Join an event that enterprise leaders have been trusted for nearly 20 years. VB Transform brings together people who build real enterprise AI strategies. learn more
Inspiration can come from a variety of locations, whether it is the architecture and design of an agent system.
in VB conversion, Capital 1 He explained how he built the agent platform for his automotive business. Milind Naphade, Technology’s SVP and Head of Capital One’s AI Foundation, said during VB Transform that the company hopes agents will function similarly to human agents in terms of solving problems with customers.
Naphade said Capital One began designing agent offerings 15 months ago “before agents become buzzwords.” In Capital One, it was important to learn from how human agents would ask customers for information to identify problems when building an agent system.
Capital One also turned its attention to another source of the agent’s organizational structure, itself.
“We were inspired by how Capital One itself works,” says Naphade. “In Capital 1, you need to manage your risk as you are sure you are within other financial services. Then there are observations, evaluations, questions, and audits by other entities.”
>> See all 2025 coverage here <This same structure applies to agents that Capital One wants to monitor. They created an agent that evaluates existing agents. Existing agents were trained in capital policies and regulations. If this evaluation agent detects a problem, it can kick back the process. Naphade said he considers it a “team of experts, each with a different expertise and is gathered to solve problems.”
Financial services organizations recognize the possibility that agents can provide human agents with information to solve customer problems, manage customer service and attract more people to their products. Other banks you like bny This year we deployed an agent.
Car dealer agent
Capital One has deployed agents in the automotive business to help bank dealer clients help their clients find the right car and car loans. Consumers can see the vehicle inventory of dealers ready for test drives. Naphade said dealer customers reported a 55% improvement in metrics such as engagement and serious sales leads.
“They can produce a much better, serious lead through this more conversational, more conversation,” he said. “They can make agents work 24/7. If your car breaks down in the middle of the night, chat is for you.”
Naphade wants Capital One to bring this type of agent into the travel business. With its new lounge open at JFK airport in New York, Capital One offers extremely popular credit cards for travel points. However, Naphade noted that banks need to carry out extensive internal testing.
Bank Agent Data and Models
Like many companies, Capital One has a lot of data on AI systems, but you need to know how best to bring that context to your agent. You also need to experiment with the model architecture that is best suited to your agents.
A team of applied researchers, engineers and data scientists from Naphade and Capital One used methods such as model distillation for a more efficient architecture.
“An understanding agent is a large part of our cost, because it’s something that needs to be clarified,” he said. “It’s a bigger model, so we’re going to distribute it and try to get a lot of bang for our spending. And there are also multi-token predictions and aggregated prefills. There are many interesting ways to optimize this.”
Regarding data, Naphade said the team had undergone several “experiments, tests, evaluations, loops, and all the right guardrails human iterations” before releasing the AI application.
“But one of the biggest challenges we faced was that there was no precedent. We couldn’t go,” Nafade said.
Editor’s Note: As a thank you to our readers, we have opened the Early Bird Registration for VB Transform 2026 – just $200. This is where AI ambitions meet operational reality and you’ll want to be in your room. Book your location now.
