Wells Fargo’s AI assistant just crossed 245 million interactions – no human handoffs, no sensitive data exposed
1 min read
Summary
US bank Wells Fargo has seen its Google LLM-driven assistant double its original projections, without exposing sensitive customer data to a language model, according to a Venturebeat article.
The bank uses a privacy-first pipeline, with a customer’s voice interactions being transcribed locally, then scrubbed and tokenised by Wells Fargo’s internal systems, including a small language model for personally identifiable information detection.
With the data then forwarded to Google’s Flash 2.0 model, only the user’s intent and relevant entities are identified, meaning no sensitive data ever reaches the model.
Called Fargo, the system helps customers with banking needs via voice or text.
From 21.3 million interactions in 2023, to more than 245 million in 2024, Wells Fargo’s internal stats show a dramatic rise in usage, with Spanish language adoption also surging, accounting for more than 80% of usage since its September 2023 rollout.
The bank’s approach is to build “compound systems,” where the orchestration layer determines which model to use based on the task, explains Wells Fargo CIO Chintan Mehta.