Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
Most large enterprises approach AI the same way they approach enterprise software: buy a capable product, configure it, and deploy. Capital One has taken a fundamentally different path, and the architecture it has built offers a revealing look at what serious, production-grade AI actually requires in a regulated, high-stakes industry.
At VB Transform 2026, Kel Vanee, MVP of machine learning engineering at Capital One, laid out how the bank constructed a multi-agent AI platform built around deeply customized open-weight models rather than frontier models sourced from OpenAI, Anthropic, or Google. The distinction is not cosmetic. It reflects a deliberate bet that proprietary data, not raw model capability, is the real competitive moat in financial services.
Proprietary Data as the Real Differentiator
The core of Capital One’s strategy is fine-tuning open-weight models on data that no external provider can replicate. “We view our data as a huge advantage and something that nobody else has, something that the general frontier models cannot provide,” Vanee said. “So we are taking that data and deeply customizing these models.”
This matters because frontier models, trained on broad internet-scale data, are generalists. They carry no institutional knowledge of Capital One’s internal policies, product nomenclature, or the specific patterns that distinguish legitimate customer behaviour from fraud. Fine-tuning on proprietary data closes that gap, and Vanee noted a compounding benefit that was not initially anticipated: improvements made for one use case produce measurable gains across the bank’s broader AI portfolio. “As we customize those open-source models for one use case, we actually see benefits across our whole portfolio,” he said. “We are training that model to be an expert at Capital One use cases, policy, and nomenclature. As we do that training, we see a general lift.”
The bank’s early investments in cloud infrastructure and data transformation, made well before the current generative AI wave, are what made this speed of execution possible. Without that foundation, fine-tuning at enterprise scale would have been far slower and more expensive.
How MACAW Handles Millions of Fraud Calls
The clearest illustration of Capital One’s approach is MACAW, its multi-agent customer workflow system, deployed for bank fraud calls that number in the millions annually. Individual interactions range from roughly four minutes to as long as sixty minutes. When the bank initially tried handling these calls with a single large language model, the results were insufficient. The complexity and variability of fraud conversations simply exceeded what a single model, responding to a single prompt, could manage reliably.
MACAW solves this by decomposing the task across specialised agents, each with a narrow, well-defined role. An understanding agent first interprets customer intent. A reasoning agent then generates a structured summary based on specific instructions. A validation agent fact-checks that summary for accuracy. Finally, an explaining agent formats the output into a document that is passed to human customer-service staff.
The practical effect is significant. Several hundred agents who handle complex fraud calls previously had to reconstruct long, back-and-forth conversations manually after each call. MACAW generates post-call summaries automatically, freeing those agents to focus on judgment-intensive work rather than documentation.
The same architectural logic underpins Chat Concierge, Capital One’s customer-facing auto-shopping assistant. That product runs on a customised version of Meta’s open-weight Llama model and uses an identical division of labour: one agent conversing with the customer, one building an action plan from business rules, one evaluating accuracy, and one validating and explaining the result.
Agentic AI Turned Inward: Optimising the Infrastructure Itself
Beyond customer-facing products, Capital One has applied the same agentic logic to internal engineering problems. One example is an autonomous system built to tune backend hosting infrastructure, a task that is tedious, technically demanding, and consequential for latency and cost.
Vanee explained that in the world of large language model deployment, new optimisations arrive constantly, and they do not always combine cleanly. Two individually beneficial optimisations can, when applied together, produce a performance regression. The agentic system navigates this by running through a search space defined by the researcher, handling all the mechanics of setting up and executing experiments, and then presenting a full summary of results. “This agentic system will run through a search space that is designed by the researcher, handle all the mechanics of setting up that experiment and running the experiment, and then put a whole summarization of the results in front of the researcher,” Vanee said.
The result is that researchers can identify the optimal configuration for latency without spending their time on the mechanical work of running trials. It is a meaningful example of agentic AI augmenting expert human judgment rather than replacing it.
What Comes Next: Routing Across Models and Proactive AI
Vanee identified two trends he expects to define the next phase of enterprise AI. The first is model routing, where an abstraction layer evaluates multiple models dynamically and selects the best one for a given query. “We actually think that you can get better accuracy than any individual model simply by routing across a broader set of available models, because different models are going to excel in different areas,” he said. This approach treats the model landscape as a portfolio rather than a single dependency, which also provides cost flexibility as pricing across providers shifts.
The second trend is proactive, event-driven AI. Rather than waiting for a human to initiate an interaction, these systems would detect conditions that warrant action and respond autonomously. Vanee was clear that this capability demands rigorous testing and monitoring before deployment, but he sees it as particularly powerful for fraud detection and large-scale monitoring. “Proactive AI is going to be a really important trend,” he said.
For technology and AI leaders across financial services in Southeast Asia, including institutions operating under the oversight of Malaysia’s Securities Commission, Bank Negara Malaysia, or Singapore’s Monetary Authority, Capital One’s architecture offers a concrete and instructive reference point. The bank’s experience suggests that the institutions most likely to extract durable value from AI are not those that adopt the most capable off-the-shelf model, but those that invest in the governance infrastructure, proprietary data pipelines, and multi-agent orchestration needed to make AI reliable enough to trust with millions of customer interactions.
Read More: Benchmark scores lied about Qwen 3.8-Max. Here is what actually predicts your AI bill
