Provided by Capital One
At VB Transform 2026, Kel Vanee, MVP of Machine Learning Engineering at Capital One, spoke with Sam Witteveen, Senior Technology Contributor at VentureBeat, about how the financial services company built a scalable, multi-agent AI architecture powered by deeply customized open-weight models instead of relying solely on off-the-shelf foundation models.
“At Capital One, we’re not just using AI, we’re building AI,” Vanee said.
Capital One’s early investments in data modernization and cloud adoption created the technical foundation needed to respond quickly to the latest advances in artificial intelligence, Vanee said. That foundation enabled the company to make several strategic architectural decisions, including developing a centralized enterprise AI platform with built-in governance, customizing open-weight models with proprietary data, and creating an internal multi-agent orchestration framework.
Customizing open-weight AI models with proprietary data
Instead of depending exclusively on commercial frontier models, Capital One uses its rich, proprietary data to fine-tune and customize open-weight AI models for financial services applications.
“We believe that our data is a huge advantage, something that no one else has and that the typical frontier models can’t provide. So we use that data to deeply customize these models,” Vanee explained. He added that real-time data is essential for delivering relevant context during live interactions with customers and employees.
Capital One has also discovered an important enterprise benefit from this strategy: model customization can improve performance across multiple use cases.
“We found that customizing these open-source models for one use case actually provides benefits across the portfolio,” Vanee said. “We’re training that model to become an expert in Capital One’s use cases, policies and nomenclature. As we’re doing that training, we’re seeing improvements across the board.”
How Capital One’s multi-agent AI workflow operates
Vanee highlighted a fraud-related customer service workflow that manages millions of calls each year. These conversations can last anywhere from approximately four minutes to an hour. Capital One’s initial experiments with a single large language model did not provide the reliability or control the company needed. With its Multi-Agent Workflow, or MACAW, the company routes each interaction through specialized AI agents with governance controls and built-in guardrails.
“The MACAW workflow consists of many different agents,” Vanee said. “The first is the understanding agent, whose purpose is to look at what the customer is saying and try to understand what they mean.”
Next, a reasoning agent receives specific instructions to generate a summary. A validation agent fact-checks that summary for accuracy. Finally, an explanation agent converts the approved summary into a structured document containing the required details and delivers it to the appropriate assignee.
In consumer banking, this multi-agent workflow supports hundreds of customer service representatives who handle complex scam-related calls. Automatically generated post-call summaries document lengthy customer interactions that representatives previously had to reconstruct manually, improving efficiency and consistency.
Capital One’s multi-agent AI architecture also powers Chat Concierge, an automated shopping assistant for customers. The system uses a version of Meta’s open-weight Llama model customized with Capital One’s proprietary data. Its specialized agents divide responsibilities: one communicates with the customer, another creates an action plan based on business rules, a third evaluates accuracy, and another explains and validates the final results.
Using agentic AI to optimize latency and costs
Beyond customer-facing applications, Capital One uses agentic AI to automate repetitive employee tasks and help researchers focus on higher-value work. One example is an autonomous optimization system that orchestrates the company’s backend AI hosting infrastructure.
Vanee explained that the large language model ecosystem introduces new optimization techniques every day, but those techniques do not always work well together. Combining two individually effective optimizations can sometimes reduce overall performance.
“This agent system runs through a search space designed by researchers, handles all the mechanics of setting up and running an experiment, and presents an overall summary of the results to the researchers,” Vanee said.
The system enables researchers to identify combinations of optimizations and configurations that produce measurable improvements, helping Capital One achieve lower latency and more efficient AI infrastructure costs.
The future of AI: model routing and proactive, event-driven systems
Looking ahead, Vanee expects model-routing abstraction layers to become increasingly important. These systems can evaluate and route requests across multiple AI models based on factors such as cost, accuracy and task complexity.
“We actually think we can get higher accuracy than any individual model just by routing across a broader set of available models, because different models are better at different areas,” he said.
Vanee also anticipates a shift toward proactive, event-driven AI systems. Rather than waiting for a human prompt, these systems can identify situations that require attention and take action based on predefined rules, policies and safeguards.
“What I think is going to be big in the future is more proactive, event-driven AI,” Vanee said. He emphasized that deploying these systems will require rigorous testing, monitoring and oversight.
“This will allow for more oversight and surveillance at scale, which will empower us as we fight fraud and respond to these opportunities,” Vanee added. “So proactive AI is going to be a very important trend.”
Advancing AI innovation in financial services
Capital One’s strategy reflects a broader lesson for enterprise technology leaders: delivering measurable value with AI requires more than adopting off-the-shelf software. Organizations need deeply customized models, strong governance, reliable data and scalable AI architectures.
By combining open-weight model customization, proprietary data and multi-agent orchestration, Capital One has created a repeatable blueprint for deploying enterprise AI across financial services use cases.
“All of these factors were absolutely critical to differentiating ourselves in this space and meeting our quality standards as well as the cost and latency thresholds we set for ourselves,” Vanee said.
As Capital One expands its AI capabilities, its centralized enterprise platform approach can help translate advances in machine learning into safer, faster and more personalized experiences for millions of customers.
Sponsored articles are content created by companies that pay us to post or have a business relationship with VentureBeat, and are always clearly marked. For more information please contact us [email protected].
Source: venturebeat.com