Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide.
Generative AI is here, but we don’t quite know what to do with it. Across both industry and the military, enthusiasm has outpaced results. Companies report rising rates of abandoned AI projects, while defense organizations often repeat the mantra of decision advantage without offering clear pathways for implementation. The problem is rarely technical. Instead, success depends on whether institutions are willing to adapt their structures and processes to make new tools work.
The U.S. military’s joint planning process is a prime example. The joint planning process is the U.S. military’s doctrinal approach to developing logical solutions to complex problems. Operational planning teams were designed before the widespread adoption of generative AI, and without deliberate adjustments they risk relegating powerful tools to the margins. Recent experiments at the Marine Corps Command and Staff College, where students integrated a large language model into multiple planning exercises, shed light on how these tools can shape mission analysis, team collaboration, and subject-matter expertise.
To explain this, first we describe how students actually used generative AI in planning, and where the tools made the greatest impact. Second, we draw out lessons on when AI supports creative thinking and when it is sidelined by group dynamics or expertise. Finally, we offer practical recommendations for restructuring operational planning teams: redesigning workflows, drafting prompt templates, building training pipelines, and integrating checkpoints that ensure AI is more than an optional add-on. To get the most out of generative AI, the operational planning team itself should change.
In its 2024–2025 academic year, the Marine Command and Staff College conducted five planning exercises. Throughout four of those planning exercises, a small number of students were given access to a model-agnostic large language model tool. This tool gave students experience employing large language models in operational planning and it gave the school the ability to investigate how best to incorporate AI into the curriculum. In this paper, we (one faculty member and one graduate) provide the lessons we learned in response to both of these questions: How should students incorporate the tool into planning? How should the faculty incorporate AI into the planning curriculum?
Broadly, through self-reporting via surveys and through the vendor’s ability to observe usage rates, we determined that planning team members were far more likely to use the tool early in the joint planning process and especially during the mission analysis step. Additionally, even though we saw a decrease in usage rates later in the planning process, for those who did use the tool in later steps, they self-reported higher quality answers to their prompts. This suggests, perhaps, that a self-selecting subset of the group persisted in using the tool, and through that persistence, obtained better results.
Lessons Learned
An operational planning team is a temporary group of officers and staff brought together to solve a specific planning problem. They are usually formed during the joint planning process and include members from different functional areas (operations, intelligence, logistics, communications, and others) so that all perspectives are represented. Their job is to analyze the mission, develop courses of action, compare options, and produce recommendations for the commander. In plain terms: they are ad hoc planning groups that pull expertise from across a staff to design and refine a plan for a particular operation. The fundamental insight under which each of the specific lessons nests is unsurprising, but important: If you want to maximize AI performance in your operational planning team, you’re going to need to restructure your operational planning team.
Lesson 1: Large Language Model Value Is Most Obvious During Divergent Thinking
In the field of creativity, convergent and divergent thinking represent two different means of developing ideas. This distinction is captured in the Marine Corps’ planning publications and the Joint Officer Handbook: Divergent thinking develops new insights and ideas, while convergent thinking aims at arranging those ideas into coherent categories.
The iterative process of employing divergent thinking to broaden the scope of possible solutions, then convergent thinking to derive a solution, takes place throughout the planning process. But divergent thinking is most clearly identifiable in the mission analysis step of the joint planning process before the operational planning team organizes those new ideas and insights into logical groups (especially in course of action development and analysis).
At Marine Command and Staff College, we found that students were more likely to use the large language model tool during the mission analysis phase — which corresponds to divergent thinking — than during the later steps of the joint planning process.
This finding is closely linked with the second. We will introduce the second lesson learned and then provide a recommendation that equally applies to Lessons 1 and 2.
Lesson 2: Large Language Model Value Is Obvious Outside a Member’s Subject Matter Expertise
Large language model failures are well-known and well-documented. Students at Marine Command and Staff were well aware that large language models have epistemic limits and that they hallucinate. One outcome of this latent knowledge is that students were more likely to use the tool when operating outside their area of expertise.
In one instance, an Army artillery officer told us that when, for the purposes of the exercise, he was placed in a billet outside his artillery area of expertise — say, as a sustainment or information planner — he was much more eager to use the tool. He expected that the tool would be able to provide him basic knowledge from that unfamiliar career field better than it would be able to provide him more advanced knowledge from within his own field.
In the planning exercises earlier in the academic year, students were assigned to small operational planning teams and, as a result, were often asked to fill a billet that did not align with their military specialty. In the largest and final planning scenario of the academic year, the planning staffs were much larger, allowing many students to be assigned within their area of professional expertise. In these cases, the perceived need to consult generative AI for technical clarification or external insights was low. When students did consult the tool, the outputs were often seen as redundant rather than complementary.
