Capital One Builds Multi-Agent AI Platform with Custom Open-Weight Models

At VB Transform 2026, Kel Vanee, MVP of machine learning engineering at Capital One, shared insights into the bank’s journey in building a scalable multi-agent AI platform. Instead of relying on off-the-shelf foundation models, Capital One has developed a deeply customized system using open-weight models and proprietary data.

Vanee emphasized that the groundwork for this approach was laid years ago with early investments in data transformation and cloud adoption. "At Capital One, we're not just using AI; we're building AI," he stated. This technical foundation enabled the company to make several deliberate architectural decisions, including:

• Centralized, Enterprise-Wide AI Platform: A unified platform with built-in governance ensures consistency and security across the organization.
• Deep Customization of Open Models: Proprietary data is used to fine-tune open-weight models, creating specialized agents for various use cases.
• Multi-Agent Orchestration Harness: A custom-built system that coordinates multiple AI agents for complex workflows.

The Power of Proprietary Data

Capital One views its rich proprietary data as a significant advantage. "We see our data as something unique and invaluable," Vanee explained. "General frontier models can't provide the same level of context and nuance." Real-time data is crucial for bringing fresh context during live customer interactions, ensuring that AI agents are always up-to-date.

One key benefit of this approach is its extensibility across the enterprise. "As we customize these open-source models for one use case, we see benefits across our entire portfolio," Vanee noted. For example, a model trained to handle bank fraud can also improve other customer service workflows by learning Capital One’s specific policies and nomenclature.

Case Study: Customer Service Workflow

A prime example of this approach in action is the customer service workflow for handling bank fraud, which processes millions of calls annually. Interactions range from four minutes to as long as sixty minutes, making it a complex challenge for AI.

Initially, using a single large language model proved insufficient. "We needed something more nuanced and specialized," Vanee said. Enter MACAW (Multi-Agent Customer Assistance Workflow), a multi-agentic system designed to handle these interactions efficiently.

"The MACAW workflow is made up of several different agents," Vanee explained. These include:

• Understanding Agent: Analyzes customer input to determine the nature of the issue.
• Policy Agent: Ensures that all actions comply with Capital One’s policies and regulations.
• Context Agent: Provides real-time data and context to inform the interaction.
• Response Agent: Generates appropriate responses and next steps.

Each agent is specialized and works in concert to provide a seamless customer experience. Governance and guardrails are built into the system to ensure security and compliance.

Under the Hood

The technical implementation of Capital One’s multi-agent AI platform involves several key components:

• Data Pipeline: A robust data pipeline ensures that real-time data is continuously fed into the system, keeping models up-to-date.
• Model Training: Proprietary data is used to fine-tune open-weight models, creating specialized agents for various tasks.
• Orchestration Layer: The orchestration layer coordinates the interaction between different agents, ensuring smooth and efficient workflows.
• Governance Framework: A comprehensive governance framework ensures that all AI activities comply with regulatory requirements and company policies.

Vanee highlighted the importance of continuous learning and improvement. "We are always iterating and refining our models based on new data and feedback," he said. This iterative process helps Capital One stay ahead in the rapidly evolving field of AI.

Key Takeaways

• Customization Over Off-the-Shelf: Capital One’s success with multi-agent AI underscores the value of deeply customizing open-weight models using proprietary data.
• Real-Time Data is Crucial: Real-time data enhances the effectiveness of AI agents, ensuring they are always up-to-date and contextually relevant.
• Extensibility Across Use Cases: Customized models can benefit multiple use cases within an organization, providing a general lift in performance and efficiency.

By building a multi-agent AI platform that leverages proprietary data and open-weight models, Capital One is setting a new standard for AI in financial services. This approach not only enhances customer experiences but also ensures compliance and security, making it a model for other organizations to follow.