Capital One is building its multi-agent AI infrastructure on customized open-weight models instead of relying on proprietary foundation models from major cloud providers or AI labs. The choice reflects a broader strategic shift among large enterprises to own their AI stack rather than depend on third-party vendors.
Kel Vanee, MVP of machine learning engineering at Capital One, revealed the architecture at VB Transform 2026, explaining that the bank prioritized customization and control over convenience. "At Capital One, we're not just using AI, we're building AI," Vanee said. This signals the bank views AI as a core competitive capability rather than a commodity tool.
Open-weight models, which are publicly available but require companies to fine-tune and deploy them on their own infrastructure, give Capital One several advantages. The bank gains control over model behavior, can optimize for its specific use cases like fraud detection and customer service, and reduces dependency on external API providers. It also sidesteps the cost and latency issues of calling cloud-based APIs for routine AI tasks.
The multi-agent architecture lets Capital One deploy specialized models for different functions. One agent might handle customer inquiries, another manages transaction analysis, and a third processes compliance checks. This modular approach scales better than a single monolithic model and allows the bank to update specific agents without retraining the entire system.
Capital One's foundation for this move runs deeper than just recent AI trends. Vanee emphasized that years of prior investment in data transformation and cloud migration positioned the bank to move quickly when generative AI arrived. That groundwork matters because building an in-house AI platform requires robust data pipelines, cloud infrastructure, and ML engineering talent. Companies that skipped those investments now scramble to play catch-up.
The financial services industry faces unique constraints around AI deployment. Banks deal with sensitive customer data, regulatory oversight, audit requirements, and compliance frameworks that make black-box solutions from third parties risky. Building proprietary models on open-weight architectures lets Capital One maintain full visibility into model decisions and audit trails, critical for meeting banking regulations like GLBA and fair lending rules.
Capital One's approach also reflects confidence in the maturity of open-source AI. Models like Meta's Llama, Mistral, and others have reached production quality. Companies no longer need to choose between building from scratch or buying from OpenAI. The open-weight ecosystem now offers a third path: take a proven model, customize it, and own the deployment.
This strategy puts Capital One in a different camp than many traditional financial institutions that signed big deals with OpenAI or Microsoft for GPT integration. Capital One bets that control and customization outweigh the convenience of managed services. That positions the bank to differentiate on AI rather than compete on price with other OpenAI customers.
The shift also reflects labor market realities. Capital One, headquartered in Richmond, Virginia, has deep expertise in data engineering and machine learning. Building custom models keeps that talent engaged in high-value work rather than prompt engineering off-the-shelf systems.
As more enterprises discover the limitations of fine-tuning closed models or hitting rate limits on API calls, Capital One's bet on open-weight customization may become the template other banks follow.