A good AI demo begins with a clean prompt and ends with a small miracle. Ben Van Roo's working life has happened after that ending. The answer must enter an old database. The user needs permission to see one document but not another. The model has to run in a data center, on a laptop, or inside a network with no path to the public cloud. Someone must be able to inspect what happened and decide whether to approve it. At 2 a.m., when the connection disappears, the work cannot.
This is the less cinematic side of artificial intelligence, and Van Roo keeps returning to it. He is the co-founder and CEO of Legion Intelligence, the San Francisco company first called Yurts. Legion builds a layer between models and the daily systems of defense, government, and large enterprises. Its premise is almost stubbornly practical: intelligence is useful only when it survives contact with the organization.
The idea looks less like a sudden founder revelation than the accumulated residue of his career. Van Roo has spent years moving between mathematics and messy institutions, between people who design systems and people who must live inside them. Each stop made the same point in a different accent. Reality does not arrive in a tidy schema.
Origin systemThe detour into mathematics
Van Roo was born in Madison, Wisconsin, and grew up in a family shaped by military service. He has joked that bad eyesight prevented him from following the Air Force route around him. So he geeked out on mathematics instead. The alternative path kept him close to the same questions - logistics, readiness, decisions under constraint - but gave him a different instrument for answering them.
One campus, four lenses
2001
2001
2003
2007
At the University of Wisconsin-Madison, he collected ways of seeing. Computer science explained what machines could do. Environmental engineering put technical systems in physical context. An MBA supplied the organizational and economic frame. A PhD in operations research trained him to turn crowded decisions into models. He later described the value of that range plainly: he had worked with better mathematicians, computer scientists, and researchers, but could get his head around complexity by being a little of all three.
“There is a lot of value in being able to get your head around complexity by being all of those things.”Ben Van Roo
Graduate school led into RAND Corporation, first while he studied and then for three years after the doctorate. There, abstractions acquired geography. Van Roo worked on civilian and military problems and traveled to installations and operations in Afghanistan, Iraq, Kuwait, Qatar, Korea, Japan, and around the United States. His published work ranged across Air Force repair networks, modernization road maps, freight transportation, and service-parts systems.
A mathematical model might describe a repair network. The hard part was everything the model did not naturally hold: infrastructure, literacy, local politics, behavior, and the cost of being wrong. Operations research taught him optimization. Fieldwork taught him humility about the boundary of the optimization.
Change of tempoFrom policy time to startup time
By 2010, Van Roo wanted a faster feedback loop. He moved to San Francisco and joined the software world. At Chegg, he worked across supply chain, analytics, machine learning, and data science as the education company grew and entered public markets. Early natural-language processing was already appearing in the work around 2012 and 2013. It was rudimentary beside today's systems, but the direction was legible.
- RAND CorporationModels meet military logistics and field conditions.
- CheggData science meets growth, operations, and public-company accountability.
- Primer AILanguage models meet national-security users.
- YurtsA company starts before ChatGPT makes the category obvious.
- Legion IntelligenceThe mission expands toward governed teams of agents.
In 2017 he left the scale of a public company for Primer AI, then a small startup working on natural-language processing. He built and led its national-security division. The models improved quickly. Open-source systems that once seemed academic became surprisingly capable. Yet users still had to step away from the six or seven applications involved in their jobs, ask the AI for something, and manually carry its answer back.
That gap stayed with him. The industry was concentrating on the model while the human experience sat to one side. A clever system could perform a task. It did not necessarily change how a day of work moved. Van Roo's question became less about what an AI could say and more about where its answer could go.
The work after the answer
Legion's opportunity lives in the connective tissue: the data, controls, integrations, and approvals between a model and an outcome.
Founding thesisThe company before the craze
Van Roo pulled together engineers he had worked with before and co-founded Yurts in 2022. This was before ChatGPT's public release turned generative AI into a mainstream fascination. The timing matters because it clarifies the bet. Yurts was not formed to wrap a newly famous chatbot. It was formed around the enterprise problem Van Roo had already watched: models were getting better, while everything required to make them usable remained difficult.
