Capital One's MACAW: Open-weight models, multi-agent orchestration, and enterprise AI governance in banking

Most large financial institutions default to licensing frontier models from OpenAI or Anthropic. Capital One went the other direction. At VB Transform 2026, Kel Vanee, the bank's MVP of machine learning engineering, detailed how Capital One built a scalable multi-agent AI platform around deeply customized open-weight models and an in-house orchestration harness called MACAW. The approach treats proprietary data as the core differentiator and uses a disciplined agent workflow to handle millions of fraud-related calls per year.

MACAW: A four-agent workflow for post-call summaries

Capital One's initial attempt at using a single large language model for fraud-call summarization proved insufficient. The MACAW workflow solves that by routing each interaction through four specialized agents with built-in governance and guardrails VentureBeat.

• Understanding agent interprets the customer's intent from the conversation.
• Reasoning agent generates a summary based on specific instructions.
• Validation agent fact-checks the summary for accuracy.
• Explaining agent formats the validated summary into a document shared with customer-service agents.

This workflow helps several hundred agents who handle complex fraud calls, replacing the manual reconstruction of long, back-and-forth interactions that could last anywhere from four to sixty minutes.

Why open-weight models and proprietary data

Rather than relying on off-the-shelf frontier models, Capital One fine-tunes open-weight models using its own data. "We view our data as a huge advantage and something that nobody else has, something that the general frontier models cannot provide," Vanee said. Real-time data is critical for bringing fresh context during live customer or associate interactions.

An unexpected benefit emerged: extensibility across the enterprise. As Capital One customizes an open model for one use case, the training on Capital One-specific policy, nomenclature, and use cases produces a general lift across the entire portfolio. The same agentic division of labor also powers Chat Concierge, a consumer auto-shopping assistant built on a customized Meta Llama model, with one agent conversing with the customer, one building an action plan from business rules, one evaluating accuracy, and one explaining and validating the result.

Beyond customer-facing agents: internal infrastructure tuning

Capital One also applies agentic AI to internal operations. The bank built an autonomous agentic optimization system that tunes backend hosting infrastructure. In the LLM world, new optimizations arrive daily, but combining two good optimizations can cause performance regression. The agentic system runs through a researcher-designed search space, handles experiment mechanics, and presents a summarization of results, helping researchers find the configuration that gives the best latency.

What this means for AI builders

Capital One's architecture demonstrates that enterprise AI value can come from data-tuned open models and disciplined orchestration, not just from frontier foundation models. For teams building in regulated industries, the MACAW approach offers a repeatable blueprint: start with a centralized platform, customize open-weight models with proprietary data, enforce governance through agent-level guardrails, and reuse tuned models across use cases. The bank's experience also suggests that routing across multiple models can yield better accuracy than any single model, a pattern more builders should consider.

Caveats and limitations

The evidence comes primarily from VentureBeat's coverage of Capital One's sponsored presentation at VB Transform 2026. Implementation specifics, benchmark results, and cost comparisons were not disclosed. Vanee's predictions about proactive, event-driven AI and multi-model routing are stated as forward-looking views and would require rigorous testing and monitoring in production. The article itself is marked as sponsored content, so readers should weigh the source accordingly.