Wells Fargo

Wells Fargo serves tens of millions of retail banking customers who expect self-service access to bill payments, fund transfers, and account management around the clock. Scaling that support intelligently — without proportional staffing increases — required AI capable of handling natural-language requests at enterprise volume. The hard constraint was regulatory and trust-based: any architecture routing customer queries through external language models risked exposing personally identifiable information (PII), an unacceptable compliance posture for a federally regulated institution. Without a secure approach to AI orchestration, the bank faced a ceiling on automation ambition that no single off-the-shelf model deployment could overcome.

Wells Fargo CIO Chintan Mehta led the design of a layered poly-model, poly-cloud orchestration pipeline that separates intelligence tasks by security sensitivity. Speech is transcribed locally first; a small language model then screens the transcript for PII before any external call is made. Google Gemini Flash 2.0 handles intent classification, receiving only sanitized inputs, while all sensitive data processing stays on internal infrastructure — no customer data ever reaches an external model. The architecture is deliberately multi-model by task: Gemini handles customer-facing interactions, Claude and OpenAI models support internal developer workflows, and Meta's Llama serves certain back-office processes. For high-context, unstructured-data applications, the bank uses Gemini 2.5 Pro's one-million-token context window, which Mehta noted "absolutely killed it" for eliminating preprocessing overhead.

The AI assistant processed 245.4 million customer interactions in 2024, an 11.5x year-over-year increase from a baseline of 21.3 million interactions the prior year — all without exposing customer data to external models. A separate multi-agent deployment re-underwrote 15 years of archived loan documents autonomously, a back-office task that would have required large teams of human analysts under any prior approach.

• 245.4M customer interactions handled (2024)
• 11.5x year-over-year growth
• 21.3M prior-year baseline
• 15 years of loan archives processed via autonomous agent network

• Orchestration architecture, not model selection, is the primary performance lever — the gap between top LLMs is small; how they are composed into secure pipelines is the durable differentiator.
• PII screening at the pipeline boundary (using a local small model before any external API call) is a practical pattern for regulated industries to deploy cloud AI without compliance exposure.
• Context window size is a genuine differentiator for document-heavy workflows; matching model to task — rather than defaulting to one model for everything — reduces preprocessing cost and latency.
• Multi-agent systems make previously infeasible back-office work tractable at scale, unlocking backlogs that no single deployment could address.