The company designed for awkward places. Its platform could be deployed in a public cloud, a private data center, on bare metal, at the edge, or in an air-gapped environment. Customers could use commercial, open-source, custom, or government models. Controls governed who could reach which data. Retrieval systems showed the material behind an answer. Human reviewers could remain inside the loop.
One workflow, four operating conditions
This architecture sounds technical because the problem is technical. It is also a philosophy of customer power. Models will change. Vendors will change their prices and policies. A model-agnostic layer lets the workflow persist while the intelligence underneath it rotates. In Van Roo's framing, the durable asset is not a conversation with one model. It is the organization's ability to connect many models to its records and rules without surrendering control.
The team brought the software to military exercises and secure facilities, often with its own hardware. Van Roo summarized the attitude plainly: any exercise, any event, anywhere they could go, they would be there. This is more than founder theater. In a complicated environment, watching a failure gives better product information than receiving a polished feature request. The dropped network, slow approval, and strange legacy interface become architectural requirements.
“The challenge is no longer about demos and pilots; it's about delivering tangible results.”Ben Van Roo
Capital and consequenceWhat the $40 million buys
In December 2024, Yurts announced a $40 million Series B led by XYZ Venture Capital, with Glynn Capital, Nava Ventures, Bloomberg Beta, and Mango Capital participating. The round brought reported total investment to $58 million. By then the company said it had deployed generative AI on Department of Energy bare-metal infrastructure and on Defense Department networks, including a secret-level environment.
December 2024
investment
ChatGPT launched
The funding story is easy to compress into numbers. The operating story resists compression. A system used around sensitive information must respect access rules before it can be helpful. A generated summary must point back to records if a user is expected to defend it. An agent that can act inside a workflow needs an identity, a scope, an audit trail, and moments where a person says yes. These are not decorative safety features. They are the conditions under which the software is permitted to exist.
The company changed its name to Legion Intelligence in April 2025. Yurts suggested portable shelter. Legion suggested coordinated action. Van Roo and co-founder Jason Schnitzer described an ecosystem of agents working alongside decision-makers, moving across CPUs and GPUs, models, tools, and environments. The visual identity used the Roman numeral for 5,000. Beneath the symbolism was a product shift from a secure AI platform toward orchestrated workflows with explicit human control.
The durable ideaJudgment stays in the room
Van Roo's recent writing extends the argument from individual productivity to collective work. Giving every staff officer an assistant may increase the amount each person can produce. It does not guarantee that eleven people agree on the same figures under time pressure. The harder problem is shared context: versioned artifacts, attributable actions, contradictions surfaced early, and agents governed by the same decision process as the humans beside them.
This keeps the person in his technical story. Legion's language emphasizes human-led AI, but Van Roo's career makes the commitment more concrete than a slogan. RAND showed him how much context sits beyond a model. Chegg taught him about scale and operating accountability. Primer revealed the distance between capable NLP and ordinary work. Legion is an attempt to compress that distance without pretending it can eliminate judgment.
He remains connected to Wisconsin through university boards and alumni work. He has called himself a small-town kid who got lucky at an institution where people helped him. The line is warmer than his usual talk of infrastructure, yet it expresses the same systems instinct. Outcomes come from networks: mentors, teams, institutions, tools, constraints, and the bridges among them.
The AI industry is fond of the frontier, the place where a benchmark moves and a new capability appears. Van Roo's attention is trained a few miles behind it, where the supply lines run. His company may swap one model for another, but it cannot swap away the database, the permission, the approval, the disconnected laptop, or the person responsible for the call.
That is the quiet argument inside Legion. A model's intelligence is temporary leverage. An organization's ability to put intelligence to work - securely, repeatedly, and with someone accountable - is infrastructure. The miracle is only the opening scene. Van Roo is building for everything that follows.