Do you want smarter information in your reception box? Sign up for our weekly newsletters to obtain only what matters for business managers, data and security managers. Subscribe now
Inspiration can come from different places, even for the architecture and design of agent systems.
At VB Transform, Capital One explained how he built his agentic platform for his automotive activity. Milind Naphade, please of technology and head of AI foundations at Capital One, said in the transformation of VB that the company wanted its agents to operate similarly to human agents, in that they solve problems alongside customers.
Naphade said that Capital One began to design its agent offers 15 months ago: “Before the agency became a fashionable word.” For Capital One, it was crucial that, to build its agent systems, they learn how their human agents ask customer information to identify their problems.
Capital One also turned to another source of organizational structure for its agents: itself.
“We were inspired by the functioning of Capital One himself,” said Naphade. “In Capital One, as I am sure in other financial services, you must manage the risks, then there are other entities that you observe, assess, question, question and check.”
>> See all of our Transform 2025 coverage here <This same structure applies to agents that Capital One wishes to monitor. They created an agent who assesses existing agents, which was trained in the policies and regulations of the capital. This evaluator agent can resume the process if he detects a problem. Naphade said to consider him as “a team of experts where each of them has a different expertise and meets to solve a problem”.
Financial service organizations recognize the potential of agents to provide their human agents with information to solve customer problems, manage customer service and attract more people in their products. Other banks like BNY have deployed agents this year.
Automobile dealerships
Capital One has deployed agents in its automotive activity to help bank dealership customers to help their customers find the right car and car loan. Consumers can consult dealership vehicle stocks for testing. Naphade said their concessionary customers have declared an improvement of 55% of parameters such as commitment and serious prospects.
“They are able to generate much better serious tracks thanks to this more conversational and natural conversation,” he said. “They can operate agents 24/7, and if the car breaks down at midnight, the cat is there for you.”
Naphade said that Capital One would love to bring this type of agent to his travel activity, in particular for his customer -oriented commitments. Capital One, which has opened a new show at JFK airport in New York, offers a very popular credit card for travel points. However, Naphade stressed that the bank must carry out large internal tests.
Data and models for banking agents
Like many companies, Capital One has a lot of data for its AI systems, but it must find the best way to bring this context to its agents. He must also experience the best model architecture for his agents.
The team of researchers, engineers and applied data from Naphade and Capital One used methods such as the distillation of the model for more effective architectures.
“The agent of understanding is the essential of our cost because it is the one who must disamboret,” he said. “It is a larger model, so we try to distribute it and do much more overflow for our money. Then there is also a multi-token prediction and aggregate prefabillation, many interesting ways that we can optimize this.”
In terms of data, Naphade said that his team had undergone several “iterations of experimentation, tests, evaluation, man in the loop and all straight guards” before publishing its AI applications.
“But one of the biggest challenges we faced is that we had no precedents. We couldn’t go and say, oh someone else did this way, so we couldn’t ask how it worked for them?” Said Naphade.
Publisher’s note: as thanks to our readers, we have opened the recording of highlights for VB Transform 2026 – only $ 200. This is where the ambition of the AI ​​meets operational reality, and you will want to be in the room. Book your place now.
