Multi-Agent Customer Assistant — the Adoption Lesson

A bank’s customer query routing had weak intent recognition, sending customers to the wrong place and generating avoidable handoffs. I designed a supervisor-and-specialists architecture on Vertex AI, orchestrated with LangGraph — a supervisor agent classifying intent and routing to specialized agents per banking function, with anything consequential (payments, account changes) always routed to a human.

One architectural decision worth calling out: I led a migration from Text-to-SQL to Text-to-API. Letting a model compose arbitrary database queries turns your security boundary into a generated string — injection risk, unbounded queries, no stable contract to test, an unauditable trail. Text-to-API changes the model’s job to choosing a governed capability and filling in validated parameters — the same authorization, validation, and audit logging every other consumer of those endpoints gets.

Intent recognition accuracy improved substantially and hallucinations dropped significantly through iterative fine-tuning. But the number I lead with is customer satisfaction — because the real lesson was that being more accurate didn’t automatically mean people trusted the system more. A confident wrong answer with no visible way out is worse, experientially, than a hesitant one with a clear handoff. Since then I design the failure experience — what a user sees and can do when the system is unsure — before I design the happy path.