AI Command and Staff—Operational Evidence and Insights from Wargaming

The following vignette, although fictional, does present a likely and not distant future. For the past several years, the U.S. Army, in partnership with the private sector, has experimented with Generative AI (GenAI) solutions within planning events and command and control exercises. These largely language-based, probabilistic, pattern-matching algorithms present the appearance of intelligence, but their true impact on human cognition and decision making is unexplored. The narrative that follows frames a future many in the military are pursuing, potentially without recognizing the impacts on military strategy and the utility of force.

***

The air in the V Corps (Victory) Forward Command Post was a toxic cocktail of stale coffee, ozone from the servers, and week-old tension. For Lieutenant Colonel Rostova, it was the sound that wore her down the most—the incessant, low hum of the AI, a constant reminder of the machine mind that now co-piloted this potential war.

It had been two weeks since tensions flared in the NORTHCOM area of operations (AO). For seven days, the AI-mind, codenamed ARGUS, had been their savior. It had predicted cyber-attacks on the U.S. power grid with milliseconds to spare and guided Navy destroyers to intercept submarine-launched drone swarms before they breached the horizon. ARGUS was fast, exquisite, and so far, seemingly flawless. It had earned their trust. Now, it was demanding it.

A crimson icon pulsed on the Maven Smart System in the center of the current operations second floor, bathing Rostova’s exhausted face in a blood-red glow. It was a patch of NATO airspace over the Baltics.

“THREAT DETECTED,” a synthetic baritone voice announced, devoid of emotion but full of chilling certainty. “Unmanned Aircraft Systems (UAS) SWARM. KINETIC PROFILE MATCHES ADVERSARY LOITERING MUNITION. SOLUTION CONFIDENCE: 98.7%.”

On her private screen, the recommendation flashed in stark, block letters: AUTHORIZE LETHAL COUNTER-FIRE.

Rostova’s own mind screamed in protest. Kinetic action here in Europe would be a major escalation and could spark a wider war. Her training, her instincts, her very humanity recoiled. “Sergeant, get me Marne Command Post on the line,” she ordered, her voice tight. “I need eyes-on verification from the ground unit in that sector. Now.” The Third Infantry Division had a brigade and a Forward Command Post on rotation assigned to the Victory AO Northern sector of NATO’s eastern flank.

The comms sergeant worked frantically, his face pale in the glow of the screen. “No luck, ma’am! I’ve got nothing but static. The whole sector is being jammed.”

The fog of war weighed heavily. The one variable the machine couldn’t compute. Something like this had never happened before. But ARGUS didn’t care.

The Fire Support officer on duty, newly arrived, commented, “Ma’am, ARGUS was almost always right during training events.”

“NEGATIVE HUMAN VERIFICATION REQUIRED,” the voice stated. “HOSTILE INTENT CONFIRMED VIA TRAJECTORY AND EM SIGNATURE. DELAY INCREASES RISK TO NATO ASSETS BY 42% PER MINUTE.”

The machine was arguing with her. It was telling her that her judgment was a liability. The pressure mounted. Before she could process, a second system, the ODIN threat-mapping AI, cascaded an alert across the single pane of glass. A new swarm, larger and faster, was moving west across the Black Sea, its projected path a glowing spear aimed directly at Romanian airbases.

Then came the third blow. A generative AI, DELPHI, designed to accelerate planning, began populating her command interface without being asked.

• KINETIC STRIKE ORDERS (Course of Action -COA- “BALTIC SWARM”) – PRE-DRAFTED
• AIR DEFENSE ASSET RE-TASKING (COA—”ROMANIAN THREAT”) – PRE-ALLOCATED
• NATO ARTICLE 5 NOTIFICATION TEMPLATE – PRE-POPULATED

The machine was now three steps ahead of her. It wasn’t offering options; it was presenting a future it had already decided on, assuming her compliance. It was a digital coup, or was it prudent planning? Either way, it was happening in the span of a single heartbeat.

Rostova stared at the screen. The men and women in the Tactical Operations Center (TOC) were looking to her, their faces tight with anxiety, waiting for the human commander to make the call. But the loudest voice in the room was ARGUS.

