Why Capital One Is Building Its AI Advantage on Open Models and Multi-Agent Systems

Capital One’s approach to artificial intelligence offers a useful lesson for businesses navigating the gap between AI experimentation and enterprise-scale deployment. Instead of relying exclusively on a single, off-the-shelf frontier model, the bank is building a centralized AI platform that combines customized open-weight models, proprietary data, specialized agents, and enterprise governance.

The strategy reflects a broader shift in how organizations think about AI. The question is no longer simply which model produces the best answer. It is how to build a system that delivers accurate results, controls operating costs, meets latency requirements, and fits the rules of a complex business.

At VB Transform 2026, Kel Vanee, an MVP of machine learning engineering at Capital One, explained how the company is pursuing that goal. The bank’s experience spans fraud-related customer service, automotive shopping assistance, and research systems that optimize the infrastructure running AI workloads.

The underlying idea is straightforward but consequential. A competitive AI advantage may come less from owning the most powerful individual model and more from combining the right models, data, workflows, and controls into a reliable operating system for the business.

From Cloud Foundations to an Enterprise AI Platform

Capital One’s current AI strategy did not emerge overnight. According to Vanee, the company’s earlier investments in data transformation and cloud adoption created the technical foundation it needed to move quickly when generative AI became a major business opportunity.

That history matters. Deploying AI across a financial institution is not simply a matter of connecting an application to a language model. The organization must manage sensitive information, integrate existing systems, maintain operational reliability, and establish clear boundaries around what automated systems can do.

Capital One responded by making three connected architectural decisions.

First, it built a centralized, enterprise-wide AI platform with governance capabilities incorporated into the system. Second, it began customizing open-weight models using proprietary company data. Third, it developed its own multi-agent orchestration framework to divide complicated tasks among specialized AI components.

These decisions reinforce one another. The platform provides a common foundation for deployment and oversight. Customized models bring business-specific knowledge into the workflow. Agent orchestration coordinates the steps required to turn that knowledge into useful outcomes.

This architecture also gives the bank a way to reuse its investments across different applications instead of building every AI solution from scratch.

Why Proprietary Data Can Matter More Than Model Size

The AI industry often focuses on benchmark rankings, parameter counts, and announcements of increasingly capable foundation models. Capital One’s strategy places greater emphasis on the relationship between a model and the business using it.

Vanee described the bank’s proprietary data as a significant competitive advantage because general-purpose frontier models do not automatically possess the institution’s internal knowledge, policies, and terminology.

Capital One therefore fine-tunes open-weight models with its own data rather than relying exclusively on models in their original form.

The distinction is important. A general-purpose model may understand banking concepts, but that does not mean it knows how a particular bank handles a complex fraud investigation, documents a customer interaction, or applies an internal business rule.

Customization can help close that gap.

There is also a potential multiplier effect. Vanee explained that improvements made while training a model for one Capital One use case can benefit other applications across the company's portfolio. The model is not merely learning one narrow task. It is becoming more familiar with the organization's policies, terminology, and recurring business requirements.

That creates the possibility of reusable institutional expertise.

Real-time data adds another dimension. Historical training data cannot always provide the latest context for a live customer interaction. Bringing current information into the workflow can help the system respond to circumstances that have changed since the model was trained.

However, proprietary data is not automatically an advantage simply because it exists. Its value depends on its quality, relevance, accessibility, and governance. Capital One’s approach suggests that the competitive opportunity lies in combining company-specific information with a deployment architecture capable of using it appropriately.

Why One Large Language Model Was Not Enough for Fraud Support

One of the clearest examples of Capital One’s strategy comes from its consumer banking fraud customer-service workflow.

The system supports a process that handles millions of calls each year. Individual interactions can last approximately four minutes or extend to as long as an hour. These conversations may involve complicated explanations, multiple questions, and details that must be documented accurately.

Capital One initially explored using a single large language model for the task. According to Vanee, that approach proved insufficient.

The problem illustrates a limitation of treating AI as a single-step answer generator. Summarizing a long conversation requires several distinct capabilities: understanding the customer's intent, identifying relevant information, reasoning about the material, checking factual accuracy, and presenting the results in a usable format.

