ABSTRACT

Across the United States defense enterprise, policy, doctrine, and risk frameworks now converge on the proposition that generative AI can accelerate staff work only when embedded inside redesigned processes, workforce pipelines, and governance mechanisms; authoritative anchors include the Department of Defense (DoD) Data, Analytics, and Artificial Intelligence Adoption Strategy, November 2, 2023, the Chief Digital and Artificial Intelligence Office (CDAO) Generative AI Toolkit and “Guidelines and Guardrails” operationalization, December 11, 2024, the CDAO Task Force Lima Executive Summary, December 11, 2024, and the NIST AI Risk Management Framework 1.0, January 26, 2023 with its generative profile, NIST AI 600-1, July 26, 2024.

Doctrinally, the Joint Planning architecture codified by the Joint Chiefs of Staff in JP 5-0: Joint Planning (doctrine portal) and the Marine Corps’ MCDP 5: Planning (October 4, 2018) provides the baseline workflow—mission analysis, course of action development, comparison, and decision—into which generative systems must be instrumented. Empirical and analytic studies over 2024–2025 by the RAND Corporation reinforce the organizational preconditions for effective human–machine teaming, e.g., Exploring AI to Mitigate Bias in IPB, August 6, 2024, Improving Sense-Making with AI, 2025, One Team, One Fight: Human-Machine Integration, 2025, and An AI Revolution in Military Affairs, 2025, each concluding that gains arise from targeted task design, measurement, and trust-building rather than from tool access alone.

Within defense education, the Marine Corps University has articulated transformation lines of effort for AY 2026–2029 that encompass faculty development and modernization of professional military education delivery, as recorded in the EDCOM/MCU Campaign Plan 2026–2029, May 20, 2025; while not a generative AI manual, this plan foregrounds curriculum adaptation and assessment mechanisms that align with the institutional conditions necessary to integrate new analytical tools into planning pedagogy. The Marine Corps doctrinal corpus that frames staff workflows—the MCWP 5-1/5-10 Marine Corps Planning Process and MCWP 5-10 display page—defines mission analysis as a hypothesis-generating stage and distinguishes conceptual design from procedural sequencing, an alignment that makes the divergent-thinking steps of planning the most receptive insertion points for large language models. In parallel, the DoD has codified cyber and assurance expectations for AI capabilities in the AI Cybersecurity Risk Management Tailoring Guide, July 14, 2025, clarifying that deployment must be paired with verifiable controls, test artifacts, and traceable governance—requirements that directly affect how operational planning teams prepare, store, and audit model-assisted artifacts.

Macro-level diffusion evidence in the broader economy provides external validity for adoption patterns observed in defense contexts. The OECD’s cross-country enterprise surveys report uneven but accelerating uptake, with sectoral and size-class divides that mirror capability and skills gaps; see OECD Digital Economy Outlook 2024, Volume 1, May 14, 2024, Fostering an Inclusive Digital Transformation as AI Spreads Among Firms, October 25, 2024, and Emerging Divides in the Transition to Artificial Intelligence, June 23, 2025. The OECD’s book-length study The Adoption of Artificial Intelligence in Firms, May 2, 2025, with the associated PDF full report, documents managerial hurdles—skills, data readiness, and change management—consistent with the conclusion that institutional adaptation determines returns. Workforce-skills implications quantified by the World Economic Forum’s Future of Jobs Report 2025, January 7, 2025 corroborate rising demand for “creative thinking” and “AI and big data” capabilities, mapping directly onto mission-analysis needs in joint planning. Risk governance guidance from NIST—the base AI RMF 1.0, January 26, 2023 and the generative profile NIST AI 600-1, July 26, 2024—specifies control families for validity, robustness, transparency, and bias management; these can be operationalized as explicit planning-cell checkpoints to log model prompts, retain rationales, and subject outputs to adversarial review.

For line organizations, the practical bridge from policy to practice has been clarified by CDAO artifacts, including the DoD Compliance Plan for OMB M-24-10, September 19, 2024 and the Statement on DoD’s Compliance, September 24, 2024, both of which reference generative guidelines and testing infrastructures that can be adapted to educational planning cells and operational staffs. The DoD’s public communication “DoD Releases AI Adoption Strategy,” November 2, 2023 and the accompanying Strategy Fact Sheet, October 25, 2024 emphasize an “AI-ready workforce,” data foundations, and scaled deployment—organizational vectors that align with staff-planning realities where multiple warfighting functions must coordinate under time pressure. In military pedagogy and research, Marine Corps University Press analyses—e.g., Automation and the Future of Command and Control (JAMS 11-1)—and Design-thinking and innovation essays interrogate how organizational structure and doctrine mediate technology effects, supporting the proposition that generative AI’s value turns on workflow integration, not tool novelty.