This observation conforms with insights from outside the military: large language models perform better at general knowledge tasks than they do at domain-specific knowledge tasks. An expert is less likely to gain novel insights at the expert level than a user is likely to get at the novice level.
This was not a failure of the tools but a reflection of how to frame and encourage AI adoption.
This phenomenon underscores a key insight: AI adoption requires leaders to define specific use cases where the tool offers clear advantages, whether by accelerating consensus-building, expanding the range of planning options, or revealing overlooked historical parallels.
In response to Lessons 1 and 2, we recommend designing workflows within the operational planning team that either require or encourage team members to compare their analysis to AI-generated recommendations during all phases of the planning process. Operational planning team leads should consider developing in advance draft prompt templates team members might use to evaluate their own work. Without structured prompts, even the most advanced tools will be underused.
Lesson 3: Enthusiasm Without Experience Is Not Enough
During the exercise, many participants were eager to explore what generative AI could offer. However, with limited prior experience with the specific tool, and with generative AI more generally, most struggled to create meaningful outputs. Enthusiasm quickly gave way to frustration when the pathway from prompt to useful product was unclear.
The relationship between limitations in AI talent and ineffective implementation is a well-known problem in industry. Initial curiosity should be matched by progressive training, mentorship, and clear examples of success. Otherwise, the momentum of early adoption stalls. Successful AI integration in industry often depends less on technical capacity than on sustained workforce development and trust-building. In other words, merely procuring the tool and providing access to team members is not enough to enable success.
In response to lesson 3, we recommend pairing AI tools with structured onboarding, scenario-based training, and peer coaching to build confidence and competence. Over time, these supports can transition from prescriptive checklists to more organic use.
In the absence of AI expertise within the organization, leaders should consider devoting resources to upskilling a small percentage of the team so those AI-capable team members can conduct the onboarding, scenario-based training, and peer coaching required for success.
Lesson 4: Small Group Dynamics Shape Behavior
One important finding was that small group collaboration often displaced the need to use generative AI tools. Students naturally turned to one another to resolve questions, debate options, and refine products. This tendency reinforces the principle that AI should be integrated into the fabric of collaborative work, not layered on top as an optional enhancement.
In innovation terms, this is an example of the structure innovation that often must accompany the product performance innovation — structure innovation refers to how you “how you organize and align your talent and assets.” In the absence of such integration of talent and tools, human relationships and established processes will dominate attempts to introduce new technological capabilities.
Develop planning templates and facilitation guides that require teams to incorporate AI outputs at designated checkpoints, such as initial mission analysis, course of action comparison, and risk assessment, ensuring the model becomes integrated with the team, rather than a mere curiosity.
Conclusion
At the Marine Command and Staff College, successful employment of the AI tool was never measured by whether students developed flawless plans. It was measured by whether they wrestled with a new tool, made mistakes, and began to see what generative AI might add to the art of planning. The most important discovery was not about the limits of the technology, but about the limits of the team. Simply handing a planning group access to a model produced flashes of insight, but those insights faded unless the group itself was reorganized to take advantage of them.
That reorganization requires more than enthusiasm. It requires writing prompt templates into the early stages of planning, training students not just to “try the tool” but to build skill and confidence with it, and carving out deliberate checkpoints where AI is required to test assumptions and sharpen ideas. These adjustments sound small, but together they signal a shift: The operational planning team is no longer a static structure but a malleable system, one that can be re-engineered for an age of human-machine collaboration.
The lesson is clear: AI on its own does not transform planning — it transforms teams who are willing to change how they work. Leaders who want more than experiments ought to be prepared to redesign their planning groups so that human judgment and machine-generated insight reinforce one another. Otherwise, generative AI will remain a curiosity at the margins of doctrine. But with deliberate restructuring, teams can take advantage of the insights AI offers.
Joseph O. Chapa is a U.S. Air Force officer who holds a Ph.D. in philosophy from Oxford. He’s published several times in War on the Rocks as well in other outlets such as the New York Times. His book, Is Remote Warfare Moral? was published with PublicAffairs (an imprint of Hachette) in 2022. He also publishes a weekly newsletter on philosophy and technology.
Sofia Cantu is a U.S. Army officer and air defender with experience in tactical air defense operations and staff planning. She holds a Master of military studies from the Marine Corps University, where her thesis analyzed AI-enabled decision support systems, workforce integration, and ethical considerations in Army modernization efforts.
The views expressed are those of the authors and do not necessarily reflect those of the U.S. Air Force, the U.S. Army, the Department of Defense, or any part of the U.S. government.
**Please note, as a matter of house style War on the Rocks will not use a different name for the U.S. Department of Defense until and unless the name is changed by statute by the U.S. Congress.
Image: Gunnery Sgt. Daniel Wetzel via U.S. Marines