A single model might attempt all these steps at once, but a complicated workflow benefits from making the responsibilities explicit.

Capital One developed a multi-agent workflow called MACAW to address this challenge.

The workflow helps several hundred customer-service employees who specialize in complex fraud-related calls. Instead of reconstructing a lengthy conversation manually after it ends, employees receive a structured summary to support their work.

The business value is not limited to reducing typing. Better documentation can help employees review what happened, understand the customer's situation, and continue the process with more complete information.

Capital One's approach also illustrates why validation deserves its own place in enterprise AI architecture. Generating a plausible summary and producing an accurate summary are different objectives. Separating generation from checking makes that distinction explicit, although the validation process itself still needs to be tested and monitored.

How the Same Architecture Supports Automotive Shopping

Capital One also applies its multi-agent approach to Chat Concierge, a customer-facing automotive shopping assistant.

This application uses a customized version of Meta's open-weight Llama model alongside Capital One's proprietary data. Its workflow divides responsibilities among agents that communicate with the customer, develop an action plan based on business rules, evaluate accuracy, and explain and validate the result.

The underlying architecture resembles the fraud-support system, but the customer experience is different.

In fraud support, the primary goal is to turn a complex conversation into reliable documentation for an employee. In automotive shopping, the system helps a customer navigate a purchasing-related experience through an interactive conversation.

This is one of the strongest arguments for building a reusable AI platform. The applications do not need to share identical business logic to benefit from common architectural principles.

A company can reuse orchestration patterns, governance mechanisms, evaluation practices, and model-customization capabilities while adapting individual agents to specific workflows.

That creates a more scalable development model than treating every new AI application as an independent experiment.

An AI Research Agent That Tests Infrastructure Optimizations

Capital One's use of agentic AI extends beyond customer-facing applications. The company has also developed an autonomous system to help researchers optimize the backend infrastructure that hosts AI workloads.

The challenge is that performance improvements do not always combine neatly.

An individual optimization may improve latency in one configuration but perform poorly when combined with another optimization. Testing every possible combination manually can consume considerable engineering time.

Capital One's system addresses this by exploring a search space defined by a researcher. It handles the mechanics of configuring and running experiments, then summarizes the results for the researcher to review.

The system is designed to help identify combinations of optimizations and configurations that deliver the best latency.

This is a different kind of AI productivity. Rather than answering customer questions or generating content, the agent automates repetitive experimental work and helps engineers focus on interpreting results and making decisions.

It also demonstrates an important distinction between autonomous execution and unrestricted autonomy. The researcher defines the search space, and the system operates within that scope. Human expertise remains central to defining the objective and assessing the findings.

For organizations deploying AI at scale, this kind of internal automation can be valuable because infrastructure efficiency affects the economics of many downstream applications.

The Economics of Custom Models and Multi-Agent Workflows

Capital One's architecture reflects a broader enterprise calculation. The cost of an AI application is not determined by model pricing alone.

Organizations must also consider response latency, integration work, data access, validation, monitoring, governance, and the human effort required to correct errors or complete unfinished tasks.

A multi-agent workflow can introduce additional model calls and processing steps. That may increase raw inference costs compared with a simple single-model interaction. The approach makes economic sense when the additional structure improves reliability, automates meaningful work, or reduces expensive manual effort.

The goal is therefore not to maximize the number of agents. It is to assign each step a clear purpose and ensure that the combined workflow performs better against measurable business objectives.

Open-weight models offer another potential advantage: more flexibility to customize the model and control how it is deployed. Capital One's strategy demonstrates how that flexibility can be combined with proprietary data and enterprise orchestration.

But open-weight does not automatically mean cheaper or easier. Fine-tuning, evaluation, deployment, infrastructure, and ongoing maintenance all require investment. The economic case depends on the workload, available expertise, usage volume, and quality requirements.

Capital One's decision to build its own platform suggests that the company sees enough value in customization and control to justify that engineering effort.

Why Model Routing Could Become the Next Competitive Advantage

Looking ahead, Vanee identified model routing as an important development in enterprise AI.

