How Capital One Built Its Own AI Brain Instead of Buying One Off the Shelf

Key points

• Capital One's multi-agent AI system, named MACAW, handles millions of bank fraud calls a year across interactions lasting four to sixty minutes.
• The bank fine-tunes open-weight models, AI models whose underlying code is publicly available, using proprietary customer data that outside providers cannot access.
• Chat Concierge, Capital One's car-shopping assistant, runs on a customised version of Meta's Llama, a widely used open-weight large language model.
• Kel Vanee, the bank's VP of machine learning engineering, said routing queries across multiple models can beat the accuracy of any single model alone.
• Vanee predicted that proactive, event-driven AI, software that acts without waiting to be asked, will be a major near-term trend in financial services.

Most big banks quietly license an AI model from one of the tech giants and call it done. Capital One took a different path.

At VB Transform 2026, Kel Vanee, the bank's VP of machine learning engineering, explained to VentureBeat why Capital One built its own AI infrastructure from the ground up rather than rely on ready-made products.

"At Capital One, we're not just using AI, we're building AI," Vanee said.

What did Capital One actually build?

The centrepiece is MACAW, short for Multi-Agent Capital One Workflow. A multi-agent system is software made up of several separate AI programs, each with a specific job, passing work between them like a relay team.

Fraud calls were the proving ground. A single large language model, the kind of AI that reads and writes text, could not handle a fraud conversation that might run an hour. So Capital One split the work across four agents.

First, an understanding agent works out what the customer actually wants. A reasoning agent then drafts a summary of the call. A validation agent checks that summary for accuracy. Finally, an explaining agent turns it into a clean document for the human specialist who takes over.

Several hundred customer-service agents who handle complex fraud cases no longer reconstruct long calls by hand. MACAW writes the post-call notes for them.

Why build on open-weight models instead of buying a finished product?

Capital One could have licensed a frontier AI model, a polished commercial product from companies like OpenAI or Google. Open-weight models, whose underlying code is publicly available and can be modified, gave the bank something a vendor contract cannot: access to its own data.

Decades of transaction records and customer interactions cannot be shared with a commercial AI provider. By training open-weight models on that private data, the bank bakes its expertise directly into the AI.

"We view our data as a huge advantage," Vanee said. "Something that nobody else has."

A useful side effect emerged. Training the AI to understand Capital One's internal policies and terminology lifted performance across other tasks too, not just the one it was built for.

Chat Concierge, the bank's customer-facing car-shopping assistant, runs on a customised version of Meta's Llama and uses the same four-agent structure as the fraud workflow. Llama is the open-weight model we first covered on 16 August 2026 in the context of developers gaining easier routes to deploy it.

What comes next for the bank's AI?

Vanee outlined two priorities. Model routing sends each query to whichever AI model is best suited to answer it, rather than sending everything to one. He believes routing across a broader set of models delivers better accuracy than any single model, at lower cost.

The second priority is proactive AI. Today most AI waits for a question. Vanee wants systems that detect a problem and act before anyone asks, spotting suspicious activity and flagging it without waiting for a customer to call. That ambition sits alongside a caution worth noting: Vanee said deploying such agents will demand rigorous testing and monitoring.

"Proactive AI is going to be a really important trend," he said.

For anyone paying attention to where banking AI is heading, the honest watch point here is cost. Capital One's build-not-buy bet only holds if proprietary data keeps producing measurable accuracy gains that a licensed model cannot match. The moment that gap closes, the arithmetic changes fast.

Common questions

Does this affect how Capital One customers experience the bank?

Yes, directly. MACAW already handles post-call notes for fraud specialists, meaning agents spend less time on paperwork and more time on complicated cases.

Is an open-weight model less secure than a commercial one?

Not by default. Security depends on how a model is deployed and governed. Capital One built governance and guardrails into the platform itself, so every agent operates within defined rules regardless of which underlying model powers it.