The hours soldiers spend feeding systems that already know most of what they ask.
Annual mandatory training, per soldier. The 2018 count is the last one anybody published. In 2025 the Army rewrote the training regulation, cut it from more than 250 pages to 132, and left about sixteen requirements standing. Nobody has said what those sixteen cost a soldier in days.
Start with the working demo → Two minutes. It drafts a real Army document and makes you check it.
The pile has two kinds of work in it, and cutting only fixes one kind. Mix them up and "cut the paperwork" quietly turns into "cut the counseling," and now everybody is worse off.
Cybersecurity refreshers, harassment prevention, the off-duty driving assessment. Each one showed up after something went wrong, and whether it still earns its hours is a fair argument.
Our levers here are keep it or cut it. If the training is worth doing, it's butts in seats, and no model can sit in the seat for you.
Citations, counselings, evaluation support forms, monthly readiness reports. A necessary evil, heavy on the necessary: recognizing a soldier and counseling one are things we should be doing more of.
What's on the chopping block here is the drafting hour. The requirement stays. Every demo on this page lives in this column, and nothing here argues for recognizing fewer soldiers.
Same four questions for every task. Read across them and you can tell which ones are ready to go and which ones are somebody's wish.
It is on the page because soldiers nominate it first and we cannot do it. I left it in because pulling it would look like I stacked the deck, which I would have.
The Army's own Enterprise LLM Workspace rewrote 300,000 personnel descriptions in about a week, work the service put at roughly 50,000 hours of review. That one already happened, on this exact kind of document.
The rest are my estimates, built from the drafting times on the cards above. The Army doesn't publish annual totals for any of these documents, so everything below is per 100,000. Scale it to your formation.
These only hold if verification stays real, about ten minutes a document, fact by fact. They also assume nobody backfills the time. If the S3 fills it in by Monday, nothing came back.
Every guard below costs somebody something. I would rather say that part myself than have a sergeant major say it for me. Click one to open it.
What goes wrong. A form that takes ten seconds instead of ten minutes is a form a worried staff can add three more of. Soldiers stay just as buried, holding a faster shovel.
The guard. Any drafting tool ships coupled to a cap or a sunset review on the requirements it serves. Somebody has to own that review and defend the deletions, and nobody gets promoted for it.
We ran this experiment already. DTS automated the paper voucher three decades ago without asking whether the voucher was the problem. It hardened around its own workflow and outlived a $374M attempt to replace it.
What goes wrong. A model drafts a readiness figure that is not true, or summarizes a counseling in a way that shifts what a leader meant. The output looks right, which is why it gets signed.
The guard. Whoever signs it owns it. That is already the rule and it should stay the rule. But if you own it and nobody gives you the twenty minutes to actually read it, congratulations, you own the mistake.
Members of Congress are tracking error rates in the VA drafting pilot right now, and they should be.
What goes wrong. Sign an unread draft and you have told the same lie Wong and Gerras documented, just faster. This is the one that would sink the whole idea.
We have all clicked "I acknowledge" without reading it. Do not build another one of those.
The guard. Verification has to cost something. A fact-by-fact check instead of one click, plus drafts sampled against the records they cite, the way finance already spot-checks vouchers. That is friction on purpose: ten minutes a leader will want to skip at 1900 on a Friday.
People who monitor reliable automation stop checking it. The demo plants one deliberate error so you can watch this step catch something.
What goes wrong. We claw back a week of a soldier’s year and fill it with new taskings by Monday. The time was returned on paper and spent before anybody saw it.
The guard. Say where the hours go before the tool ships. Rehearsal, maintenance, sleep, the counseling conversation itself. A commander then has to defend empty space on a calendar, and every pressure in the building runs the other way.
Everything factual on this page traces to one of these. The estimates are labeled as mine.