Me, Myself, and AI · MIT Sloan Management Review

Making Magic With Gen AI: Capital One’s Prem Natarajan

January 3, 2024·29 min·2 clips

Prem Natarajan reveals how Capital One balances product innovation and engineering to harness generative AI's potential.

This episode features a conversation with Prem Natarajan, Capital One's chief scientist and head of enterprise AI, hosted by professors Sam Ransbotham and Shervin Khodabandeh. The discussion focuses on implementing generative AI within a major financial services organization. Natarajan explains that Capital One's approach balances product innovation, scientific research, and engineering discipline. He details the company's "AI Factory" model, which standardizes tools and platforms for efficiency. A core principle is treating generative AI as a team sport requiring collaboration across diverse experts. The episode references specific applications like a tool that helps contact center agents summarize customer interactions. Natarajan emphasizes the importance of rigorous evaluation frameworks to measure AI performance beyond simple accuracy. He argues that successful AI integration requires reimagining business processes, not just automating old tasks. The conversation covers the challenge of managing both the cost and the computational latency of large language models. Natarajan shares that Capital One runs extensive internal experiments, or "AI hackathons," to explore use cases. He states the company's strategy involves a mix of building proprietary models, fine-tuning open-source models, and using external APIs. A surprising insight is his view that over-reliance on a single large vendor for AI tools can create strategic vulnerability. He claims the most valuable generative AI applications often start as solutions to internal employee pain points. Natarajan observes that current AI excels at tasks involving reasoning over text but struggles with complex numerical reasoning. He predicts a shift from today's general-purpose models to a future of more specialized, efficient "small language models." The hosts and guest discuss the critical need for proactive security measures and "red teaming" in generative AI deployments. Natarajan concludes that the biggest barrier to AI adoption is often organizational culture, not technology. The tone is educational and conversational, grounded in real-world corporate experience. The style is pragmatic, focusing on executable strategy rather than theoretical speculation. Listeners interested in the operational challenges of enterprise AI adoption would find this episode valuable. Those seeking debates on AI ethics or futuristic predictions might find the content too focused on practical implementation.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Me, Myself, and AI