Case Study: Capital One’s AI agents boost car dealership sales 55%

Capital One has built an agentic AI system for its auto business that mimics the company’s own organizational structure, with specialized agents working together like human teams. The approach has delivered impressive results, with dealership clients reporting a 55% improvement in engagement and sales leads, demonstrating how financial institutions can leverage AI to enhance customer service while maintaining rigorous oversight and risk management.

What you should know: Capital One began developing its agentic platform 15 months ago, before “agentic became a buzzword,” focusing on creating agents that problem-solve alongside customers like human agents do.

• The company studied how its human agents interact with customers to identify problems and incorporated those learnings into the AI system design.
• An evaluator agent trained on Capital One’s policies and regulations monitors other agents and can halt processes if problems are detected.
• Milind Naphade, SVP of Technology and Head of AI Foundations at Capital One, describes it as “a team of experts where each of them has a different expertise and comes together to solve a problem.”

How it works: The auto dealership agents assist the bank’s dealership clients in helping customers find the right car and car loan, with consumers able to browse vehicle inventories ready for test drives.

• The system enables more conversational, natural interactions that generate better quality leads.
• Agents operate 24/7, providing support even for situations like car breakdowns at midnight.
• Capital One used methods like model distillation (a technique that creates smaller, more efficient AI models from larger ones) for more efficient architectures, with the “understanding agent” representing the bulk of costs due to its disambiguation requirements.

The big picture: Capital One drew inspiration from its own organizational structure, recognizing that financial services require risk management and entities that “observe, evaluate, question and audit.”

• This same hierarchical oversight structure was applied to the AI agents, creating built-in governance and risk controls.
• The approach reflects how financial institutions can maintain regulatory compliance while deploying AI at scale.

What they’re saying: Naphade emphasized the challenge of pioneering without precedents in agentic AI development.

• “They’re able to generate much better serious leads through this more conversational, natural conversation,” he said about the dealership results.
• “But one of the biggest challenges we faced was that we didn’t have any precedents. We couldn’t go and say, oh somebody else did it this way, so we couldn’t ask how it worked out for them.”

What’s next: Capital One wants to bring similar agent technology to its travel business, particularly for customer-facing engagements related to its popular travel rewards credit card.

• The company recently opened a new lounge at New York’s JFK Airport and sees potential for AI agents in travel services.
• However, Naphade noted that extensive internal testing is required before deployment.

Technical details: The development process involved multiple iterations of experimentation, testing, evaluation, and human-in-the-loop validation before release.

• Capital One’s team of applied researchers, engineers, and data scientists used techniques like multi-token prediction and aggregated pre-fill for optimization.
• The company leveraged its extensive data resources while experimenting with optimal model architectures for agent systems.