Different models have different strengths. One may perform particularly well on a specific reasoning task, while another may be more suitable for a different request because of its cost, latency, or accuracy.

Instead of sending every request to the same model, a routing layer can select among multiple models according to the task and the platform's evaluation criteria.

Vanee argued that routing across a broader range of models could deliver better accuracy than relying on any individual model alone.

The logic is compelling, but the outcome depends on execution. A routing system needs reliable ways to classify requests, compare model performance, account for costs, and handle cases in which the initial selection is wrong. Routing also introduces operational complexity and additional monitoring requirements.

Still, the direction is significant. As enterprises gain access to more capable models, selecting the appropriate model for each task may become as important as improving any single model.

Capital One's existing emphasis on centralized infrastructure and governance provides a foundation for exploring this approach.

From Reactive Chatbots to Proactive, Event-Driven AI

Vanee also expects proactive, event-driven AI to become more important.

Traditional conversational AI generally waits for a person to submit a prompt. An event-driven system could instead monitor defined conditions and initiate an action when a relevant event occurs.

In a financial-services context, that could support broader monitoring and help identify situations that warrant attention, including potential fraud-related activity. The specific applications would depend on the signals available, the rules governing the workflow, and the safeguards surrounding automated decisions.

The potential benefit is a shift from responding to individual requests toward monitoring larger numbers of events continuously.

But proactive AI also raises the stakes. A system that waits for a user to ask a question has a naturally defined starting point. A system that initiates action must determine which events matter, when intervention is justified, and what it is authorized to do.

False positives, missed events, inappropriate escalation, and unintended actions can all create problems. Rigorous testing, monitoring, and governance are therefore essential, particularly in financial services.

The opportunity is substantial, but so is the responsibility. Proactive automation must be designed around clear permissions and measurable safeguards rather than treated as a simple extension of chatbot functionality.

The Competitive Landscape Is Shifting From Models to Systems

Capital One's strategy offers a different way to understand competition in enterprise AI.

Frontier-model providers compete to improve the general capabilities of their models. Open-weight model developers provide another route for organizations seeking flexibility and customization. Enterprise platforms focus on integrating AI into workflows, managing deployment, and controlling operational risk.

These layers overlap, but they are not interchangeable.

A bank may use an open-weight model as a foundation, customize it with internal information, and combine it with multiple specialized agents. It can then introduce routing, validation, and governance around the resulting system.

The differentiation emerges from the complete architecture rather than any one component.

This does not establish that Capital One's approach is universally superior to using a hosted frontier model. Some applications may be simple enough that a single model and a well-designed prompt are entirely adequate. Other workloads may benefit from managed services that reduce the burden of operating models internally.

The right choice depends on the task, security requirements, deployment constraints, performance targets, and total cost of ownership.

Capital One's experience instead demonstrates that organizations with substantial proprietary data and complex workflows have a reason to consider building beyond the default model interface.

The Risks Behind a More Sophisticated AI Stack

The benefits of a customized, multi-agent platform come with several important risks.

Model and data quality. Fine-tuning cannot compensate for every weakness in the underlying data. Incorrect, incomplete, or outdated information can undermine the entire workflow.

Validation limits. A validation agent is not an infallible source of truth. If the generating and validating components share similar weaknesses, they may fail in correlated ways. Independent checks and representative evaluations remain important.

Operational complexity. Multiple agents, model versions, routing rules, and dependencies create more components to monitor and maintain. A sophisticated workflow can become harder to debug if responsibilities are poorly defined.

Cost and latency. Additional agents may increase the number of model calls and the time required to complete a task. The architecture must demonstrate that its quality or productivity gains justify those costs.

Governance and accountability. Financial-services organizations need clear rules for data access, human escalation, logging, and permitted actions. Automation should not obscure who is responsible for a consequential decision.

Changing model capabilities. New models and optimizations arrive frequently. An architecture built around one model or configuration may become less competitive over time. Model routing and repeatable evaluation can help, but they also require ongoing engineering.

These are not reasons to avoid multi-agent AI. They are reasons to treat it as production infrastructure rather than a collection of experimental prompts.

