Capital One Built Its Own Multi-Agent AI Platform and It Handles Millions of Fraud Calls a Year

Quick Facts

• Capital One uses Meta’s open-weight Llama model, customized with proprietary data, instead of closed-source alternatives like GPT-4 or Claude.
• The company’s MACAW multi-agent workflow routes customer fraud calls through four specialized agents: understanding, reasoning, validation, and explaining.
• Capital One’s internal AI tool has been used more than 10,000 times and supports thousands of agents across the company.

Capital One built its own multi-agent AI platform from the ground up, and the bank chose open-weight models as its foundation for one specific reason: closed-source models cannot be meaningfully customized.

Kel Vanee, MVP of machine learning engineering at Capital One, made that case at VB Transform 2026, explaining how the bank’s proprietary data is its primary competitive weapon. “We view our data as a huge advantage and something that nobody else has, something that the general frontier models cannot provide,” Vanee said. “So we are taking that data and deeply customizing these models.”

Capital One’s Chief Scientist and Head of AI, Prem Natarajan, put it directly: “Your data advantage is your AI advantage.”

Why Open-Weight Models

Capital One evaluated closed-source models and ruled them out. The bank’s AI team selected Meta’s Llama and began building on top of its public cloud infrastructure. The bank does not simply pull models from Hugging Face. Milind Naphade, SVP of AI Foundations, said the team starts with open-source models and customizes them “to the point where it’s almost unrecognizable.”

To meet regulatory requirements, Capital One’s team fully dissects each model, studying its architecture and tracing its lineage. That level of scrutiny is not possible with a closed system.

The MACAW Workflow

The bank’s multi-agent system is called MACAW, short for Multi-Agentic Architecture workflow. It was built to handle customer fraud calls, a process that can last anywhere from four minutes to one hour and involves millions of calls annually.

A single large language model could not handle the complexity. MACAW routes each interaction through four specialized agents. An understanding agent reads customer intent. A reasoning agent produces a summary. A validation agent checks that summary for accuracy. An explaining agent formats the final output into a document shared with human bank agents.

“The MACAW workflow is made up of a number of different agents,” Vanee said. The system was also presented at NVIDIA GTC 2026, where Capital One described how it supports several hundred bank customer service agents handling complex fraud resolution calls.

Infrastructure and Stack

Capital One’s technology stack combines in-house tools, open-source toolchains, and NVIDIA’s inference stack. The retrieval-augmented generation service runs on AWS OpenSearch and deploys LLM inference through Amazon SageMaker, backed by NVIDIA A100 GPUs. The retrieval layer supports keyword, vector, and hybrid search with reranking models.

Capital One worked directly with NVIDIA to hit performance targets and to prioritize features in the Triton server and TensorRT LLM. Real-time data feeds are built into the system to bring fresh context into live customer and associate interactions.

Governance Built In

The platform separates agent design from runtime governance. Policies, guardrails, and security controls are embedded across agent boundaries. The team also built for observability and evaluation of stochastic, multi-agent workflows.

“The main breakthrough for us was realizing that this had to be dynamic and iterative,” Naphade said. The system accounts for the reality that customers often need multiple conversation turns before their intent is fully understood.

Capital One also released an open-source security tool called VulnHunter in July 2026. The agentic tool scans source code for vulnerabilities, maps attack paths, and proposes fixes before code ships to production. It is available on GitHub under an Apache 2.0 license and was used internally to identify and fix vulnerabilities across thousands of repositories spanning tens of business areas.

With more than 100 million customers, the scale of Capital One’s AI deployment sets a high bar. The bank’s internal AI tool has logged more than 10,000 uses and supports thousands of agents company-wide. “At Capital One, we’re not just using AI, we’re building AI,” Vanee said.

Read more: Why Capital One built its multi-agent AI platform around open-weight models