The Most Expensive Word in Government Is ‘Why?’

The first useful thing to know about ConductorAI is that it began with a very small act of bureaucracy. Zachary Long, then working in national security, wanted to post notes from a meeting in Slack. The answer was no. He asked the security team why. They cited policy. He asked to see the policy. Nobody could produce it. A decision had been made, but the reason for it had evaporated somewhere between the rulebook and the person enforcing the rule.

This is funny until the blocked item is not a Slack message. Suppose it is a technical manual an ally needs, a document that might contain export-controlled information, evidence buried in a million-page investigation, or a disclosure request waiting on one qualified reviewer. The American administrative state is full of decisions that are both essential and hard to explain. The people making them are not necessarily slow. They are often searching for the one paragraph that gives them permission to move.

The wrong question was almost right

Long founded ConductorAI in 2023 with Benjamin Fichter and Eric Schwartz. Long and Schwartz had each spent seven years at Palantir; they met on a Department of Justice project involving large fraud prosecutions. Fichter, a college connection of Long’s, helped originate the company idea. Their experience gave them an unusual view of institutional software: the database is rarely the whole problem. The harder problem is the unwritten sequence by which an expert turns evidence, policy, precedent, and caution into a defensible yes or no.

The founders’ first thesis was to help people classify a document correctly. It was directionally sound and commercially awkward. Few users begin a working day eager to consult a classification guide in the abstract. Pain arrives at the boundary. Someone wants to send the file to a foreign partner. A deadline appears. The question changes from “What is this?” to “Can I share this?”

That distinction changed the company’s pitch. ConductorAI began selling at what Schwartz calls the friction point - the transfer, release, approval, or review that cannot proceed until policy has been translated into action. It is a useful product lesson because it is so ordinary. A capability is not yet a market. A blocked handoff often is.

Policy, but executable

The product is called Conduit. It ingests policy documents, previous approvals, emails, office files, images, and other institutional material. Experts then encode a review as a sequence of AI-assisted decisions. One step may identify the relevant jurisdiction; another checks a sanctioned-party list; another marks text for redaction; another asks a human reviewer to resolve an ambiguity. The platform can run these sequences on demand or through an API.

For an ITAR or EAR review, that means more than asking a language model whether a passage looks sensitive. ConductorAI describes model-agnostic agents trained around the regulations and an organization’s historical patterns. The system suggests whether content is controlled, identifies the relevant jurisdiction, highlights the exact sections that deserve human attention, and returns a document-level recommendation with a rationale. Each action is logged.

That last feature is not clerical garnish. In high-stakes work, the answer and the provenance are a pair. A clever model that says “no” without pointing to the governing language has created a new review problem. Conduit’s more interesting promise is to show the math - down to a page and passage - so a reviewer can defend the decision to a supervisor, an auditor, or an allied government.

The offices nobody puts in the movie

ConductorAI’s customers sit in defense, intelligence, civilian government, and the industrial companies that build alongside them. The use cases include foreign disclosure, partner information sharing, export licensing, security classification, FOIA review, e-discovery, and investigations. These are consequential offices with spectacularly uncinematic work.

On the Crossing the Valley podcast, Long and Schwartz offered two numbers that explain the market better than any total-addressable-market slide. At Nellis Air Force Base, they said, one person handles a technical-manual transfer review process spanning 500,000 pages. Elsewhere, one GS-13 adjudicates roughly $40 billion in export-license approvals a year. Rotational staff can require six months of training for a twelve-month seat. The scarce asset is not another PDF. It is experienced judgment.

The company says those exercises reduced reviews that normally consumed a day to about an hour. Its current website claims policy approvals can run ten times faster. In August 2026, ConductorAI announced a contract with U.S. Army Cyber Command through the Army Cyber Technology and Innovation Center, focused on accelerating the release of sensitive information to partners and the public.

A machine that knows when to show its work

ConductorAI occupies an odd corner of the software market. It overlaps enterprise search, e-discovery, governance and compliance tools, workflow automation, and the custom systems built by large government integrators. Its differentiator is not simply “agents.” Plenty of vendors have agents. The harder bundle is domain knowledge, page-level citation, repeatable approval logic, and deployment where the documents actually live.

Conduit is advertised as FedRAMP High accredited and deployable across NIPR at IL-5, SIPR at IL-6, and JWICS. It can run in cloud-native and on-premises environments. Those acronyms are part of the product. An approval tool that cannot enter the secure network merely relocates the bottleneck.

The same is true of procurement. In 2025 ConductorAI partnered with Carahsoft, which made the platform available through reseller channels and government vehicles including SEWP V and ITES-SW2. Booz Allen Ventures also made a strategic investment. In public-sector software, being useful and being buyable are separate engineering problems.

There is a natural boundary to the approach. Conduit works best when an organization has real governing material, knowledgeable experts, recurring decisions, and enough historical practice to encode. If the policy contradicts itself, the source material is poor, or nobody owns the judgment, software cannot manufacture institutional clarity. The human-in-the-loop is therefore less a reassuring slogan than an operating requirement. Automation handles the search and the repeatable steps; accountable people handle the exception.

What the rest of us can steal

ConductorAI’s playbook travels better than its classified deployment stack. Any organization with a long-lived rulebook eventually develops a second, invisible rulebook made of precedents, email threads, and the memory of the person everyone calls. The opportunity is not to dump all of that into a chatbot. It is to find one recurring decision and map how a trusted expert actually makes it.

The government is not slow because nobody has invented a faster text generator. It is slow because action depends on permission, permission depends on interpretation, and interpretation is trapped inside documents and people. ConductorAI has raised serious money to make that chain legible: the Air Force’s $1.07 million Phase II award in 2023, then a $15 million Series A in 2025. The cost bought the company a chance to prove that the least glamorous layer in government might be one of the most valuable.

Long’s original Slack message is still a useful image. Somewhere, a person knew the answer was no. Somewhere else, a document may have explained why. The distance between those two facts was the whole problem. ConductorAI is building in that distance.