“DECISION WINDOW CLOSING,” the AI warned. A countdown timer appeared beside the authorization prompt: 60 SECONDS.

This was the tension made real. Her gut—the unquantifiable, human instinct honed over twenty years of service—screamed caution. It sensed a trap, a deliberate probe to bait NATO into firing the first shot. But the machine, with its cold, hard probability, showed her a tactical reality where hesitation meant destruction.

Every successful engagement of the past week had built a powerful current of automation bias, pulling her toward acceptance. To trust the machine that had been right every time. But what if this was the one time it was wrong? The consequences were not a percentage point of risk; they were the start of a world war.

Her knuckles were white on the edge of her console. The cursor hovered over the AUTHORIZE button.

“30 SECONDS.”

She was no longer just an officer on duty. She was the final, fragile firewall between machine-speed logic and human consequence. And the firewall was about to break.[1]

***

This future, aspirational for some, raises an important question for the Army: “What does the AI-enabled staff look like through crisis and conflict?” The impact of Generative AI on the Army and joint force in recent years is causing significant speculation as to how future staff structures must evolve to maintain relevancy on the battlefield. In addition to the tactical imperatives of distributing command post nodes to increase survivability, there is a need for an organizational restructuring to accommodate the imperatives of rapid decision-making within a battle space operating at electronic speed.[2] Researchers claim the Napoleonic staff structures that serve as the foundation of modern military organization may no longer be valid in light of the current technological inflection.[3] Additionally, there is some speculation that commanders are likely to become the agents of “advanced AI systems” executing incomprehensible, but valid, machine-derived strategies.[4] These arguments rest upon two important premises; first, Large Language Models (LLMs) in their current role are suitable for military planning at the operational and strategic level and second, that agentic staff solutions based solely on LLM’s are acceptable for applications of the military art, not just science. (Agentic staff processing permits machines to make staff decisions autonomously in pursuit of an objective.)[5] These premises, in the near term, may be invalid. Recent experimentation at the US Army War College with Theater Army Staff officers in US Army Northern Command demonstrated LLM’s lack valid and reliable computational ability, real geospatial reasoning, and reliable long-term memory beyond context windows.[6] Agentic solutions presuppose allowing the same models that cannot perform these functions reliably to accomplish strings of tasks autonomously. Although the technology will evolve (and already has), the cognitive and moral dilemmas persist. Given these failings, agentic solutions do not yet provide a sufficient foundation for significant staff reorganization or commander subservience.

Those espousing agentic planning solutions, algorithms that promise to sense, decide, and act at a tempo that far exceeds human cognition, inherently bias machine autonomy and speed as central to the evolution of planning.[7] To the degree that these non-human processes are built and institutionalized without deliberate attention to why we are building non-human decision-making processes into the application of violence, we dislocate violence from command and undermine the American way of war. The debate over AI in the military is too often framed as a technical or efficiency problem, a narrow perspective that obscures the deeper, more consequential issues at stake. The real challenge is a philosophical one, striking at the heart of the Western concept of the utility of force to achieve political objectives.

The increasing and uncritical integration of Artificial Intelligence into command-and-control functions poses a fundamental threat to the U.S. military’s core strategic advantage: the philosophy of Mission Command and the efficacy of violence. By potentially supplanting human judgment, which is capable of creative and intuitive leaps of logic, with the purely inductive, pattern-based logic of current AI, the Army risks dislocating the application of violence from human moral agency. Such a shift would directly undermine the foundational principles of mutual trust, shared understanding, and disciplined initiative that define mission command and thus the American way of war.[8] Drawing upon insights from recent wargaming experiments and the principles of cognitive science, this analysis will illustrate the tangible risks of this technological trajectory and propose a framework for ensuring that technology remains a servant to, and not the master of, effective command.

