The Checklist Tax: The First Problem for AI in the US Army
Abstract: Mandatory training in the US Army grew from about five days per year in the 1980s to more than forty days by 2018, to which the Army’s answer was to cut requirements outright. Most of the surviving administrative work – e.g., soldier counseling, award citations, readiness reporting, evaluation support documents – consists of rote drafting that a generative artificial intelligence (AI) model is perfectly suited for. With AI models already accredited and fielded by the Department of Defense, the Army can keep these requirements and give working hours back to soldiers. Drawing on Leonard Wong and Stephen Gerras’s Lying to Ourselves, this “checklist” quietly erodes honesty. Handing this rote work to an AI model – with human verification – is one way to stop the Army from spending its integrity on paperwork.
As a companion to this article, the author built the Checklist Tax site, which includes a working demo for how the military could improve its award citation process.
Somewhere on an Army post this week, a captain clicks through an online training module they have finished at least three times prior in the past year. They know the information. What they need is the certificate at the end – a name, a date, and a tracker somewhere flipping from red to green – so they click until they get it. On the corner of their desk sits the range packet they owe to the battalion commander by Friday as a soldier waits outside their office with a pay problem that has been shorting their family for two months. They close the laptop and pick up the range packet.
Multiply this hour across a few dozen requirements and more than a million people and it becomes the “checklist tax”: the hours Army leaders spend feeding systems that already know most of what they ask. It is one of the largest and least examined costs the Army carries. After six years in uniform, I never met an officer who believed these modules made a unit more ready or more honest. We completed them because a tracker demanded it, and we got good at doing them fast.
A Bill Measured in Days
A traditional National Guardsman gets roughly 39 paid training days in a year; the required slate could swallow the entire training calendar before a soldier ever touched their rifle.
There are numbers for this going back decades. Officers in the 1980s remember a mandatory training slate of about five days a year. By 2018, an Army War College count put it above forty. That remains the most recent public figure. A traditional National Guardsman gets roughly 39 paid training days in a year; the required slate could swallow the entire training calendar before a soldier ever touched their rifle.
The Army knows. In 2024, it cut what the service described as roughly 350 hours of online coursework and scrapped the Structured Self-Development program. In April 2025, Chief of Staff General Randy George dropped nearly a dozen more mandatory programs, made six others discretionary for commanders, and shrank the training regulation from more than 250 pages to 132, telling commanders to put the recovered time into warfighting. Command Sergeant Major Chris Mullinax, the top NCO in the operations and training directorate, described the intent in plainer terms than the regulation did: “There’s no distractions, there’s no burdens and our war fighters are focused on war fighting.” Around sixteen requirements survived, but the Army has not published the cost to soldiers in terms of working days. Deletion recovers real hours and the instinct behind it is sound. It also runs out at the edge of the requirements the Army is willing to lose.
Whose Hours Are They?
The latest Quadrennial Review of Military Compensation found enlisted pay above the 80th percentile of comparable civilian earnings. As such, the waste hides not in the paycheck but in the hours. Pay is annual and fixed, meaning the marginal hour of a soldier’s time costs the requirement writer nothing; a staff section in the Pentagon can levy an hour against a million people and never see a bill.
Those hours land hardest on the people the Army can least afford to lose. The captain or staff sergeant with real options outside the gate serves at a discount no pay chart records and accepts it for reasons that have little to do with money. Wasting their time taxes the people with the most alternatives. The institution keeps no account of the cost.
Why Requirements Pile Up
Most of the requirements exist for a reason. Annual cybersecurity training came into effect after a breach. Harassment prevention education followed a scandal. The off-duty driving assessment exists because of a steady toll of soldiers killed in privately-owned cars on their own time. Cutting a requirement recovers the hours with the understanding that it has effectively mitigated the risk it targets. Every so often, it does. Usually, deletion is simply the only lever anyone has, because keeping the requirement meant paying for it in people’s time, something that was never funded.
The Cost of the Checklist
A young officer who learns to click “complete” on training they skipped is rehearsing a small dishonesty. The Army will one day need that same officer to face a larger one: a readiness number a commander wants rounded up, a sworn statement for an investigation, a report from a patrol that never left the wire.
Saving soldiers’ time is the easy case. The harder cost to the checklist is to honesty, something that shows up in the Army’s own research. In 2015, Leonard Wong and Stephen Gerras of the Army War College published a study with a blunt title: Lying to Ourselves. They documented that the volume of mandatory requirements had outrun the available hours so badly that officers routinely certified tasks they had not done – training conducted, briefings delivered, equipment inspected – and had grown comfortable doing it. The officers signing off were conscientious people who had been handed more compliance than any calendar could hold, and signing the form was the rational way to get through the week.
The damage goes deeper than wasted time. Each requirement the force cannot honestly satisfy teaches the people in it that a signature is a formality. That lesson does not stay confined to trivial training modules. A young officer who learns to click “complete” on training they skipped is rehearsing a small dishonesty. The Army will one day need that same officer to face a larger one: a readiness number a commander wants rounded up, a sworn statement for an investigation, a report from a patrol that never left the wire.
Automation stops at drafting administrative work, however. It cannot take the captain’s module for them, and it cannot complete skipped training. A requirement that exists only to be falsely certified needs deletion or honest hours dedicated to it.
The Tools Already Exist
Annual trainings like Cybersecurity Awareness and administrative requirements like awards citations are different categories of work for the force.
