How Capital One prioritizes data and customizes AI models

Approaching the build versus buy question with generative AI and agentic AI technologies, Capital One has found that the answer is customization.

In a recent Targeting AI episode, Capital One's executive vice president, chief scientist and head of AI, Prem Natarajan, said the financial institution's GenAI strategy is focused on the ability to use the data it has while customizing open source or open weight models as opposed to closed models.

"The open weight stuff is important to us because our starting hypothesis was that all the data we have … is going to be a crucial way to improve the performance of these models," Natarajan said. "Over the last couple of years, that hypothesis has borne out to be true."

He added that using proprietary data rather than proprietary models has given Capital One an AI advantage.

Moreover,  customizing open source models has provided Capital One with the ability to ensure models remain optimized for the tasks they were designed for, while also improving their performance on the tasks the bank needs them to be optimized for.

"What we have found through a combination of these techniques … is that with [many] use cases we're able to take to production only because of the customization improvements we've been able to get," Natarajan said.

With this technique, Capital One has created some AI applications that are helpful to both employees and customers. For example, the banking institution has a GenAI-powered agent assist tool that associates use for customer service cases. Also, the financial giant uses AI-based for auto loan customers.

Esther Shittu is an Informa TechTarget news writer and podcast host who covers AI software and systems. Shaun Sutner is senior news director for Informa TechTarget's AI Business site. He is a veteran journalist with more than 30 years of news experience. Together, they host the Targeting AI podcast.