Within this verified landscape, four operational lessons emerge with direct applicability to joint planning teams.

• First, divergent-thinking phases such as mission analysis are the natural insertion points for large language models; this aligns with the conceptual design emphasis in MCDP 5 and with RAND findings that AI can expand hypothesis spaces and contextual frames in early sense-making (Improving Sense-Making with AI, 2025).
• Second, the biggest marginal gains accrue outside a planner’s core specialty, consistent with enterprise evidence that AI adoption disproportionately benefits organizations with complementary skills and data readiness (OECD, May 2, 2025) and with planning research on bias mitigation in IPB (RAND, August 6, 2024).
• Third, enthusiasm unbacked by structured onboarding typically stalls; public DoD implementation artifacts establish training pipelines, testing infrastructure (JATIC references), and governance checkpoints as preconditions for deployment (CDAO resources hub).
• Fourth, small-group dynamics can displace tool use unless facilitation forces incorporation; doctrinal guidance and NIST risk profiles together justify mandatory checkpoints where teams must compare human analysis to logged model outputs with traceable rationales (NIST AI RMF resources).

Claims about specific classroom usage rates or proprietary telemetry from internal educational pilots at Marine Corps University or Marine Corps Command and Staff College are not available in public, citable institutional documents as of October 14, 2025; where such data are necessary to support numerical assertions, No verified public source available. Accordingly, this analysis confines itself to publicly documented policy, doctrine, governance, and research that together delineate how generative AI should be embedded into the JP 5-0 planning flow.

The verified corpus enables practical prescriptions: designate an “AI analyst” within each warfighting function; pre-publish prompt templates mapped to doctrine paragraphs; require mission-analysis, COA development, and risk-assessment checkpoints that log prompts, responses, and evaluator notes; adopt NIST’s validity, robustness, and transparency controls as acceptance criteria for staff products; instrument education programs per the EDCOM/MCU campaign plan to sequence training from prescriptive to adaptive tasks; and measure adoption with explicit artifacts—prompt logs, survey instruments, cross-case comparisons—linking to governance documentation under the CDAO umbrella. These measures operationalize the consistent message across DoD, NIST, OECD, and RAND outputs: planning teams must change their structure and process to convert generative tools into repeatable decision advantage.

---

Plain-Language Summary for Non-Experts

Generative artificial intelligence is a tool that can produce text, lists, and options when it is given a clear prompt. Military planning teams can use it to speed up parts of their work. They must also control how and when they use it. This chapter explains the basics, shows where the tool helps, shows where it fails, and lists the controls that keep it safe and accountable. Every claim below links to an official, public source.

Generative artificial intelligence will be shortened here to AI. A planning “team” means a small group of officers and staff who prepare options for a commander. “Mission analysis” means the early part of planning when a team defines the problem and lists what might work.

The United States military has written rules for planning. The joint rulebook is JP 5-0: Joint Planning from the Joint Chiefs of Staff. The Marine Corps version is MCDP 5: Planning. These books say teams should analyze the mission, create different courses of action, compare them, and then write orders. They also say teams must challenge their own assumptions. These points are important, because they match what AI can do well when teams ask for many options early and then check those options later.

The National Institute of Standards and Technology publishes public guidance on AI risks that any organization can use. The main guide is NIST AI Risk Management Framework 1.0 (January 2023). It tells teams to govern, map, measure, and manage risks. There is also a profile just for generative systems: NIST AI 600-1: Generative AI Profile (July 2024). That profile says organizations should log prompts and outputs, track model versions, check for changes in data or behavior, and provide transparency so people can review what the system did. These steps fit military planning, because they create a record that can be audited.

The Department of Defense has a public toolkit for generative systems that explains how to make projects safe and reviewable. The toolkit is posted by the Chief Digital and Artificial Intelligence Office and is called Generative AI Version 1.0: Responsible AI Toolkit (December 2024). The CDAO also explains the toolkit in a public blog note, “GenAI Toolkit operationalizes Guidelines & Guardrails memo,” December 11, 2024. The toolkit contains simple checklists: suitability (is the task right for the tool), feasibility (is the data and setup in place), and advisability (is it acceptable to use the tool here). It also suggests how to record use and how to train users.

