Capital One has built its multi-agent AI platform on open-weight models rather than relying on commercial off-the-shelf foundation models, according to Kel Vanee, the bank's MVP of machine learning engineering. The approach reflects a strategic pivot toward customization and control over vendor lock-in.
Speaking at VB Transform 2026, Vanee highlighted the bank's philosophy: "At Capital One, we're not just using AI, we're building AI." This distinction matters. While many enterprises adopt pre-built models from OpenAI, Anthropic, or Google, Capital One invested in deeply customized open-weight alternatives that align with its specific operational needs.
The choice stems from practical advantages. Open-weight models offer transparency into how the system works, reducing black-box dependencies that plague enterprise deployments. Capital One can audit, modify, and fine-tune these models on proprietary financial data without sharing sensitive customer information with third-party API providers. For a bank handling millions of accounts and transactions, data residency and security control are not negotiable.
Multi-agent architecture compounds these benefits. Rather than a single monolithic AI handling all tasks, Capital One deployed specialized agents that work in concert. One agent might handle customer inquiry routing. Another validates transactions. A third manages compliance checks. This modular approach improves reliability, allows easier updates to individual components, and creates clear audit trails for regulatory oversight. If one agent fails, others continue functioning.
The bank's timeline matters. Vanee noted that Capital One's "early investments in data transformation and cloud adoption" provided the foundation for rapid AI deployment once generative AI matured. Most large enterprises struggle here. They lack the data infrastructure, cloud-native architecture, or engineering talent to move fast. Capital One had already solved these problems years earlier, which accelerated its AI strategy when the current wave hit.
This strategy aligns with broader industry trends. Banks face intense pressure from regulators who demand explainability and control over automated decision-making systems. Open-weight models allow Capital One to document exactly how decisions get made, where data comes from, and what triggers customer-facing actions. This transparency becomes invaluable during regulatory audits or when defending decisions to customers.
The open-weight approach also hedges against vendor risk. If a third-party API provider changes pricing, imposes usage limits, or discontinues a service, Capital One's core systems remain operational. The bank owns the models running its infrastructure.
However, this path demands engineering resources. Building, training, and maintaining custom multi-agent systems requires deep expertise in machine learning operations, infrastructure scaling, and data governance. Capital One employs significant ML talent, giving it advantages smaller institutions lack. The bank's early cloud investments also mean it already had the DevOps and monitoring capabilities to run these systems reliably.
Capital One's approach reflects a maturation of enterprise AI strategy. Earlier adopters often grabbed whatever large language model seemed cutting-edge. Capital One instead treated AI as core infrastructure requiring the same rigor applied to payment systems or fraud detection. By controlling the model layer directly, the bank ensures its AI strategy remains aligned with business goals, regulatory requirements, and security standards for years to come.