GOVERNANCE

Capital One builds its own multi-agent platform on open-weight models — validation stays between agents, never a human clicking

In an article published August 13, 2026 by VentureBeat — partner content funded by Capital One itself — Kel Vanee, the bank's head of machine learning engineering, explains why Capital One chose to customize open-weight models rather than buy a generic orchestration platform, and to build its own multi-agent harness, MACAW, for its bank-fraud calls. Every call passes through a chain of four specialized agents — understanding, reasoning, validation, explanation — but the article never once mentions a human click before a decision reaches a customer: only one AI checking another AI's work.

On August 13, 2026, VentureBeat published an interview conducted at its VB Transform 2026 conference with Kel Vanee, who leads machine learning engineering at Capital One, interviewed by Sam Witteveen. The angle of the interview is blunt: "At Capital One, we don't just use AI, we build it," Vanee sums up. Concretely, the bank made three deliberate architectural choices rather than buying an off-the-shelf solution: a centralized, enterprise-wide AI platform with governance built in from design, open-weight models (including Meta's Llama) fine-tuned on its own data rather than a generic frontier model, and its own multi-agent orchestration harness, named MACAW. "We see our data as a massive advantage that nobody else has, that generic frontier models can't provide. So we take that data and deeply customize these models," Vanee explains — also noting an unexpected side effect: training an open-weight model on the vocabulary and internal policies of one specific use case improves its performance across the bank's entire portfolio of use cases, not just the targeted one. MACAW illustrates the architecture on the most sensitive ground: bank-fraud calls, several million a year, ranging from four to sixty minutes long. A single large language model proved insufficient; the bank therefore split the work across four specialized agents that hand off on each call — an understanding agent that interprets the customer's intent, a reasoning agent that builds a structured summary, a validation agent that fact-checks that summary, and an explanation agent that formats it before passing it on. That document then lands in the hands of several hundred specialists in complex fraud calls, who no longer have to reconstruct the call history by hand.

The architecture doesn't stop at the call center. Chat Concierge, Capital One's conversational car-buying assistant aimed directly at customers, runs on the same customized version of Llama and the same division of labor: an agent that talks with the customer, an agent that builds an action plan from business rules, an agent that assesses the accuracy of the result, an agent that explains and validates it. Capital One applies this same logic internally too — an autonomous agentic system that tests combinations of backend infrastructure optimizations, runs the experiments in place of researchers, and hands them a summary of results, because two individually good optimizations can conflict once combined. Vanee anticipates two developments: routing across multiple models to balance cost and accuracy rather than betting on one, and a shift toward "proactive, event-driven" AI that acts as soon as it detects a condition, without waiting to be asked — a shift he himself describes as requiring "rigorous testing and monitoring," not something to take for granted. What the interview never clarifies is how a human concretely intervenes before a decision from these agents reaches a customer: MACAW's validation agent checks the accuracy of a summary, but that's one AI checking another AI, not an advisor approving an action before it goes out. For Chat Concierge, the wording is identical — an agent "assesses" and "validates" the result — without it being clear whether a human still sits somewhere in that loop before the assistant acts toward the customer. One detail that also matters: this article is partner content, funded by Capital One itself, published to coincide with its own conference — real quotes, a real event, but a story the company is telling about itself, not independent reporting.

There's something reassuring, reading this interview, in seeing a bank of this size arrive at the same conclusion AppH has defended from the start for French SMBs: a generic off-the-shelf model isn't enough, you need a platform built for your own business, with your own governance, not borrowed from someone else. It's no accident Capital One chose to customize open-weight models with its own data rather than rent someone else's intelligence — it's the same logic that pushed AppH to build a vertical per trade rather than one generic chatbot dressed up differently for each sector. But it's also necessary to name what this interview doesn't say, and to name it precisely because it's content Capital One itself funded to tell its own success story: the "validation" Vanee describes is an AI checking another AI, never explicitly a human clicking before a decision touches a customer. That may well be exactly what happens behind the scenes — the article doesn't contradict it, it simply never confirms it either. At AppH, this question doesn't need to be guessed at behind the scenes of a sponsored article: no action with a real consequence — a quote, a reminder, an accounting consolidation — goes out without a human clicking to approve it, logged, verifiable. Building your own harness of governed agents is real engineering progress, at the scale of a bank as much as an SMB; but governing by design and having a human approve remain two different things, and only the second really answers the question of who said yes before the action went out.

Reviewed by an AppH human