AI Reshapes Military Headquarters: Efficiency, Compression, and Human Judgment

Last year at Nellis Air Force Base, an artificial intelligence system was tasked with building courses of action—a complex operational problem that military planners have spent generations mastering. The AI completed the task in under a minute, yielding solutions rated at 97 percent viable and tactically valid (essentially, whether a proposed plan could work given operational constraints like fuel and force composition). By contrast, human counterparts required roughly 19 minutes and achieved only a 48 percent viability score.

This comparison requires two caveats. First, the tool achieving that 97 percent viability originated from a single vendor that normalized the data, cleaned spreadsheets, and translated narrative material before the software even encountered the problem. Second, the process was merely a two-week sprint, hardly a full deployment. Still, one operator involved described the machine output as an “80 percent solution”—a solid starting point around a maturing capability that still required experienced planners to finalize the details.

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 speeds that fundamentally alter how headquarters are staffed. Public discourse on military AI typically centers on autonomy at the point of attack, whether an algorithm should select targets or release weapons. This transformation, however, unfolds several steps behind the trigger, inside staff sections responsible for planning and analysis.

Recently, an article in these pages examined the promises and pitfalls of automating staff work across the national security bureaucracy. Operational military headquarters face a distinct challenge: they must deploy these technologies against adaptive adversaries, often in contested and degraded environments.

Artificial intelligence creates an opportunity to build smaller, more expert staffs that disperse more easily and reduce physical and electromagnetic signatures. Compression comes with a cost, though. If you remove too much human expertise and redundancy, a headquarters may become less resilient to combat stressors. The line between optimization and single points of failure is fine.

AI should shrink the portion of a headquarters dedicated to producing information, while leaving intact the functions built to coordinate that information and answer for it. To test this concept, the Pentagon should design a pilot program, designating a chosen headquarters as a testbed for alternative AI-enabled staff designs.

Transforming the Military Staff

Military staffs exist precisely because no single commander can absorb the sheer volume of data needed to understand what is possible: intelligence shifting by the hour, changing friendly force statuses, terrain logistics, and continuous decision-making from headquarters. Armies solve this by distributing cognitive loads across staff that each hold slices of information, feeding judgments to a commander.

AI has begun to perform the work this division of labor was invented for. Agent Network, launched in June under the Pentagon’s AI Acceleration Strategy, scans intelligence and operational systems continuously and converts findings into commander options 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 vast bodies of material, reconcile competing inputs, and revise analysis continuously as conditions change. If machines can perform the intermediate cognitive work staffs were built for, the size and structure of headquarters surrounding a commander deserves serious reconsideration. Done well, the result could be a smaller staff carrying more expertise per officer, with compression concentrated in functions like intelligence synthesis and maintaining running estimates where large staffs currently spend considerable time tuning streams of information into analysis. However, resiliency remains crucial. Redundancy among people is vital for headquarters to survive severed network links or losing a section chief mid-crisis, and a small staff may lack the depth to absorb such shocks.

As we think about where AI can be most useful to military staffs, we should start with the running estimate. It is the basic unit of staff work, the mechanism through which a section converts raw information into judgment a commander can use. The Army found that commanders routinely leave briefings without the knowledge they need 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 to reconcile with every other one, compounding the coordination burden. This is particularly visible in Combatant Commands.

Most 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 an easy place to start cutting. Collecting data and assembling briefing products consume large shares of staff hours without drawing on anyone’s judgment, 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 excel at. At Nellis, the software produced options at a speed a planning cell could not match. Experienced operators were still needed to catch hidden assumptions and physical limits that made some 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 translates these categories into billet cuts. While a machine absorbs much of a planner’s workload, the work left behind increasingly depends on operational experience. Compression should fall hardest on those who produce information, and lightest on those 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, concentrating experienced officers around the judgments machines cannot make, all without becoming intellectually brittle.

The Dual Challenges

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.

Compression creates two risks. The first is that officers will trust polished machine output 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 degrades.

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 data had been carefully prepared. In a real-world scenario with incomplete intelligence, the outputs could be elegant but flawed, something only an experienced planner could catch. As models improve, this problem will grow harder to manage because increasingly coherent output can conceal a weak assumption behind persuasive reasoning. 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 officers would trust outputs they 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. Furthermore, staffs in high-operations tempo environments face extreme cognitive loads. An adversary can exploit that deference without ever touching the software by feeding decoy formations and spoofed signals. A headquarters that has compressed its human analytical capacity too aggressively may simply become highly 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. Systems that perform beautifully in a cloud-connected exercise can become liabilities the moment headquarters personnel lose the practiced ability to plan without them. Habitual reliance on machine-generated analysis risks eroding the independent judgment officers need most—precisely in the degraded environments where that judgment matters most.

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 Paradigm Shift

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, interfaces showing 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. There is already 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 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—will reveal the differences. Resilience after losing a network node and performance under deliberate deception must also belong in the assessment; 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 among 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 over 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.

Source link