Capital One Blueprint for Enterprise Agentic AI: Data First, Platform Second

Capital One has been building toward agentic AI for 14 years. The bank's VP of Enterprise AI, Rashmi Shetty, is now operationalizing systems that don't just answer questions: they take action across the business.

In an interview with CIO Dive, Shetty outlined the company's core approach: invest in a robust data foundation first, then build agentic applications on a standardized, governed platform. No retrofitting governance after agents are deployed.

Key Takeaways:

• Capital One's flagship agentic use case, "Chat Concierge," helps customers navigate car purchases, connecting to dealers and advancing through the buying process -- a multiagent application running in production for over two years.
• Internally, agents improve the customer service journey. "Providing agent context becomes that much easier with a very strong data foundation."
• The platform-first approach means runtime controls, compliance checks, and cybersecurity guardrails are baked in before developers build -- not added later. Retrofitting governance to fragmented agentic apps is "far more difficult."
• Three pillars for agentic readiness: observability (tracking agent trajectories and tool accuracy), evals (sandbox simulations and vulnerability testing), and a harness strategy (standardized controls over tool permissions and actions).

Capital One's approach is a useful model: treat agentic AI as an end-to-end system, not a feature bolted onto existing workflows.

Read the full article on CIO Dive