At Nellis Air Force Base last year, an AI system was tasked with building courses of action, a complex operational problem that military planners have spent generations learning to perform. The AI system took less than a minute to complete its task with solutions rated at 97 percent viable and tactically valid (essentially, whether the proposed plan could work given operational constraints such as fuel and force composition). By contrast, their human counterparts took roughly 19 minutes and only reached 48 percent viability.
That comparison needs two qualifiers. First, the tool that hit 97 percent came from a single vendor that had normalized the data, cleaned up the spreadsheets, and translated the narrative material before the software encountered the problem. Second, the process was just a two-week sprint — hardly a full deployment. Yet still, one of the operators involved described the machine output as an “80 percent solution,” a real starting point around a maturing capability that still required experienced planners to finish the job.
What the Nellis sprint demonstrated is that once a problem is prepared for them, machines can weigh competing constraints and produce usable options for a commander at a speed that changes the underlying assumptions of headquarters staffing. Most public debate about military AI centers on autonomy at the point of the spear, whether an algorithm should ever select a target or release a weapon. The transformation underway also sits several steps behind the trigger, inside staff sections.
Recently, an article in these pages examined the promises and pitfalls of automating staff work across the national security bureaucracy. Operational military headquarters present a different problem: They must employ these technologies against an adaptive adversary, often in contested and degraded environments.
AI creates an opportunity to build smaller, more expert staffs that can disperse more easily and reduce their physical and electromagnetic signatures. But compression comes at a cost. If you remove too much human expertise and redundancy, the same headquarters may become less resilient to the stressors of combat. There is a fine line between optimization and single points of failure.
AI should shrink the part of a headquarters built to produce information, while leaving largely intact the part built to coordinate that information and answer for it. To test the efficacy of this idea, the Pentagon should design a pilot, designating a chosen headquarters as a testbed for alternative AI-enabled staff design.
The Military Staff
Military staffs exist because no commander can absorb the sheer amount of data needed to understand what is possible: intelligence that shifts by the hour, the changing status of friendly forces, and the terrain and logistics constraint, all while decisions keep arriving from headquarters. Armies solve this by distributing cognitive loads across staff that each hold slices of information to feed judgments to a commander.
AI has begun to perform the work that this division of labor was invented to accomplish. Agent Network, launched in June 2026 under the Pentagon’s AI Acceleration Strategy, scans intelligence and operational systems continuously and converts what it finds into options for commanders within seconds, while leaving target selection to human judgment. In August, the department added ChatGPT Mil to its GenAI.mil platform, extending frontier models to document-heavy planning and logistics work across the workforce.
AI can search across large bodies of material, reconcile competing inputs, and revise analysis continuously as conditions change. If machines can perform the intermediate cognitive work that staffs were built to perform, the size and structure of the headquarters surrounding a commander deserves real reconsideration. Done well, the result could be a smaller staff carrying more expertise per officer, with compression concentrated in the functions such as intelligence synthesis and maintaining running estimates where large staffs currently spend considerable time tuning streams of information into analysis for commanders. But resiliency remains important. Redundancy among people is part of how headquarters survive severed network links or the loss of a section chief mid-crisis, and a small staff may lack the depth to absorb that kind of shock.
As we begin to think of where AI can be most useful to military staffs, we should start with the running estimate. The running estimate is the basic unit of staff work, and the mechanism through which a section converts raw information into judgement a commander can use. The Army found that commanders routinely leave briefings without the knowledge they need to decide because their staff present unrefined data rather than the analysis required to turn it into a usable conclusion. Every additional specialty a modern headquarters absorbs, such as cyber or space command, adds real expertise, but it also adds another interface that must be reconciled with every other one, and the coordination burden tends to compound. This is particularly apparent in Combatant Commands.
Most of the debate over AI in military planning asks what the technology can do, but the more useful question for headquarters is what it allows a commander to take out. Routine production is the easy place to start cutting. Collecting data and assembling briefing products eat a large share of staff hours without drawing on anyone’s judgement, and military writers already treat these tasks as near-term candidates for automation. Intelligence synthesis and course-of-action development are harder because each requires interpretation, as recently shown in intelligence challenges concerning Chinese shipping. Each also rests on the high-volume comparison and constant updating that machines do well. At Nellis, the software produced options at a speed planning cell could not match. Experienced operators were still needed to catch the hidden assumptions and physical limits that made some of those options unworkable. In this work, the officer’s job shifts from building the analysis to interrogating it. Decisions about how much risk to accept and the costs of an action remain with the commander.
The debate goes wrong when automation potentially translates these categories into billet cuts. While a machine absorbs much of a planner’s workload, as detailed above, the work left behind increasingly depends on operational experience. Compression should fall hardest on the people who produce information, and lightest on the people with the expertise to challenge it. Simply giving every section an AI assistant while preserving the organizational chart misses the larger opportunity. AI should allow headquarters to become smaller and more dispersed, while concentrating experienced officers around the judgments machines can’t make, all without becoming intellectually brittle.
A headquarters built around persistent machine analysis would look different from one where each section simply receives an AI assistant while preserving the same organizational structure. A smaller staff needs less equipment and a smaller electromagnetic signature to operate, at a moment when Army planners are already redesigning command posts around dispersion and mobility, drawing lessons directly from Ukrainian forces on concealment and rapid displacement to survive detection by drones and long-range fires.
Two Risks
Compression creates two risks. The first is that officers will trust polished machine output that they did not build and cannot easily check. The second is dependence: A staff that loses the habit of planning without machines may be unable to plan when the network fails or the data goes bad.
