The Warfighter Exercise as a Real Stress Test
The Warfighter Exercise (WFX) is the corps's highest-stakes training event: a command post exercise conducted approximately every 18 to 24 months per corps HQ schedule, evaluated by the Mission Command Training Program (MCTP) at Fort Leavenworth, in which the corps staff executes the full Military Decision-Making Process under simulated operational conditions. Unlike Project Convergence, which is a deliberate experimentation environment where failure is instructive by design, the Warfighter is a readiness evaluation with real consequences for the corps commander's assessment.
That distinction matters for AI integration. Tools introduced into a Warfighter must function in the corps operations center without degrading the staff's ability to synchronize warfighting functions across the deep, close, and support areas in real time. A tool that slows a portion of the planning cycle—even if it improves plan quality at that step—can cascade friction across the rest of the staff. This is a harder test than any technology demonstration, and it reveals problems that controlled experiments do not.
III Armored Corps' current preparation, incorporating allied partner planners alongside AI planning tools, adds an additional layer: multinational interoperability. Coalition partners bring different classification accreditations, different command post systems, and different expectations about AI decision aid output. The planning session at Fort Cavazos on October 2 was, in effect, a rehearsal for a multinational Warfighter—the hardest variant of corps planning, and the most operationally representative.
How the MDMP Works—and Where AI Can Help
Army Doctrine Publication 5-0 (The Operations Process) and Field Manual 6-0 (Commander and Staff Organization and Operations) describe the MDMP as a seven-step iterative process: receipt of mission, mission analysis, course-of-action development, course-of-action analysis and wargaming, course-of-action comparison, course-of-action approval, and orders production. At corps level, a deliberate planning process spans 24 to 96 hours and involves hundreds of staff officers across warfighting function cells—fires, maneuver, intelligence, sustainment, protection, and command—each contributing analysis that must be synchronized into a single coherent order.
AI can meaningfully accelerate several specific steps. During mission analysis, AI tools can ingest and synthesize intelligence products, order-of-battle data, logistics statuses, and adjacent unit reports faster than human staff sections, compressing the time to build a common operational picture without sacrificing completeness. During course-of-action development, AI can model terrain, assess lines of communication, and surface maneuver options that a time-pressured staff might miss. During orders production, AI tools can generate draft synchronization matrices, populate fragments of the operations order, and cross-check annex consistency.
These are genuine accelerants. A well-integrated AI tool can compress the time available for the wargame step by getting the staff to a well-analyzed starting position faster. That compression has operational value.
Where AI faces hard limits is in the wargame itself.
The Wargaming Problem
The wargame step is not primarily an analytical process. It is a structured social process in which the commander builds shared mental models with key leaders, discovers staff assumptions about enemy behavior, and validates the plan under conditions of deliberate adversarial challenge. Its product is not a document—it is a trained staff that understands how the plan is supposed to unfold and what to do when it doesn't.
AI can support wargaming by running adversary behavior models, generating branch and sequel options, and simulating engagements at high tempo. What it cannot do is substitute for the human wargame record. When a staff officer plays the adversary and forces the friendly plan to an unexpected decision point, the discovery is memorable in a way that a simulation output is not. It changes how the operations officer thinks about the plan and how the commander assesses risk. An AI-generated adversary analysis, no matter how sophisticated, produces information rather than shared understanding.
The risk in current AI planning tool deployments is subtle. Under time pressure—which is the defining condition of every corps Warfighter—a staff looking at an AI-generated course-of-action analysis that includes modeled adversary action templates has a powerful incentive to treat that output as equivalent to a wargame finding. If the tool's interface doesn't clearly distinguish a modeling output from a wargame record, and if the tool's workflow doesn't require a formal human wargame step before producing a synchronized order, the staff can produce a technically complete OPORD that reflects AI analysis but not human wargaming. The plan looks sound until the adversary does something the model didn't anticipate.
This is not a hypothetical risk. It is a documented pattern across analytical tool adoptions in planning processes without workflow guards: the tool's output substitutes for the process step the tool was meant to support, rather than feeding into it.
The Data Pipeline Problem
AI planning tools are only as useful as the data they receive. At the corps operations center, data flows in from multiple classified networks: the Secret Internet Protocol Router Network (SIPRNet), the Global Command and Control System–Army (GCCS-A), the Command Post of the Future (CPOF), and—in multinational exercises—coalition networks operating at different classification levels and update frequencies. These systems were not designed for the kind of structured, low-latency data feeds that AI tools require to maintain a coherent operational picture.
In practice, AI tools at the corps operations center frequently operate on stale or incomplete data unless dedicated data management personnel maintain manual feeds. A course-of-action recommendation based on logistics statuses six hours old, or on an order-of-battle database that doesn't reflect the latest coalition intelligence, will be analytically valid but operationally wrong. The AI doesn't know what it doesn't have, and it won't signal the gap unless explicitly designed to do so.