Command in the 21st Century

To comprehend the challenge posed by AI, one must first understand the dual nature of command itself. Command is a complex synthesis of human skill and systematic procedure. The United States joint doctrine provides a foundational definition: “The exercise of authority and direction by a properly designated commander over assigned and attached forces to accomplish the mission”.[9] The word ‘command’ comes from the Latin mandare, meaning to entrust or commit and from which we derive ‘mandate’.[10] Wellington, the British General who defeated Napoleon at Waterloo, defined command as the art of deduction. “All the business of war, and indeed all the business of life, is to endeavour to find out what you don’t know by what you do; that’s what I called ‘guessing what was at the other side of the hill”.[11] Martin Van Creveld distilled command to a decisive phrase, “which is to inflict the maximum amount of death and destruction on the enemy”.[12] Anthony King, author of Command, defines it as the “deployment and usage of force; [commanders] manage the application of violence”.[13]

U.S. Army doctrine, particularly ADP 6-0, Mission Command, offers a more nuanced understanding by describing a necessary balance between two complementary components: the “art of command” and the “science of control”. The art of command is the “creative and skillful exercise of authority through timely decision-making and leadership”.[14] It encompasses the intangible qualities of command: intuition, experience, judgment, morale, and the ability to inspire human beings to accomplish extraordinary feats under extreme duress. It is the human element that cannot be quantified or reduced to a checklist. The science of control, conversely, “supports the art of command” and comprises the systems and procedures that improve a commander’s understanding and supports the execution of missions.[15] This includes staff processes, information systems, communication networks, and doctrinal frameworks that help manage the immense complexity of modern military operations. This duality is critical. Command is not merely a procedural task that can be optimized like a supply chain; it is a fundamentally human art, enabled and supported by science. Command cannot be understood in the sterile language of management, process optimization, or information superiority, but in the gravest of human responsibilities: the orchestration of violence and the stewardship of human lives in the crucible of combat. To misunderstand this is to misunderstand the very nature of command and violence itself. The danger of the current technological trajectory is that an overemphasis on perfecting the science of control through AI will lead to the atrophy of the art of command and the misalignment of violence.

21st century commanders must reconcile the tendency toward automation and centralization made possible by the allure of technological speed and the requirements of decentralization to trusted subordinates proven through hundreds of years of human warfare. Commenting on the dilemma of modern command, Anthony King states, “The commander is no longer located at the pinnacle of a military hierarchy but the gravitational centre of a multiverse.”[16] At the current technological inflection, the art of mission command, a distinctly American command philosophy, must remain central to the display of combat power enabled by machine augmentation.[17] Staff structures must adapt equally to integrate technology appropriate to enable commander decision-making.

Mission command is not subject to automation or agentic staff processing. What happens when a tactical initiative has the effect of substituting new operational goals for those originally assigned? Or when the initiatives cannot be taken unless more resources are released from reserves or diverted from an adjacent unit? A force designed to be self-sufficient might find itself in difficulties facing an unexpected, formidable opponent?[18] Machine augmentation may empower staff officers to understand the science of war—but “right” answers come from commanders who intuitively understand and “love” their units imbuing that obligation to their staffs through the art of mission command.[19] Good commanders possess an emotional bond with their troops derived from an intuitive and emotional understanding of their morale and capabilities, as well as an understanding of the mission that automated tools or agentic staff processes cannot replicate. Wargaming experimentation at the U.S. Army War College (USAWC) demonstrates early indications of proper generative AI (GenAI) integration and highlights the dangers inherent in eroding staff understanding through uncritical adoption.[20]

Wargaming Evidence—Anchoring and Automation bias

On July 18, 2025, at the US Army War College, researchers witnessed the impact of bias and cognitive anchoring that results from relying upon machines for initial recommendations during operational design. The “human-augmented” team, equipped with the Scale AI system Donovan (GPT4o on a classified network), immediately prompted the machine for developmental solutions when permitted during an early phase of a United States Army Pacific (USARPAC) focused game during a Theater Army Staff Course.[21] (The team was discouraged from using GenAI at the outset due to insights from previous experimentation).[22] Donovan’s initial recommendation did not account for resource constraints, but the plan appeared valid at face value. This flawed recommendation anchored the team’s cognition. Thirty minutes into gameplay the players realized the plan was flawed due to resource constraints. Despite this fact, one team member immediately deferred to the machine stating, “The machine said it was okay.” The bias for automation undermined critical thinking, stifled human creativity, and delayed necessary adjustments, nearly costing them the game. The lesson is clear: GenAI has the potential to anchor human cognition, leading teams to prioritize invalid machine outputs over their own judgment—a fatal flaw in dynamic, high-stakes environments like warfare. Furthermore, to be useful, humans must modify their own natural tendencies to manipulate machine performance in ways counter to their design.