Independent research backs up these steps. The RAND Corporation studied how soldiers and algorithms work together. One Team, One Fight: Volume I (June 2, 2025) explains common problems when teams add an automated teammate. Examples include unclear roles, too much trust in automation, too little trust due to unclear outputs, and confusion about who owns the final decision. The report recommends clear role boundaries, short explanations for outputs, and set routines for when the tool can act. The full PDF is also public: RAND RRA2764-1 PDF. Another RAND study shows how AI can help human analysts avoid bias during intelligence preparation. It is Exploring AI Use to Mitigate Potential Human Bias Within U.S. Army IPB (August 6, 2024) and the PDF is here. That report says the best use is to require the tool to produce alternative views that the team must consider. The human team still decides, but the process becomes more complete.

These official and research sources point to six plain lessons.

First lesson: use the tool early to broaden thinking, then check it. The early step of planning is called mission analysis. Teams try to ask the right questions and to list many possible paths. At that point, the tool can quickly generate several different versions of the problem, several sets of questions, and several options. Doctrine supports this kind of broad thinking. See JP 5-0 and MCDP 5. But the team must not accept these outputs as fact. It must review and edit them. It must also log what was prompted and what was accepted. The NIST profile says to keep those records and to track the system version and context. See NIST AI 600-1.

Second lesson: do not use the tool to replace deep, technical judgment. People with many years of training in a narrow area, such as artillery or logistics planning, will often find that the tool returns answers they already know, or answers that are not precise enough. The RAND studies caution against depending on the tool for final decisions. See One Team, One Fight (2025). They also show where the tool adds value: when the team needs alternatives to avoid bias. See AI and Bias in IPB (2024).

Third lesson: training and simple rules matter more than excitement. The CDAO toolkit is practical. It has questions that any project lead or team member can answer. The checklists convert good intentions into action. See Generative AI Version 1.0: Responsible AI Toolkit (December 2024). The NIST framework lays out the basic loop: govern, map, measure, manage. See AI RMF 1.0 (January 2023). The takeaway is simple: give users short modules on how to ask good prompts, how to spot common errors, and how to record what they did. Then require them to use a small set of forms at key times.

Fourth lesson: small groups need forcing points or they will ignore the tool. In real staff rooms, people default to known habits. Senior voices often shape the plan. If using the tool is optional, the group will skip it under time pressure. The fix is to add required checkpoints to the process. For example, after the first draft of mission analysis, the facilitator stops the meeting and requires the AI output to be shown next to the human draft. The group must then write down accept, modify, or reject, and why. A second checkpoint can be used before finalizing risk and branches. A third checkpoint can be used before sending orders forward. These steps match “govern” and “manage” in NIST AI RMF 1.0 and the practical forms in the **CDAO** toolkit. They also reduce the trust problems seen in human–machine teams, as shown in One Team, One Fight (2025).

Fifth lesson: assign clear roles so work is reviewed before it enters official products. Teams can add four roles. An AI Analyst writes and runs prompts. A Prompt Librarian stores and updates standard prompts for the most common tasks, with dates and version numbers. An Output Critic checks every AI result and either rejects it or marks the parts that can be used, with a short reason. A Telemetry Custodian keeps the logs: prompts, outputs, edits, accept or reject, and the time spent. These are simple names and simple duties. They mirror public guidance that says roles and audit trails reduce error and improve trust. See NIST AI 600-1, AI RMF 1.0, and CDAO toolkit.

Sixth lesson: keep the records so independent reviewers can check what happened. Every time the tool is used, the team should save the prompt, the output, the model version, the edits, and the final choice. This allows audits, lessons learned, and testing. Public writing on test and evaluation in defense shows why this matters: it connects use to accountability and to the law of armed conflict. See “The Practical Role of ‘Test and Evaluation’ in Military AI,” Lawfare, October 7, 2025.

The simple rule that ties these lessons together is this: use AI to propose, require people to challenge, and save a trace of what was done.

The next part gives concrete, non-technical examples that match public reporting and open sources.

Example 1: early-stage options in a planning room. A team is preparing options for air defense and logistics support. The facilitator asks the AI Analyst to run a standard prompt to list three different ways to phase the operation and five key questions for each path. The Output Critic then removes items that do not match theater guidance from JP 5-0, which sets how joint operations are planned. See JP 5-0. The team saves the prompt, the output, and the edits. This matches the NIST call to log and manage risk. See AI RMF 1.0.

Example 2: structured bias checks in intelligence preparation. Analysts often anchor on their first view. The RAND report on bias in the U.S. Army’s intelligence preparation process recommends that a tool generate alternative views that challenge the first view. See Exploring AI Use to Mitigate Potential Human Bias Within U.S. Army IPB (August 6, 2024) and PDF. A practical version is this: the AI Analyst runs a prompt that asks for five reasons the first assessment might be wrong, grouped by logistics, timing, terrain, deception, or information effects. The Output Critic marks which points are plausible and sends them to the team to check with other data. The record goes into the logs. This is a low-tech step that makes the process more complete.

