Capital One Builds MACAW Multi-Agent AI on Customized Open-Weight Models
Capital One details MACAW, a governed multi-agent AI platform using customized open-weight models for fraud calls, auto shopping and infrastructure tuning.
Summary
In a sponsored VentureBeat article presented by Capital One and published August 13, 2026, Kel Vanee, MVP of machine learning engineering at Capital One, told Senior Technology Contributor Sam Witteveen at VB Transform 2026 that the bank built a centralized, enterprise-wide AI platform around customized open-weight models, proprietary data, built-in governance and its own multi-agent orchestration harness. Earlier data transformation and cloud adoption enabled the strategy. Unlike off-the-shelf frontier models, Capital One fine-tunes open models with proprietary and real-time data for fresh interaction context. Training on one use case teaches Capital One policy and nomenclature, producing gains across its portfolio.
For fraud customer service handling millions of calls annually, each roughly four to 60 minutes, a single large language model was insufficient. MACAW uses an understanding agent to identify intent, a reasoning agent to summarize under specific instructions, a validation agent to fact-check, and an explaining agent to produce a detailed formatted document. The summaries replace manual reconstruction for several hundred agents specializing in complex fraud calls. Chat Concierge, Capital One’s customer-facing auto-shopping assistant, uses a proprietary-data customized version of Meta’s open-weight Llama with agents for conversation, business-rule action planning, accuracy evaluation, explanation and validation.
Capital One also built an agentic system to automate rote employee work. It traverses a researcher-designed search space, runs experiments, summarizes results and identifies backend hosting configurations with the best latency, addressing cases where individually useful LLM optimizations combine into performance regressions. Vanee expects routing layers to validate across multiple models for cost and accuracy, potentially outperforming any single model, followed by proactive, event-driven AI that acts when conditions warrant. Such agents require rigorous testing and monitoring but could expand fraud monitoring. The platform targets Capital One’s quality, cost and latency thresholds while delivering safer, faster and more personalized service to millions of customers.
Positives
- MACAW summarizes millions of annual fraud calls lasting roughly four to 60 minutes for several hundred specialized customer-service agents.
- Meta’s open-weight Llama powers Chat Concierge after customization with Capital One’s proprietary data.
- Real-time data supplies fresh context during live customer and associate interactions.
- Fine-tuning on Capital One policies and nomenclature produces gains beyond the original use case across the bank’s portfolio.
- Capital One’s agentic research system automates experiments and identifies backend configurations offering the best latency.
Risks & concerns
- A single large language model proved insufficient for Capital One’s complex fraud customer-service workflow.
- Fraud calls can last up to 60 minutes and previously required agents to reconstruct lengthy interactions manually.
- Combining two individually useful LLM optimizations can produce a performance regression.
- Proactive, event-driven agents will require rigorous testing and monitoring before deployment.
- Off-the-shelf frontier models lack the proprietary Capital One data the bank considers essential for differentiation.