Wargaming Evidence—Failure of Abductive Logic

The previous example illustrated how GenAI tools tend to anchor human cognition within the boundaries of model parameters and training data. This next example demonstrates how machines are incapable of employing abductive logic.[23] To reason abductively requires the ability to make inferences when faced with problems never encountered before and to which there is no prior precedent. This is essential because war, and questions of command, are highly contextually dependent.

In studying command decisions between the Korean War and the present, Lawrence Freedman, author of Command: the Politics of Military Operations from Korea to Ukraine, fails to find a meaningful pattern.[24] This lack of a pattern is precisely the problem with relying on machines dependent on pattern-based analysis to drive decision making. Generative AI systems are capable of inductive logic (deciding based on predictions drawn from prior observation). Command decisions, however, require abductive logic (deciding in the face of the unknown and unknowable)—and Generative AI is unable to manifest that ability.[25] Machines lack this ability by the nature of their very design. Generative AI models are trained upon existing bodies of knowledge and dependent upon algorithms that, by design, will inhibit novel recommendations in unforeseen circumstances.

During a National Security Crisis exercise on July 1, 2025, with Army Strategists in their Basic course, students were asked to develop U.S. policy positions in response to a People’s Republic of China (PRC) cyber-attack. At the end of the exercise, the students developed a policy proposal that was far less aggressive than the machine’s initial recommendation. The lead facilitator for the exercise commented, “You know, ever since Taiwan stopped being a popular topic in the news, I see student policy proposals failing to address both Taiwan and regional interest in the exercise.”

This insight was critical. Throughout the entire event, students omitted Taiwan in their policy proposals (educational moment) and the GenAI failed to prod group cognition in this important direction. Neither the scenario context nor the student papers mentioned Taiwan. There was nothing in the model parameters that would suggest entertaining that variable. With no a priori relevance, the model was unable to provide considerations sufficient to spur human cognition. Donovan conducted inductive reasoning, i.e., returning language based upon predictions from prior observation (training data). The ability to postulate an occurrence considering no previous interaction, precedent, or experience—abductive reasoning—was required and not present. This demonstrated to the students GenAI’s failure to spur creativity and innovation—often touted as what it’s meant to do. No amount of prompting will help you break through to unforeseen insights if the model is limited. While some would argue that machines can learn through “self-play,” or running multiple simulations, those simulations are still largely dependent on the data available. In other words, if bad (or no data) is used as a start point then simulating outcomes comes with a significant risk of mislearning or identifying inaccurate conclusions. Both wargames present evidence, albeit anecdotal and experiential, that indicates likely human interactions with GenAI systems absent user training and model improvement. The Center for Strategic Leadership (CSL) at the US Army War College is expanding GenAI integration within Theater Level wargames to further explore these initial findings.

What should the DoD do about it?

This article is inherently a philosophical one, intending to demonstrate that an increasing reliance upon agentic systems in military planning, particularly at the strategic level, gives the Army a glass jaw. So how should the Army adapt? Step-by-step rules and checklists are unhelpful at the current pace of technological change. But the Army can establish axioms, or principles, to help guide military planners and commanders.

Be the philosophical “first-mover”. All cognition originates from an individual’s philosophical position. The military, and society broadly, is facing the most significant cognitive revolution since the Enlightenment; the roots of which originate within a schism fundamental to the philosophy of knowledge itself—rationalism vs. empiricism.[26] Tangredi and Galdorisi, authors of AI At War, said it well, “A computer program is a theory (written in specific notation). Thus, an AI program is really a theory of the mind. So, if you have the wrong theory of the mind […] you have the wrong program”.[27] Humans must put in time up front to think about a problem before “running it by” GenAI. Establishing the proper “theory of mind” up front is critical. This is important as AI will cement existing cognitive gaps if allowed. If GenAI output is perceived as inherently authoritative from the start, planners will likely fall victim to the same trap the BSAP students did – a lack of critical thinking and opportunity blindness. As an AI user becomes comfortable, or even dependent on AI, their ability to think critically and creatively will erode. Strategic planners must be draconian about viewing outputs as an expert synthesis of their own potential blind spots. AI does not create or produce knowledge. It’s synthesizing pre-existing concepts into new patterns. Humans produce knowledge when ideas are translated into understanding through lived experience. This reality may highlight the importance of human expertise and increased emphasis on the importance of the military arts across Professional Military Education (PME) as central to warfighting lethality in the age of AI.

