When most enterprises are still debating whether to dip a toe into generative AI, Capital One has already built — and deployed — a full-scale, production-grade multi-agent AI system. And they didn't do it by plugging in an off-the-shelf API from a frontier AI lab. Instead, they made a deliberate, strategic bet on open-weight models, deeply customised with proprietary data, orchestrated through their own purpose-built multi-agent harness.

The results are hard to argue with. This is the story of how one of America's largest banks built a repeatable blueprint for enterprise AI — and what technology leaders can learn from it right now.

Why Capital One Chose Open-Weight Models Over Closed APIs

The first and perhaps most consequential decision Capital One made was which kind of AI model to build on.

At VB Transform 2026, Kel Vanee, MVP of machine learning engineering at Capital One, explained how the bank built a scalable multi-agent AI architecture around deeply customised open-weight models rather than relying on an off-the-shelf foundation model.

The reasoning was strategic, not just technical.

Capital One used open-weight models rather than closed ones because that allowed them significant customisation — and that customisation is critical because competitive advantage in AI strategy relies on proprietary data.

As Naphade put it: "We do believe that the open weights and open models approach… it's the only way we get to deeply customize them."

This wasn't a naive decision either.

Open models do tend to carry security risks such as model tampering and data leakage — but Capital One isn't simply downloading open models off Hugging Face and plugging them in where they fit.

Every model is hardened, governed, and deeply adapted to Capital One's specific banking context before it ever touches production traffic.

The Foundational Investments That Made This Possible

Capital One's multi-agent AI success didn't happen overnight. It was the payoff from years of infrastructure investment that most banks haven't yet made.

The groundwork was laid years ago with Capital One's early investments in data transformation and cloud adoption, which Vanee said were foundational to moving quickly when the current wave of AI arrived.

That technical foundation enabled the company to make several deliberate architectural decisions, including building a centralised, enterprise-wide AI platform with built-in governance, deeply customising open models with proprietary data, and constructing its own multi-agent orchestration harness.

The company strategically leverages its cloud maturity, proprietary data, and partnerships with academic institutions to enable advanced AI research, agentic AI workflows, and multi-agent systems that drive operational efficiency and competitive advantage across its $39 billion business.

The lesson here for CIOs and technology leaders is clear: you cannot sprint to agentic AI without first having invested in clean, governed, discoverable data. The foundation matters enormously.

MACAW: Capital One's Multi-Agent Orchestration Architecture

The technical centrepiece of Capital One's AI strategy is a proprietary multi-agent orchestration system known as MACAW — their multi-agentic workflow platform.

With Capital One's MACAW, interactions are routed through specialised agents with governance and guardrails built in. The workflow is made up of a number of different agents. The first is an understanding agent, whose purpose is to look at what the customer is saying and try to understand their intention. From there, a reasoning agent generates a summary; a validation agent fact-checks the summary; and an explaining agent turns it into a formatted document shared with human agents.

This modular, specialised approach is a significant departure from the "one big model does everything" mentality. Each agent in the chain has a defined role, a constrained scope, and built-in accountability — which is exactly what regulated industries like banking demand.

In Capital One's multi-agent system, multiple AI agents work together to not only provide information to the customer, but to take specific actions based on their preferences and needs. For example, one agent communicates with the customer, while another creates an action plan based on business rules and the tools it is allowed to use.

Critically, Capital One also built in a dedicated evaluator agent — an AI whose sole job is to review what the other agents do, checking outputs against company policy and rules before anything reaches the customer.

As Naphade explained: "We thought that was a good idea for us, to have an AI agent whose entire job was to evaluate what the first two agents do based on Capital One policies and rules."

Chat Concierge: The First Production Deployment

Capital One's first publicly deployed multi-agent workflow is Chat Concierge, and its results are striking.

Chat Concierge was Capital One's first multi-agentic workflow, deployed through the company's auto business, and designed to support both auto dealers and customers with the car-buying process.

Chat Concierge is a multi-agent chat experience for auto dealerships that handles intent disambiguation, tool invocation, and human handoffs to deliver safer, more personalised customer journeys.

The business impact has been measurable.

With rich customer data, dealers are identifying serious leads, which has improved their customer engagement metrics significantly — up to 55% in some cases, with agents able to "generate much better serious leads through this natural, easier, 24/7 agent working for them."

Beyond the auto business,

Capital One is also leveraging agentic AI to automate rote tasks for employees and help them focus on high-leverage work — including an autonomous agentic optimisation solution to tune backend hosting infrastructure.

Governance, Guardrails, and Responsible AI at Scale

Deploying AI in a highly regulated financial services environment is not for the faint-hearted. Capital One has taken a deliberately cautious, governance-first approach to scaling its multi-agent systems.

Capital One's platform-centric approach separates design from runtime governance, embedding policies, guardrails, and cyber controls across agent threat boundaries.

Natarajan stressed that responsible AI is fundamental to Capital One's design process. His teams take a "responsibility through design" approach, implementing robust guardrails — both technological and human-in-the-loop — to ensure safe deployment.

As Naphade described it: "We have multiple iterations of experimentation, testing, evaluation, human-in-the-loop, all the right guardrails that need to happen before we can actually come into the market with something like this."

The team also invests heavily in the developer experience for agent builders, observability, and evaluations for stochastic multi-agent workflows, as well as strategies for model specialisation including fine-tuning and distillation.

On the data side,

Capital One sidesteps many reliability problems like hallucinations by severely limiting permitted queries to information that exists within the company's own data, as Natarajan notes: "We don't have to answer the world's curiosity for random information."

The Technology Stack: NVIDIA, Open Source, and In-House Tooling

Capital One's AI infrastructure is a carefully assembled hybrid — not a single vendor solution.

In terms of the technology stack, they use a combination of tools, including in-house technology, open-source tool chains, and NVIDIA inference stack. Working closely with NVIDIA has helped Capital One get the performance they need, and collaborate on industry-specific opportunities in NVIDIA's library, and prioritise features for the Triton server and their TensorRT LLM.

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.

As Vanee summarised: "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."

Practical Tips: What Enterprise Leaders Can Take Away Today

Capital One's journey offers a clear, actionable framework for any enterprise looking to move from AI experimentation to production-grade deployment:

Open-weight models allow significant customisation, which is critical when competitive advantage relies on proprietary data.

Capital One's platform-centric approach separates design from runtime governance, embedding policies, guardrails, and cyber controls across agent threat boundaries

— a model every enterprise should study.

Conclusion: A Blueprint Every Enterprise Should Be Studying

Capital One's multi-agent AI platform is more than an impressive technical achievement — it is a strategic proof point.

Their approach underscores a broader truth for enterprise technology leaders: driving measurable value with AI requires moving beyond off-the-shelf software toward deeply customised, highly governed architectures.

The bank didn't chase the latest API or bolt a chatbot onto legacy systems. It invested in data foundations, chose open-weight models for maximum control, built its own orchestration layer, and embedded governance at every layer of the stack. The payoff — 55% improvements in customer engagement and an AI platform scaling across the entire business — speaks for itself.

Whether you're a CTO, a head of AI, or an enterprise architect, Capital One's playbook is the most instructive real-world case study in production-ready multi-agent AI available today. If you're ready to begin your own journey from AI experimentation to enterprise deployment, start by auditing your data infrastructure, mapping your highest-value automation opportunities, and exploring open-weight models that can be fine-tuned on your proprietary assets. The blueprint exists — now it's time to build.