP with narrow intent-mapping can outperform generative AI for high-stakes, narrow-domain tasks where hallucination tolerance is near zero — accuracy outweighs flexibility.
• A decade of proprietary data integration creates a competitive moat that off-the-shelf vendors cannot replicate quickly.
• Proactive push-style alerts — subscription changes, duplicate charges — drive sustained engagement beyond reactive Q&A.
• Extending a proven AI platform to employee-facing workflows (IT, HR) multiplies ROI without rebuilding from scratch.
• Multimodal input (voice, text, touch) is a prerequisite for broad adoption across diverse customer segments.