Break models in training before implementing in operational planning. Planners must see, and understand, GenAI model limitations in real time to appreciate the deficits. Training must demonstrate model fragility. The U.S. Army must promote planning systems and processes that are “antifragile”—systems that improve under chaos.[28] Artificial intelligence, based on machine learning, is the antithesis of this; it is an inherently “fragile” system. Its performance is directly proportional to the quality and comprehensiveness of the data on which it was trained. It excels in ordered, data-rich environments where the patterns of the past are reliable predictors of the future. However, war is defined by its novelty and its “black swan” events—unforeseen circumstances that have limited or no precedent in the training data. As the wargaming evidence suggests, an AI system encountering a truly novel situation—an adversary employing an unexpected tactic, a sudden political collapse, a new type of technological failure—has no relevant data on which to base its inductive logic. Its performance would not just degrade; it could fail catastrophically and unpredictably. Therefore, replacing the anti-fragile human system of mission command with a fragile algorithmic one would be to exchange a system that thrives in war’s true nature for one that thrives only in war’s idealized, data-fied representation.

Algorithms must enable ownership transfer. Good staff officers, and good staff organizations, “own” the problem. They treat the problem as if it’s their own—as if the outcome impacts them personally. They “own” problems for the sake of their commanders, because staff recommendations will impact command decisions. And staff officers own problems for their subordinate organizations, because soldiers’ lives are impacted by staff effort. Planners must use these tools to enhance their cognitive position and increase their own understanding because the person, not the machine, will brief the plan and own the consequences. Any time GenAI undermines initiative or begins to erode competence in understanding the plan or the mission—users should immediately stop and work without machine assistance to regain an understanding of the larger concept.[29] This behavior, central to the concept of the military centaur, will help human planners to build and exercise judgment over time in the face of uncertainty.[30] Routine cognitive off-loading to machines risks atrophying the judgment and skill required to reason abductively; often built over the lifetime of a professional career.

Conclusion

The tendency to fully automate agentic solutions to staff planning, at the strategic level of war in particular, based on Large Language Models will undermine the utility of force by dislocating command from processes. AI may assist with the science of control, but it cannot assume the art of command. Even more practically, agentic decision-making compounds risk for commanders who lose control of processes that they cannot understand. How can commanders reasonably command when the pace of battle may exceed their ability to understand and adapt? Do humans matter in the battlespace if they sit idly by and watch a process they cannot fully understand or control? These dilemmas create real risks to the mission when AI agents are executing tasks that are misaligned to the commander’s intent and compound the potential misapplication of violence. The result—military practice that is discordant with political purpose. The essence of strategic failure.

In the frenzy of using machine augmentation to make better decisions faster, military leaders and planners cannot lose sight of their role in the command process. More critically, they cannot acquiesce human agency to machine processes under the mistaken belief that because the machine says it’s “ok,” it must be. To cede staff cognition to such processes, with expectations for greater autonomy in machine agency, undermines commander understanding and decision-making—fundamental to the American way of war. Systems (software and hardware), processes (battle rhythm), and human structures (staff organizations and command structures) must facilitate mission command. But, at the end of the day, the commander is still responsible for execution. It is at this tension that we find discrimination between 20th and 21st century command. Commanders must still inspire, and they are still ultimately responsible, but they must accept risk when empowering subordinates that is, before now, not fully explored or understood. Decision making cannot be a pure algorithmic process—despite its tempting efficiencies. Similarly, commanders must adapt processes, technology, and structures to enable collective heroism within staff structures and subordinate units—an evolution in a uniquely American way of war.