At VB Transform 2026, Kel Vanee, MVP of machine learning engineering at Capital One, spoke with Sam Witteveen, Senior Technology Contributor at VentureBeat, about the bank’s approach to building an enterprise AI platform that relies on deeply customized open-weight models and a multi-agent orchestration layer rather than solely on off-the-shelf foundation models.
Technical foundation and architectural choices
Vanee said years of prior investment in data transformation and cloud adoption enabled Capital One to move quickly when current AI advances arrived. That foundation supported several deliberate architectural choices: creating a centralized, enterprise-wide AI platform with built-in governance, deeply customizing open models with proprietary data, and developing an in-house multi-agent orchestration harness.
Fine-tuning open-weight models with proprietary data
Instead of relying exclusively on commercial frontier models, Capital One fine-tunes open-weight models using its proprietary data. Vanee described the bank’s data as a significant competitive advantage that general frontier models cannot provide, and said real-time data is critical to supply fresh context during live customer or associate interactions.
He also noted an unexpected benefit: extensibility across the enterprise. As the bank customizes an open-source model for a specific use case, the improvements often carry over to other parts of the portfolio because the model becomes an expert in Capital One’s use cases, policies, and nomenclature.
Inside the multi-agent workflow (MACAW)
Vanee cited a customer-service workflow for bank fraud as an example. The system handles millions of calls per year, with interactions ranging from about four minutes to as long as sixty minutes. An initial attempt to use a single large language model proved insufficient, so the team developed a multi-agent workflow called MACAW that routes interactions through specialized agents with governance and guardrails.
The MACAW workflow uses several agents: an understanding agent that interprets customer intent; a reasoning agent that generates a summary following specific instructions; a validation agent that fact-checks the summary; and an explaining agent that formats the summary into a documented result shared among agents.
For consumer banking, this workflow supports several hundred customer-service agents who specialize in complex fraud calls. The post-call summaries produced by MACAW document lengthy, back-and-forth interactions that agents previously had to reconstruct manually.
Chat Concierge and Llama customization
The same multi-agent architecture underpins Chat Concierge, a customer-facing auto-shopping assistant. For that product, Capital One leverages a version of Meta’s open-weight Llama model customized with the bank’s proprietary data. The product uses the same division of labor: one agent converses with the customer, another constructs an action plan from business rules, a third evaluates accuracy, and a fourth explains and validates the outcome.
Agentic research system to optimize latency and cost
Capital One also applies agentic AI internally to automate repetitive tasks and let employees focus on higher-leverage work. One example is an autonomous agentic optimization system that tunes backend hosting infrastructure.
Vanee explained that in the LLM world new optimizations appear daily and are not always complementary; combining two good optimizations can sometimes degrade performance. Their agentic system explores a researcher-defined search space, handles the mechanics of setting up and running experiments, and then presents a summarized report of results to the researcher. This helps identify sequences of optimizations and configurations that deliver the best possible latency.
What’s next: model routing and proactive, event-driven AI
Looking ahead, Vanee highlighted two trends. First, platforms will increasingly validate routing abstraction layers that span multiple models to balance cost and accuracy. He argued that routing across a broader set of models can yield better accuracy than any single model because different models excel in different areas.
Second, he expects a shift toward more proactive, event-driven AI—systems that act without waiting for human prompts. Vanee stressed that deploying such proactive agents will require rigorous testing and continuous monitoring, but said these capabilities will enable larger-scale monitoring and empower efforts to fight fraud and seize other opportunities.
Lessons for enterprise AI in financial services
Capital One’s approach illustrates that extracting measurable business value from AI in enterprises often requires moving beyond off-the-shelf software to deeply customized, governed architectures. By combining fine-tuned open-weight models, a multi-agent orchestration harness, and proprietary data assets, the bank has built a repeatable blueprint for deploying scalable AI in financial services.
“All of those ingredients were absolutely critical to differentiating in this space and hitting the quality bars as well as the cost and latency thresholds we set for ourselves,” Vanee said.
As Capital One expands these capabilities to new use cases, its enterprise platform approach aims to translate technical breakthroughs into safer, faster, and more personalized experiences for millions of customers.
This article was produced from content sponsored by Capital One.