What This Means for the Economics of Financial Services

Capital One's approach points toward a future in which AI creates value through the redesign of entire business processes.

In customer service, structured summaries can reduce the administrative work associated with long interactions. In automotive shopping, a conversational assistant can organize a more complex customer journey. In infrastructure research, autonomous experimentation can reduce the manual effort needed to evaluate optimization strategies.

These examples share a common pattern. AI is being applied to a defined process with a recognizable operational objective.

The economic impact depends on what happens after deployment. If a system produces summaries that employees must repeatedly correct, the expected productivity gain may disappear. If an agentic workflow introduces excessive latency or requires costly supervision, its business case weakens. Conversely, if the system reliably reduces repetitive work while maintaining quality, the value can extend across thousands or millions of interactions.

This makes measurement critical. Organizations should evaluate not just model accuracy but also completion rates, error severity, handling time, latency, inference costs, escalation frequency, and the amount of human correction required.

A successful AI platform should make those outcomes visible. It should help leaders understand which applications generate value, which require improvement, and which should not be deployed.

The broader lesson is that AI adoption is increasingly an operating-model decision. The organizations that benefit most may be those that connect technical capabilities to measurable changes in how work gets done.

What Business Leaders Should Do Next

For businesses considering an approach similar to Capital One's, several practical steps follow from its experience.

1. Start with a well-defined workflow. Identify a repetitive or complex process with measurable costs, quality requirements, and a clear business owner. Avoid introducing multiple agents merely because the technology makes them possible.

2. Evaluate proprietary data as a strategic asset. Determine which internal information improves outcomes beyond what a general-purpose model can deliver. Establish appropriate controls for data quality, access, freshness, and use.

3. Compare single-model and multi-agent designs. Build a baseline before adding orchestration. Measure whether specialized agents and validation steps improve end-to-end results enough to justify additional latency and expense.

4. Make governance part of the architecture. Define permissions, logging, escalation rules, and human review requirements before expanding deployment. Controls should be designed into the workflow rather than added after problems emerge.

5. Measure business outcomes, not demonstrations. Track real-world task completion, correction rates, employee time saved, customer experience, and total operating cost. Strong performance on a limited demonstration does not guarantee production value.

6. Prepare for multiple models. Keep evaluation and routing capabilities flexible enough to accommodate models with different strengths. Reassess choices as performance, cost, and business requirements change.

7. Introduce proactive agents cautiously. Begin with monitoring and recommendations before granting systems authority to take consequential actions. Expand permissions only after testing demonstrates that the safeguards are effective.

The objective is not to reproduce Capital One's entire stack. It is to adopt the principle behind it: build the smallest, most controllable AI system that can reliably improve a specific business process, then reuse the capabilities that prove valuable.

The Real Advantage Is the Architecture Around the Model

Capital One's AI strategy challenges the assumption that enterprise AI leadership comes primarily from selecting the most powerful foundation model available.

Its approach combines customized open-weight models, proprietary data, specialized agents, validation, and centralized governance. That combination supports applications ranging from fraud-call documentation to automotive shopping and infrastructure optimization.

The bank's next areas of interest—model routing and proactive, event-driven AI—extend the same philosophy. Instead of treating AI as an isolated conversational interface, Capital One is developing a platform capable of coordinating models and actions across different business needs.

The strategy is not without trade-offs. Customization requires expertise, multi-agent workflows introduce complexity, and autonomous systems demand rigorous oversight. No architecture eliminates the need to measure performance against real operating conditions.

Nevertheless, the central business lesson is clear. The value of enterprise AI increasingly depends on how well organizations connect models to their data, workflows, controls, and economic objectives.

Models provide capabilities. Platforms make those capabilities reusable. Governance makes deployment more manageable. And well-designed workflows turn technical potential into business value.

For financial services companies—and enterprises in other industries facing similarly complex operational requirements—that may prove to be a more durable advantage than any single model release.

Capital One AI, multi-agent AI, open-weight models, enterprise AI platform, banking AI, AI fraud detection, proprietary training data, AI agent orchestration, financial services AI, AI model routing, customer service automation, AI governance, event-driven AI, Llama model customization, AI infrastructure optimization