AI course-of-action analysis: wargaming COAs at staff speed

Every operational plan begins with the same question: of the options available to us, which one should we execute? The Military Decision-Making Process (MDMP) formalizes how a staff works through that question – and the wargaming step, where each course of action is played against likely enemy responses, is where most of the analytical work happens. It is also the step most vulnerable to time pressure. When the planning horizon compresses from 72 hours to 12 – as it routinely does in high-tempo operations – a staff that still requires four to six hours of manual COA wargaming will arrive at the commander's decision brief with inadequate analysis or no analysis at all. AI-assisted constructive simulation for staff planning is changing that constraint, enabling staffs to generate, wargame, and compare multiple courses of action in the time previously required for one.

The COA analysis problem: why manual wargaming breaks under time pressure

The doctrinal wargaming step asks the staff to play through each candidate COA against the enemy's most likely course of action and most dangerous course of action, recording action-reaction-counteraction triplets at each decision point, populating the decision support template, and identifying branches and sequels. Done properly, this produces an analytically grounded decision brief. Done under time pressure with a tired staff, it produces a checklist exercise that confirms the plan the S3 already preferred before the wargame started.

Three failure modes are common in manual COA wargaming under pressure. First, the staff games only the most likely enemy COA, skipping the most dangerous. Second, the wargaming proceeds at a single optimistic estimate of task completion times rather than testing sensitivity to delays and setbacks. Third, the comparison of COAs uses qualitative impressions rather than quantified outcomes from the wargame. Each failure reduces the analytical value of the process and increases the risk that the selected COA carries unidentified vulnerabilities.

AI-assisted COA analysis addresses all three failure modes by automating the computationally intensive parts of the wargame – enemy COA development, action-reaction logic, outcome quantification – while keeping human judgment at the decision points that require it: mission analysis, constraint identification, and commander guidance.

COA generation: from mission to maneuver sketch

Before wargaming can begin, there must be COAs to wargame. Generating a viable COA is itself a non-trivial staff task: the S3 must derive a task organization from available assets, assign missions to subordinate elements, establish a scheme of maneuver with phase lines and control measures, and confirm that the resulting COA satisfies doctrinal feasibility criteria – correct support ratios, deconflicted routes, realistic time-distance calculations.

AI COA generators approach this as a constraint satisfaction problem. Given the mission statement, the area of operations, the commander's critical information requirements, and the assessed enemy situation, the generator applies a library of doctrinal templates – task force structures, maneuver patterns, support allocation rules – to propose candidate COAs that satisfy the specified constraints. The output is a structured COA sketch: task organization, scheme of maneuver by phase, a list of planning assumptions, and a set of critical events whose timing drives the rest of the plan.

The generator does not produce a finished COA – it produces a starting point that the staff validates and refines. The value is speed and completeness: the AI ensures that no doctrinal constraint is accidentally violated and that the task organization arithmetic – combat power ratios, lift capacity, support allocations – is correct before the staff begins wargaming. It also surfaces options the S3 might not have considered by systematically exploring the constraint-satisfying solution space.

Integrating terrain and intelligence data

COA generation quality scales with the quality of the inputs. Modern AI COA tools ingest georeferenced terrain data – digital elevation models, road network graphs, vegetation and urban classifications – to compute terrain-dependent parameters automatically: route feasibility for vehicle types, observation and fields of fire for defensive positions, line-of-sight constraints on communication relay nodes. The enemy intelligence picture, drawn from the current intelligence preparation of the battlefield, constrains the COA generator's assumptions about where enemy forces are, what they are capable of, and what courses of action they are likely to pursue.

This integration is what separates an AI COA generator from a simple template library. A template library produces the same output regardless of terrain; a terrain-aware generator produces COAs adapted to the specific operational environment, with route recommendations that reflect actual road conditions and defensive positions sited for the specific terrain features in the area of operations.

Automated red teaming: the adversary in the wargame

The core of COA analysis is the wargame: friendly action, enemy reaction, friendly counteraction, repeat until a phase boundary or decision point is reached. In manual wargaming, the S2 plays the enemy. The quality of the wargame depends directly on how well the S2 can represent the adversary – under time pressure, with incomplete intelligence, while simultaneously managing the intelligence production cycle.

Automated red teaming replaces the S2's manual enemy play with a constrained AI agent that models adversary behavior according to the templated enemy COA and assessed capability set. The red team agent is not an unconstrained optimizer – it does not find strategies that no real adversary would employ. It operates within doctrinal bounds: it uses the terrain features the enemy is assessed to have identified, it employs the weapon systems in the enemy order of battle, and it responds to friendly actions with the counter-measures that the enemy's doctrine prescribes.

The result of each wargame iteration is a chronological log of action-reaction pairs, with each pair annotated with the decision logic that produced the enemy response. This log is the input to the decision support template: named areas of interest where intelligence collection should focus, decision points where the commander must choose between branches, and trigger conditions that indicate which enemy COA is developing. Generating this log manually from a wargame played by a tired S2 under time pressure produces incomplete and inconsistent results; automated red teaming produces a complete, consistent log in minutes.

