Skip to content

banking · production case

Capital One

Chat Concierge

One of the few public, customer-facing multi-agent tools in US retail banking — a car-buying concierge, not a generic chatbot.

What it does

Chat Concierge coordinates several specialist agents to help a customer research and buy a car. Capital One’s Prem Natarajan described it publicly in 2025; NVIDIA GTC talks by Milind Naphade covered how it was built. Evident’s 2025 AI Index ranks Capital One second overall and first on talent density (about one in fifteen employees in AI roles).

Architecture

A customer channel talks to an orchestrator. Specialists handle inventory, financing, and eligibility. The bank has said the same agentic framework is being reused for other internal tools.

Evidence

Evident AI Index 2025 and Evident’s “Banks go agentic.” Independent of vendor marketing.

Caveat

Capital One publishes little operational telemetry (containment rate, handoff quality, loss given error). The architecture is public; the P&L is not.

Sources

Go deeper

Four briefings, grounded in this case and the rest of the observatory. Not a chatbot reciting the page.

← All production cases