The former is often either a response to a shortcoming or a preemption based on foreseeable and preventable risk. The Army is debating the merits of these responses. It is worth clarifying that no model can/should sit through the module on anyone’s behalf; if the Army decides the training is worth it, it is butts in seats. This category is more like physical training; outsourcing it means you are missing the point.
The latter is admin. There is far less merit to cutting counseling, promotions, and recognitions. It is work worth doing. What is on the chopping block is how we do it – e.g., the compliance labor around it, the typing, and the routing for a signature – since each is a document assembled from facts a leader already has.
The drafting tool stopped being hypothetical while the Army was busy deleting requirements. In December 2025 the Department launched GenAI.mil, an accredited platform cleared for controlled unclassified information; by February it carried ChatGPT and had spread to five of the six branches with more than a million users.
The drafting tool stopped being hypothetical while the Army was busy deleting requirements. In December 2025 the Department launched GenAI.mil, an accredited platform cleared for controlled unclassified information; by February it carried ChatGPT and had spread to five of the six branches with more than a million users. The Army fielded its own Enterprise LLM Workspace in May 2025 and kept the earlier model CamoGPT running beside it. The Secretary of Defense’s January 2026 AI strategy names enterprise workflow agents one of four pace-setting projects. The tools sit in every soldier’s browser but almost none of their capacity addresses the checklist.
The Army ran the arithmetic once already. When it needed 300,000 personnel descriptions rewritten, the Workspace did it in about a week, work the service estimated at roughly 50,000 human-hours of review. The Department of Veterans Affairs piloted drafting veterans’ disability questionnaires straight from medical records since August 2025. Alongside hiring surges and process changes, its claims backlog fell 72 percent while average decision time dropped from about 141 days to under 80. Members of Congress are watching the error rates, and they should be. What the pilot proves is narrower and still useful: drafting from existing records works inside a large federal bureaucracy.
Think of the model as a clerk. It assembles the document, gets the format right, and hands it up for signature. It never decides who earns an award, and nobody would ask a commander to type the citation themselves. Most of what the Army has to decide about AI is not novel. The Army already knows how to supervise a clerk – this one just works faster and never gets tired. It involves seeing technology through a new metaphor and an openness to acknowledging human limits as much as we acknowledge technological limits. An AI clerk also keeps better records than a tired officer signing from memory at 1900 on a Friday.
Putting the clerk to work involves three gates, and the model is not any of them. The first is money, and it is already open: the platforms are bought, accredited, and sitting on every soldier’s browser, so a drafter working from facts a leader types could be running as soon as this quarter. The second is data. These systems should be wired into the Army’s personnel and training systems, Integrated Personnel and Pay System-Army (IPPS-A) and Army Training Information System (ATIS), systems that are rate-limited by record quality. IPPS-A publishes its own running list of known errors, and a clerk working from a bad record hands you a confident, well-formatted mistake with a signature block already attached. Closing that gate takes accreditation and data cleanup, which is what the Department’s new software fast-track exists to shorten. The third gate is permission. Someone senior must put in writing that a drafted counseling, verified and signed, is still a counseling. This gate costs nothing yet has not been done.
Three Costs
Automating a bad requirement makes it cheap, and cheap things spread. If a safety assessment takes ten seconds instead of ten minutes, a brigade staff worried about missing something will add three more forms beside it. Without the discipline General George showed in cutting the slate, drafting tools leave soldiers as buried as before but holding a faster shovel. Any drafting tool the Army adopts should ship coupled to a review on the requirements it serves. The Department has already run a counter-experiment with the Defense Travel System (DTS), the Army’s system for compensating soldiers for official travel. DTS automated the paper travel voucher three decades ago, hardened around its own workflow, and outlived a $374 million attempt to replace it.
A model can be confidently wrong. It can draft a readiness report with a fake number or summarize a counseling in a way that changes what a leader meant. The signature stays human for exactly this reason. The model prepares the document; a person verifies the facts and signs, taking final ownership of the document.
One more fact should discipline this debate: soldiers are not waiting for permission. Soldiers already use generative AI models to draft evaluations and award language.
The quieter failure is the person checking the model. People who monitor reliable automation stop checking it, and a signature performed as ceremony on an unread draft would put the lie Wong and Gerras described on a faster clock. Ten minutes of real verification against an hour of drafting saved is still a good trade for the soldier. The verification step must actually cost those ten minutes: a fact-by-fact check rather than a single click, liability staying with the name on the signature block, and drafts sampled against the records they cite, the way finance offices already spot-check travel vouchers.
Freed hours have a way of getting refilled. If the Army claws back a week of a soldier’s year and immediately fills it with new taskings, the recovered week disappears into the new requirements before the soldier ever sees it.
One more fact should discipline this debate: soldiers are not waiting for permission. Soldiers already use generative AI models to draft evaluations and award language. When a soldier built VECTOR, an AI talent-management tool, the Army suspended it pending a compliance review. Delay has moved the demand underground rather than ending it. An accredited, logged version is the alternative still on the table.
Conclusion
The captain from the opening is still at their desk. The Army decided their training mattered, then built a system that made honest completion impossible, making quiet shortcuts the norm. The platforms that could do the rote part of their paperwork are accredited, fielded, and idle in their browser. Point them at the citations and the counselings, keep the judgment and the signature with the leader, and audit what they sign. They came for service. The hour returned goes where they always meant to spend it: planning training, fixing soldiers’ problems, and required training modules – if important enough – done honestly and only once.
