There is a certain kind of AI company that demos beautifully and deploys never. The demo happens on a laptop with five bars of WiFi and a friendly cloud API waiting on the other end. Then someone asks the hard question - can it run in a windowless room on a network that has never touched the internet and never will - and the demo goes quiet. Legion Intelligence is built entirely around that quiet.
Legion, a San Francisco company of about 78 people, sells a governed agentic AI platform to customers who operate in exactly those windowless rooms: the U.S. Department of Defense, the Department of Energy, U.S. Special Operations Command, and a supporting cast of aerospace, manufacturing, and financial-services enterprises. The pitch is not that the model is smarter than everyone else's. The pitch is that the whole thing works where the others can't be installed, and that a human stays in charge of it the entire time.
From a yurt to a legion
The company was founded in 2022 as Yurts. A yurt is a portable, flexible shelter, which is a reasonable metaphor for a general-purpose AI sandbox you can pitch anywhere. By April 2025 the founders decided the metaphor no longer fit. They renamed the company Legion Intelligence - after the Roman military unit - to signal a shift from "flexible sandbox" to, in their words, a disciplined ecosystem of agents behind every individual and every mission.
This is the sort of rebrand that is easy to be cynical about, because renames are cheap and armies are evocative. But the repositioning was real. Yurts had spent its early years learning that the defense customer does not want an open-ended AI toy; it wants a system of record, an audit trail, and someone accountable when an agent does something. "Legion" is a more honest description of what the product became.
This milestone reflects the trust our investors and customers have placed in us.
What it actually does
The Legion Platform lets an organization deploy, coordinate, and govern networks of AI agents. The agents can search across a customer's systems without migrating the data, answer questions in natural language, draft reports and products, and - the part that matters most to Legion's customers - execute multi-step workflows. Crucially, it is model-agnostic: you can run open-source models, proprietary models, or your own custom models, and you are not locked into any single vendor. The platform is the constant; the models are swappable.
It is designed to run across the full spectrum of deployment environments - cloud, hybrid, on-prem, air-gapped, and classified - accredited from Impact Level 2 all the way up to IL-6, which is the level for SECRET material. It ships with more than 100 integrations to the boring, load-bearing enterprise systems: Salesforce, Oracle, SAP, Slack, GitHub, Jira. Every agent action is logged and gated. That governance layer is not a footnote; it is the reason a combatant command will let the software near its data at all.
The 94% number
Legion's most-cited result comes from U.S. Special Operations Command, where it deployed an application called SOFchat across the enterprise. The company reports that targeting officers generated intelligence summaries 94% faster, that SITREP processing dropped from about eight hours to under one - an 88% reduction - and that battalion staff saved an estimated 216 hours per week on repetitive manual work.
You should read numbers like these with the usual skepticism you'd apply to any vendor's own metrics. But note what they are measuring. Not tokens per second, not benchmark scores - staff hours. If you have never written a situation report at two in the morning, the eight-hours-to-one figure is abstract. If you have, it is the entire value proposition. Legion sells to people who do the work, and it measures itself in the currency those people actually care about.
Stop talking to AI. Start commanding agents.
The moat nobody wants to build
Here is the counterintuitive thing about defense AI. The model is arguably the least defensible part of the stack, because there are excellent models available to everyone and better ones every quarter. The defensible part is the unglamorous accreditation: FedRAMP High, SOC 2 Type II, HIPAA, CMMC, and the IL-2-through-IL-6 approvals that take years and armies of paperwork. Legion has done that work. Nearly a quarter of its staff hold security clearances, and the team is heavy on U.S. Armed Forces veterans - which is less a hiring quirk than a form of product-market fit. You cannot build for the mission without people who have lived it.
The company has also chosen the hardest deployment surface on purpose. Its edge product, Centurion, is a box - hardware, software, and sustainment together - built to keep agentic AI running in DDIL conditions: denied, degraded, intermittent, and limited connectivity. The insight is simple and slightly grim: at the tactical edge you cannot download a patch, so you ship the entire self-contained system and support it. That is a much harder business than an API, and a much stickier one.
The money and the company it keeps
Legion has raised roughly $56 million, anchored by a $40 million Series B in December 2024 led by XYZ Ventures, Glynn Capital, and Nava Ventures, with earlier backing from Bloomberg Beta, Mango Capital, and Essence Venture Capital. In 2025 it announced a partnership with Palantir to deliver generative AI to Special Operations Command, and it distributes to government buyers through Carahsoft. Its competitors are the names you'd expect in this corner of the market - Palantir's own AIP, Scale AI, Cohere's government efforts, Ask Sage, and the enterprise arms of the big model labs.
What makes Legion worth watching is not that it has the best model - it doesn't claim to. It's that it picked the market everyone else finds too slow, too regulated, and too hard to demo, and made the difficulty itself the product. Governance over autonomy. Accreditation over hype. A box that works when the internet doesn't. It is a deeply unfashionable set of bets, which is usually where the interesting companies are.