Advancing the State of the Art: The Frontier of Enterprise Agentic AI

TL;DR

• A four-role multi-agent framework: an understanding agent (reads the environment or the customer and figures out what's needed), a planner agent (holds the APIs, the knowledge, and Capital One's policies), an independent evaluator agent that validates whether executing the plan would fit policy and be safe — kicking it back to the planner, or further back to the understanding agent, when it wouldn't — and an explainer agent that translates inter-agent chatter into human language. Naphade calls the evaluator the secret sauce.
• This is not RPA with new branding: "This is not robotic just done differently with agents. This is an inherently non-deterministic system of agents that works together to achieve outcomes."
• Already in production: a car-buying chat concierge running behind dealerships nationwide (vehicle needs, availability, appointments, trade-ins — fully autonomous, 24/7/365), and a consumer-bank assistant for fraud-related customer conversations (understanding the complaint, recommending the right actions for the human agent, then summarizing the conversation into a learning corpus).

Key Points

Positioning: a technology company that knows how to manage risk (~02:00–02:02)

Capital One has always framed itself as a technology-first company, and Naphade's thesis is that the winner in banking will be a technology company — one with deep technology roots and deep competence at managing risk. The transformation has been running for years: the first large bank fully on the public cloud, a modern data ecosystem, and a data-driven analytical culture dating to the company's origins. The last three and a half years brought heavy investment in generative and agentic AI, making Capital One one of the first enterprises to put agentic AI into production.

His organization, AI Foundations, is mostly AI researchers, applied AI engineers, and data scientists, chartered to deliver business impact through scientific innovation:

• 65+ publications this year at leading conferences and journals (ICML, ICLR, NeurIPS, ACM, and others).
• Academic partnerships with the University of Southern California, University of Illinois Urbana-Champaign, Columbia University, and others, focused on advancing safe AI.

Sample recent work, mentioned in passing:

• Generating curated datasets for multi-turn conversations in multi-agent environments — which he stresses is a materially different problem from synthetic data generation for single-turn or single-agent interactions.
• Routing generated data through multiple models — "think of this as a mixture of models instead of a mixture of experts" (presented at ACL).
• Critic-guided distillation for robust reasoning (at ICML).

The through-line: true differentiation comes only from customizing the entire AI stack with your own data. "Capital One's AI advantage is Capital One's data advantage being converted into an AI advantage."

The framework and its independent evaluator (~02:04–02:06)

Capital One began exploring agentic AI more than two and a half years ago and had its first multi-agent production system running last January. The overall framework (heard in the captions as "MACA" — spelling unverified) comprises four classes of agent:

1. Understanding agent — interacts with the environment or the customer to determine what is actually needed.
2. Planner agent — has access to all the APIs, all the knowledge, and all of Capital One's policies, and works out how to satisfy the need.
3. Evaluator agent — the piece he singles out as secret sauce: a completely independent agent that uses a simple world model to determine whether executing the plan would fit Capital One policy and be safe to do. If not, it kicks the plan back to the planner, which can in turn go back to the understanding agent for better input from the customer.
4. Explainer agent — the other agents talk in their own language; this one speaks in terms humans understand.

This is a non-deterministic system at play, naturally. This is not robotic just done differently with agents — this is an inherently non-deterministic system of agents that works together to achieve outcomes.

Two production cases, and a hiring pitch (~02:06–02:08)

• Car-buying concierge: dealerships across the country run Capital One on the back end, so a buyer is most likely interacting with this concierge — working out what they're looking for in a vehicle, availability, appointments, and trade-ins, fully autonomously, 24/7/365.
• Fraud-related conversations (debuted last year, serving the consumer bank): understanding the customer's complaint, recommending the right set of actions for the human agent to take, and summarizing the conversation to build a learning corpus for the next round of interactions.

He closed with recruiting. The team customizes at four layers — the foundation models themselves, the services around them, the agentic layer, and the solution layer — across three job families. The pitch: going from research to production in record time inside a regulated enterprise ("typically unheard of"), and seeing 130 million customers use your work and feed a continuous learning loop.

Quotes

"Capital One's AI advantage is Capital One's data advantage being converted into an AI advantage." (~02:03)

"This is not robotic just done differently with agents. This is an inherently non-deterministic system of agents that works together to achieve outcomes." (~02:05)