Conductor AI is automating the Pentagon's $18B classification and paperwork bureaucracy

Summary

Conductor AI is a startup built to automate the Pentagon's classification and document approval machinery — a system the government itself estimates costs $18 billion a year to run.

Zach, the founder, spent seven years at Palantir before starting the company. The core problem he saw firsthand: adjudicators responsible for security classification, foreign military sales approvals, and ITAR controls are expected to read and apply roughly 20,000 pages of policy across every case they review. The machinery exists, but it was never built to move fast.

Conductor's initial customers include the Air Force and the Office of the Secretary of Defense. The go-to-market has run through OTAs and subcontracts off existing vehicles — the company won one SBIR early on but hasn't won another since. The sales motion is opaque enough that Zach says a 23-year-old without prior government exposure would have little chance of figuring it out. That same opacity, he argues, is the moat: getting a foot in the door takes a long time, and trust compounds through internal word-of-mouth once you're in.

Model access is the near-term constraint. GPT-4.5 and Gemini 2.5 are capable enough to handle much of this work, but they can't yet be deployed on the classified networks where Conductor operates. The workaround is architectural — rather than building for real-time response, Conductor designs workflows where users submit a document and expect a reply in minutes rather than instantly. Against a baseline of weeks-long review cycles, a 10-minute turnaround is a step change. Against a live UI, it would feel slow. So the product deliberately frames around the former comparison.

Much of the near-term value isn't in resolving genuinely hard classification questions. It's in catching the obvious errors — forms filled out incorrectly, submissions that would be rejected on procedural grounds — before they enter a queue that takes four weeks and nine signatures to clear. Flag those early, and the substantive review can happen faster on requests that actually warrant scrutiny.

The physical constraints of classified networks are real friction. Conductor staff travel to secure facilities to upload forms and get software provisioned onto classified systems — not a metaphor for bureaucracy, but literal gate-keeping around where software can run. Zach treats this as the current state of the job rather than a problem that disappears soon.