Acquisition programs need to address this directly. Requirements documents for AI planning tools typically specify the tool's analytical performance—accuracy of recommendations, speed of processing, model fidelity. They less often specify the data interface contract: which feeds the tool ingests, at what latency, with what data quality checks, and—critically—what the tool's behavior is when a feed is degraded or absent. A tool that silently degrades its recommendations based on incomplete data, without surfacing that degradation to the operator, will fail in exactly the conditions where the corps most needs it.
In a denied-degraded-intermittent-limited (DDIL) communications environment—the baseline assumption for near-peer corps operations—data pipeline integrity is not a nice-to-have. It is the gating requirement for whether the AI tool functions or misleads.
Trust Calibration Across Echelons
The third integration problem is the hardest to address in acquisition documents: trust calibration. Corps planning involves not just the corps staff, but subordinate division, brigade, and battalion planners who must execute the corps order. A well-integrated AI tool at the corps operations center can compress the time to produce a corps plan—but that compressed timeline may leave subordinates with inadequate time for their own parallel planning. Speed at echelon can cascade fragility downward.
There is also the inverse problem. Subordinate formations with their own AI planning tools may optimize their plans against locally available data without fully accounting for corps synchronization requirements—deep fires timing, adjacent unit boundaries, sustainment constraints. When a division AI tool produces a course-of-action recommendation that maximizes the division's task accomplishment without fully modeling corps-level dependencies, the corps has a synchronization failure even if both AI tools functioned correctly. This is the classic emergent problem of locally optimal, globally suboptimal behavior, applied to AI-assisted planning.
Both failure modes trace to the same absence: the corps and its subordinate formations haven't established shared standards for what AI tool output means in the planning process. Does an AI-recommended course of action carry the same analytical weight as a staff-briefed option? Does it require the same wargame validation? Can a subordinate formation incorporate an AI-generated synchronization suggestion into its own plan without explicit corps approval? These are doctrinal and training requirements that acquisition programs cannot resolve on their own—but acquisition programs can create the conditions that force the conversation. A tool that requires a human to explicitly accept or reject an AI recommendation, with a recorded rationale, generates the audit trail that doctrine writers and trainers need.
What Acquisition Programs Need to Specify Now
The Warfighter exercise is the right place to develop these requirements through honest failure. The important outcome from III Armored Corps' current preparation isn't a summary of time saved or tasks accelerated. It's a detailed after-action record of where AI output was misunderstood, ignored without justification, or accepted without adequate validation—and specific interface and workflow changes that would prevent each failure mode.
Program offices can get ahead of that process by specifying three things current requirements documents typically leave open:
Data interface contracts. Define which feeds the AI tool ingests, at what latency, with what quality checks, and what the tool does when a feed is degraded or absent. Require the tool to surface data gaps to the operator rather than silently operating on incomplete information.
Wargame boundary enforcement. The tool's workflow must distinguish AI-generated adversary models from human wargame records. Require a formal human wargame record as a prerequisite before the tool produces a synchronized order template. This is an interface design requirement, not a modeling requirement—it governs what the operator must do before the AI can proceed.
Trust calibration metrics. Include in Warfighter exercise after-action reviews a structured assessment of AI tool reliance: instances where AI output was accepted without operator review, instances where it was rejected with inadequate justification, and instances where subordinate tool outputs conflicted with corps-level recommendations. These data points drive the training program that determines whether a corps can actually use AI planning tools under operational conditions, not just demonstrate them.
The Integration Problem Is Tractable
These are not unsolvable problems. They are engineering and process problems that require precise requirements, not better models. Spartan X's work on AI verification—distinguishing model outputs from validated records, surfacing confidence bounds, and enforcing human-in-the-loop checkpoints before consequential outputs are accepted—addresses exactly the wargaming boundary enforcement gap described here. The same principle applies to the DDIL data pipeline problem: edge-deployed compute with explicit data provenance and degradation signaling turns a silent failure mode into a visible one. The corps operations center doesn't need AI tools that operate confidently on incomplete data; it needs tools that report what they don't know.
Sources and further reading
- Army Doctrine Publication 5-0: The Operations Process (June 2019, change 1)
- Field Manual 6-0: Commander and Staff Organization and Operations (May 2014, change 2)
- Army Futures Command AI Integration Center / PEO IEW&S: Project Linchpin portfolio and AI planning tool program of record materials
- Mission Command Training Program (MCTP), Fort Leavenworth: Warfighter exercise overview, publicly available
- DVIDS: "III Armored Corps incorporates allies and AI into Warfighter planning," October 2, 2026 (Image ID 9998400, Staff Sgt. Christopher Neu, III Armored Corps)
- DVIDS: "2-12 FA Table 13 Training," October 6, 2026 (Image ID 9998300, Spc. Kristen Cruz, 14th Public Affairs Detachment)