The Nellis experiment demonstrated the potential of AI-enabled planning, producing sophisticated plans almost immediately and outperforming human planners on the exercise’s viability metric. But those results emerged from a controlled, perfect environment where the data had been carefully prepared. A real-world scenario with incomplete intelligence could produce elegant but flawed outputs that only an experienced planner could catch. As models improve, this problem may grow harder to manage because increasingly coherent output can conceal a weak assumption behind persuasive reasoning, and an officer reviewing a polished recommendation at machine speed faces a fundamentally different cognitive task than one who built that recommendation from scratch.
There are several reasons that officers would trust outputs that they themselves did not create or verify. Center for Security and Emerging Technology researchers Lauren Kahn, Emelia Probasco, and Ronnie Kinoshita have documented how little protection a “human in the loop” actually provides when interface design pushes users toward deference.
And staffs in high operations tempo environments are exposed to extreme cognitive loads. An adversary can exploit that deference without ever touching the software by feeding the system decoy formations and spoofed signals. A headquarters that has compressed its human analytical capacity too aggressively may simply become very efficient at processing a corrupted picture of the battlefield. China’s own procurement documents describe AI decision-support systems partly as a means of compensating for perceived weaknesses in their officer corps, making People’s Liberation Army headquarters an inviting target for deception (and a bet the United States should not copy).
A second category of risk concerns dependence itself. Any force preparing to contend with a peer adversary should assume that data quality will degrade during actual conflict, and systems that perform beautifully in a cloud-connected exercise can become liabilities the moment headquarters personnel lose the practiced ability to plan without them. Staffs will lose access to systems they have come to depend on in degraded environments. Habitual reliance on machine-generated analysis risks eroding the independent judgment officers need most in exactly the degraded environments where that judgment matters.
The most serious version of this problem concerns how junior officers learn the profession in the first place. Planning skill develops through repetition and the experience of constructing an assumption to discover why a plan can falter if not wargamed. If machines absorb that labor before an officer has done it often enough to internalize why it works, the force risks losing the expertise required to supervise the machines a decade later.
Testing the Redesign
The remedy is to compress the production of information but never its interpretation through established evaluation criteria. That means a small analytic cell charged with contesting the machine’s picture against independent sources, and interfaces that show what inputs drive a recommendation, and what evidence would overturn it. It also means exercises that grade headquarters on whether they caught an adversary’s attempts to corrupt their data.
The way to learn how far compression can safely go is to test it, and there is some progress here. The department’s 2026 AI Acceleration Strategy explicitly calls for iterative experimentation across its pace-setting projects, including Agent Network and a program built to accelerate simulation and feedback loops. The Army’s Next Generation Command and Control effort has already moved from a proof of concept to a division-level exercise in under a year, testing whether commercial AI tools can support division-scale operations with a materially different staff architecture.
The logical next step is to let selected headquarters redesign around AI-enabled workflows, rather than layering tools onto an unchanged organizational chart, comparing units that preserve current staff sections against units that consolidate functional cells around persistent agents and units that distribute personnel geographically while keeping a small command group forward. Resilience after losing a network node and performance under deliberate deception should also belong in the assessment because a headquarters that produces an excellent plan in a clean information environment may prove fragile against a peer adversary determined to intentionally corrupt the data environment. This is exactly what Scale AI is doing for the Air Force in the Survivable Airborne Operations Center Program.
As machines increasingly perform tasks that once consumed an entire section’s day, the staff can immediately evolve from a large production bureaucracy into a smaller organization built around expertise, judgment, and combat effectiveness. A headquarters that processes more information with fewer people can adapt faster, disperse more widely, and let commanders spend a greater share of their attention on the decisions that genuinely require command. However, a defense culture designed to compete for resources amongst services and even within services tends to be reluctant to embrace new and emerging technologies. Key assignments, promotions, and staff billets traditionally favor those that promote the current system versus asking strategic questions about defense’s role in national resourcing priorities. Overcoming that resistance will require joint-manning documents and promotion boards that credit officers for redesigning the staffs beneath them so that shrinking a headquarters becomes a career asset.
Additionally, professional military education and unit training should preserve deliberate periods of manual planning even as operational headquarters lean further into AI-enabled workflows. Military doctrine involving training and education must shift to support these emerging realities.
The larger competition over military AI will extend beyond autonomous weapons and the race for the most capable model, turning on the design through which humans and machines think together under pressure, and on which military reaches there first. Its remaining task is institutional and cultural, requiring that we experiment aggressively enough to discover which forms of headquarters these capabilities make possible, and stay disciplined enough to preserve the judgment that gives command its purpose in the first place, keeping the commander human and redesigning what surrounds them.
Josh Richards is the chief commercial officer of Pacific Peering. He serves on the U.N. Joint Task Force on SMART Cables as a member of the Steering Committee, and chairs the Business Development Committee. He is a security fellow with the Truman National Security Project, a tech policy fellow with the Aspen Institute, and a senior fellow with AI 2030.
Gabe Arrington is a senior military officer, accomplished national security and business strategist, visionary technologist, celebrated international best-selling author, and entrepreneur. He is the former chief of disruptive technology for the United States Air Force, is a non-resident fellow at Baker Institute of Public Policy, focusing on energy studies, a senior fellow with AI 2030, the director of Texas AI Alliance’s Advocacy Groups, and a graduate of both the Massachusetts Institute of Technology’s Seminar XXI and Center for a New American Security’s national defense fellow.
Image: Debora Henley via DVIDS.