Editorial illustration for Capital One finds open AI models deliver enterprise-wide benefits
Capital One Ditches Proprietary AI for Open Models
Capital One finds open AI models deliver enterprise-wide benefits
Capital One's AI strategy runs against the industry default. Instead of building its multi-agent systems on top of a single frontier foundation model licensed from OpenAI or Anthropic, the bank has spent years customizing open-weight models with its own data and running them through an orchestration layer it built in-house. Kel Vanee, MVP of machine learning engineering at Capital One, laid out the thinking behind that choice at VB Transform 2026, in a conversation with VentureBeat senior technology contributor Sam Witteveen.
The decision wasn't made overnight. Vanee traced it back to Capital One's early bets on data transformation and cloud migration, investments that gave the bank the technical footing to move fast once generative AI tools matured. That groundwork shaped three specific architectural calls: a centralized AI platform with governance built in from the start, deep customization of open models using proprietary data, and a homegrown harness for coordinating multiple agents at once.
Vanee's framing of the company's role was direct. Capital One isn't just deploying someone else's AI, he said, it's building its own. What that building process looks like, and why it pays off beyond any single use case, is where the conversation went next.
Capital One’s approach underscores a broader truth for enterprise technology leaders: driving measurable value with AI requires moving beyond off-the-shelf software toward deeply customized, highly governed architectures. By combining fine-tuned open-weight models, a multi-agent orchestration harness, and proprietary data assets, the bank has established a repeatable blueprint for deploying scalable AI in financial services.
Why this matters
Capital One's bet is a useful data point for anyone weighing build-versus-buy on foundation models. The bank isn't customizing open-weight models to save money on tokens; it's treating fine-tuning as an asset that compounds across the business. Train a model on Capital One's policy language and use cases once, and Vanee says the gains show up in other parts of the portfolio, not just the original project.
That's a different argument than the usual open-weight pitch, which tends to focus on cost or control. It's closer to compounding institutional knowledge into the weights themselves.
For developers and founders, the lesson isn't "open models beat closed ones." It's that ownership of the customization layer has value that doesn't show up on a pricing page. If you're locked into someone else's foundation model, you're renting improvements. If you own the weights, you're banking them. Worth asking, though: how much of this only works because Capital One had years of clean data and ML infrastructure already in place before any of this started.
Related Reading
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps Windows Copilot AI unable to pinpoint image source in user test
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
From Archives NVIDIA RTX 5090 Powers On-Device Agentic AI with Meta's Muse Glimmer
From Archives The Vergecast: Google AI Shake-Up, Model Wars, and Ads in Your BMW
Common Questions Answered
Why does Capital One use open-weight models instead of licensing frontier models from OpenAI or Anthropic?
Capital One chose to customize open-weight models with its own data and run them through an in-house orchestration layer rather than relying on a single licensed frontier model. This approach allows the bank to create deeply customized, highly governed architectures that deliver measurable enterprise-wide value beyond what off-the-shelf software can provide.
How does Capital One's multi-agent orchestration layer work with fine-tuned models?
Capital One built an in-house orchestration layer that coordinates its customized open-weight models to work as a multi-agent system. This proprietary infrastructure allows the bank to combine fine-tuned models with its proprietary data assets to create a scalable AI deployment blueprint for financial services.
What are the compounding benefits of fine-tuning open-weight models at Capital One?
According to Kel Vanee, Capital One's MVP of machine learning engineering, when a model is trained on the bank's policy language and use cases, the gains extend across multiple parts of the business portfolio, not just the original project. This demonstrates that fine-tuning creates a reusable asset that compounds value throughout the enterprise rather than being limited to a single application.
How does Capital One's AI strategy differ from the industry default approach?
While most enterprises default to building multi-agent systems on top of a single frontier foundation model licensed from providers like OpenAI or Anthropic, Capital One has spent years customizing open-weight models with proprietary data and governance structures. This customized approach prioritizes measurable value and scalability over convenience of off-the-shelf solutions.
Further Reading
- How Capital One drives returns on its AI investments - CIO
- Five insights into how Capital One is gaining momentum with enterprise AI - VentureBeat
- How open models solved Capital One's AI problems - The Deep View
- Capital One: The Ongoing Story Of How One Firm Has Been Pioneering Data Analytics & AI Innovation For Over Three Decades - Forbes
- How AI is transforming financial services & banking - Capital One Tech