Most dangerous enemy course of action coverage

Because automated red teaming is fast, the staff can afford to game each friendly COA against both the most likely and most dangerous enemy COAs without sacrificing the overall planning timeline. The most dangerous enemy COA – the one with the highest potential to defeat the friendly plan – is precisely where manual wargaming most often gets cut under time pressure, because it takes longest to develop and is most psychologically uncomfortable to play. AI red teaming removes both constraints: the most dangerous COA takes no longer than the most likely, and the software does not exhibit optimism bias.

Statistical outcome modeling: COA robustness under uncertainty

A deterministic wargame produces a single outcome – this COA succeeds or fails under these specific assumptions. That outcome is as uncertain as its least reliable input assumption, and in planning under uncertainty, the least reliable assumptions tend to be the most consequential: enemy strength, enemy reaction time, weather, and friendly task completion rates.

Statistical outcome modeling runs each COA through a Monte Carlo simulation – typically several hundred to a few thousand iterations – varying the uncertain input parameters across their assessed probability distributions. The result is not a single outcome but a distribution: win probability for the overall operation, a distribution of time-to-objective, a confidence interval on logistic consumption, a probability that a particular phase completes before the enemy can reinforce. This distributional output gives the commander and staff a quantified measure of COA robustness rather than a point estimate.

The comparison of robustness distributions across COAs is the analytical contribution that AI-assisted analysis makes most clearly. It is straightforward to claim, in a manual comparison, that COA 1 is more flexible than COA 2. It is qualitatively different to show that COA 1 achieves the mission with 78% probability across the uncertainty range while COA 2 achieves it with 61% – with COA 2's advantage concentrated in the narrow favorable scenario the planning staff preferred.

Staff decision support: from analysis to brief

The output of COA analysis must reach the commander in a form that supports decision-making. In the traditional MDMP, this means the decision brief: the S3 presents each COA with its wargaming results, the S2 presents the enemy analysis, and the staff recommends a COA based on a comparison matrix weighted by the commander's criteria.

AI decision support tools integrate directly into this brief preparation. The COA comparison matrix is populated automatically from the wargaming outputs – criteria scores are derived from the quantified analysis rather than from staff impressions. The decision support template is generated from the wargame log. Natural-language summaries of each COA's key advantages, vulnerabilities, and critical assumptions are drafted by the AI and reviewed by the S3 before the brief. The commander receives a brief grounded in quantified analysis rather than in the qualitative impressions of a staff that has been awake for twenty hours.

The staff's role shifts from computation to judgment. Officers who previously spent hours doing arithmetic – time-distance calculations, support ratio checks, synchronization matrix population – now spend that time on the questions where professional judgment is irreplaceable: is the intelligence picture reliable enough to support this COA? Does the commander's intent, as stated, actually align with the recommended COA? Are the critical assumptions realistic given what we know about this adversary from operational experience? The AI handles the computation; the staff handles the reasoning.

Integration with AI OpFor wargaming and simulation

COA analysis AI does not operate in isolation. In the most capable implementations, the COA wargaming engine is connected to the same simulation environment used for exercise play – the same terrain model, the same unit capability database, the same adversary behavioral model. This means the automated red team in the COA analysis step and the AI OpFor system in the training exercise use consistent models of adversary behavior, producing a coherent thread from planning to execution.

When the plan is executed in a constructive simulation exercise, the decision points identified during COA analysis become the trigger conditions monitored by the exercise control system. The branches identified during wargaming become the contingency branches the staff must recognize and execute during the exercise. The after-action review can compare what the COA analysis predicted with what the simulation produced, creating a feedback loop that improves both the staff's planning proficiency and the fidelity of the AI models used in the next planning cycle.

This integration also supports the multi-domain operations context. A multi-domain wargaming environment spans land, air, maritime, space, and cyber effects; a COA that looks sound in a land-only model may fail when air defense suppression timelines or cyber effect windows are introduced as constraints. AI COA tools integrated with a multi-domain simulation can surface these cross-domain timing conflicts during the planning step, not after the exercise has exposed them.

Implementation considerations for defense organizations

Deploying AI COA analysis in a real staff environment requires attention to several integration requirements that are not always apparent from capability demonstrations. First, the tool must be operable in a disconnected or degraded network environment – operational planning does not wait for cloud connectivity. Edge-deployable inference and a local terrain and intelligence data store are requirements, not enhancements.

Second, the tool must export its outputs in formats that integrate with the staff's existing workflow: standard military graphics formats for COA sketches, spreadsheet-compatible formats for comparison matrices, and plain-language summaries that can be incorporated into the OPLAN or OPORD without reformatting. A tool that requires the staff to re-enter its outputs into a separate system will be bypassed under time pressure.

Third, and most fundamentally, the tool must be trusted. Trust in AI decision support tools is built through transparency – the staff must be able to see the logic behind every recommendation, trace every output to its inputs, and override any AI-generated result without friction. A tool that produces correct outputs through an opaque process will not be trusted; a tool whose reasoning is legible to a senior officer will be used and will improve planning outcomes.