On August 31 the Defense Department switched on two more frontier models inside GenAI.mil, its central generative-AI portal: OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government, joining Google’s Gemini. The number the department released alongside them is the one worth keeping. GenAI.mil has onboarded more than 1.7 million unique users out of 3 million personnel.
That is 57% of the workforce of the largest employer in the United States, disclosed by the employer, with a denominator attached.
What the department actually said
GenAI.mil launched last year with Gemini as its only model. The Pentagon CTO put usage at 1.5 million people on June 12; the department had already reported personnel building roughly 100,000 custom agents on the platform by April. Both new models cleared Impact Level 5, the highest authorization for non-public sensitive unclassified data. Personnel currently reach GPT-5.4 Terra, with 5.6 Terra sequenced next.
Then the department described the work. ChatGPT Mil, per DOD, supports “document-heavy, routine unclassified work such as administrative tasks, logistics, planning, and policy,” built around chat, files, projects and custom GPTs. A person familiar with the OpenAI side told DefenseScoop the practical uses look like filling out travel forms and responding to contracting opportunities.
The Grok release is more specific still. DOD said it would let the military “execute missions faster and with greater precision across numerous operational contexts, ranging from market research analysis for acquisition professionals to supply chain management for logisticians.”
Acquisition professionals. Logisticians. Those are not capability categories. Those are federal job series.
Anthropic’s Claude is absent from all of this. The Trump administration designated the company a supply-chain risk after it refused unrestricted military use; a federal judge ruled that designation unlawful last week and the White House said it will appeal.
Two halves of a measurement, five days apart
Compare the shape of this disclosure to the one we covered on August 26, when Salesforce reported 3.2 billion agentic work units and never said whose work those units used to be. Salesforce published volume with no denominator. The Pentagon has published a denominator with no volume: 1.7 million people are logged in, and nothing states how many hours of travel-voucher processing, market research or policy drafting moved off human desks.
Neither organization is hiding anything in particular. Nobody has the instrumentation. Adoption is trivial to count because the login table counts it. Displaced task-hours are hard to count because no system was ever built to measure them, and the first organization to publish that figure will have created a liability, not a metric.
So 1.7 million is the largest verified single-employer genAI adoption figure on record, and it is also the least informative kind of number: it says the tools are in the building. It does not say what left.
The cohort point is worth holding onto. The largest published private-sector Copilot and enterprise-assistant deployments this year have been reported as seat counts sold, not as verified unique users against a full-workforce denominator, and vendors have no incentive to publish the second figure because it is where the adoption story usually thins out. DOD had no such incentive either. It published anyway, because the number is a procurement success story on its face. That accident is why this is the cleanest penetration figure available anywhere, and why it is worth anchoring future claims against: when a vendor next says an enterprise “rolled out AI company-wide,” 57% of headcount in under a year is now the benchmark that phrase has to beat.
The exposure is real and it will not arrive as a layoff
The mechanism here is different from the private-sector stories in this corpus, and the difference matters for anyone reading this from a federal desk.
Start with which work is exposed. Stanford’s revised Canaries paper in August widened the young-worker employment gap to 19% and, more usefully, named the mechanism: the jobs falling are the ones built on written-down knowledge. Federal administrative work is the most codified work in the American economy. Travel vouchers follow the Joint Travel Regulations. Contracting responses follow the FAR. Policy memos follow a format manual. Every one of those is a documented procedure with a documented output, which is exactly the profile the Stanford revision identified as first to go, and exactly the list DOD published.
Now the timing. Federal civilian employment does not contract through WARN notices and press releases. It contracts through hiring freezes, attrition and unbackfilled vacancies, which produce no announcement and no dateable event. A GS-0300 administrative slot that quietly stops being reposted is invisible in every layoff tracker we cite. The signal to watch instead is the federal job postings series for the 0300 clerical-administrative family, the 1102 contract specialist series and the 0346 logistics management series, measured against separations. That gap is where this shows up, and it will show up before anyone announces anything.
The last piece is portability. Around 950,000 of the department’s personnel are civilians, and a large share of what they do has a direct private-sector twin: procurement, scheduling, records, logistics coordination. A contract specialist who has spent two years running acquisition market research through Grok is learning a workflow that transfers. One who has spent two years being routed around by it is not. The people inside GenAI.mil right now are, without being told so, sorting themselves into those two groups.
What we would want to see next
One number would change the read: task-hours, not users. Until some organization publishes hours-displaced against a headcount denominator, adoption figures like 1.7 million tell us the direction and nothing about the rate. We flagged the same gap when IBM stood up an OpenAI practice and counted certified consultants rather than the work they replaced. The Pentagon, uniquely, has the instrumentation and the mandate to produce that figure. It has not.
Sources: TechCrunch, DefenseScoop