Example 3: real-world conflict shows why speed and trace matter. Open sources show that uncrewed aircraft and counter-drone measures are central in the war in Ukraine. NATO publishes regular material about support to Ukraine and air and missile defense efforts. See the public summary “NATO’s support for Ukraine” and the Secretary General Annual Report 2024 (April 26, 2025). For technical trends and tactics, NATO’s lessons community has shared open papers that describe widespread use of reconnaissance and strike drones by both sides. One example is “Tactical Developments in the Third Year of the Russo-Ukrainian War” (2025). In this kind of environment, planning teams benefit from a tool that can quickly list alternative paths and risks. But they must also keep proof of why an option was accepted or rejected. That proof supports later reviews and helps leaders understand trade-offs. The Lawfare article explains why test and evaluation connects technology to lawful and responsible use. See Lawfare, October 7, 2025.

Example 4: what adoption looks like outside the military. Governments and firms are adopting AI, but the rate is uneven. The Organisation for Economic Co-operation and Development has published new data on use by firms and the public sector. See “The Adoption of Artificial Intelligence in Firms” (May 2, 2025) and the PDF report. Another OECD study shows gaps in adoption and notes 8.3% of firms in the United States reported using AI to produce goods or services in April 2025. See “Emerging divides in the transition to artificial intelligence” (June 23, 2025). Public-sector use across countries is also tracked. See “Governing with Artificial Intelligence” (September 18, 2025). These civilian findings support the point that adoption requires training, clear roles, and management attention, not just access to a tool.

With the examples in mind, the final parts of this summary list the practical steps any planning team can take, written in simple terms and mapped to public guidance.

Practical step 1: create a short list of standard prompts, and give each one a version number. Store prompts for common tasks: framing the problem, listing assumptions, stress testing assumptions, sketching branches, and outlining risks. The Prompt Librarian publishes the list with a date. When a prompt changes, update the version number. This matches the NIST call for transparency and version control. See NIST AI 600-1.

Practical step 2: add required checkpoints to the timeline. After the first mission analysis draft, the team must run the framing and assumptions prompts, compare results, and write down accept, modify, or reject. Before finalizing courses of action, the team must run the stress-test prompt and write down what changed. Before the risk register is final, the team must run the risk prompt. Without these records, the plan cannot move forward. This matches the “govern” and “manage” steps in AI RMF 1.0 and the forms in the **CDAO** toolkit.

Practical step 3: keep human review in charge, every time. The Output Critic must sign off on any AI text before it enters an official product. This aligns with the human–machine teaming guidance in One Team, One Fight (2025). It also aligns with the CDAO rule that tools should be suitable, feasible, and advisable for the task. See Generative AI Version 1.0 (December 2024).

Practical step 4: teach the basics to all users in short blocks. The training should cover safe prompting, common failure signs (like confident but wrong answers), how to log work, and when to stop using the tool and switch back to human-only drafting. The CDAO materials include practical checklists. See CDAO blog index (August 13, 2025) and the toolkit PDF.

Practical step 5: measure what matters. The Telemetry Custodian should report four things after each planning cycle: how many prompts were used, how much human editing was needed, how often outputs were accepted or rejected, and how long it took to get a usable draft. Use these numbers to update prompts and training. This is consistent with “measure” in AI RMF 1.0.

Practical step 6: prepare for review. Save the prompt, the output, the edits, and the reason for the final choice. If an exercise or operation is reviewed later, these records help explain the decision path. The Lawfare article explains how test and evaluation connects these records to responsible use. See October 7, 2025.

The final part explains why this matters to the public, to elected officials, and to people who read and share news.

Why it matters to the public. Good planning reduces mistakes in operations. If AI is used without records, the public cannot know whether it helped or harmed. If it is used with clear roles, checkpoints, and logs, leaders can show that people stayed in control. This improves trust.

Why it matters to elected officials. Budgets and laws must match how AI is actually used. Officials can ask for simple things: show the standard prompts with version numbers; show the checkpoint forms; show the logs from the last three planning cycles; and show how training changed prompt quality over time. These are straightforward asks that match public guidance in NIST AI RMF 1.0, NIST AI 600-1, and CDAO materials.

Why it matters to readers on social media. Many posts either hype AI or fear it. The truth is more practical. The tool can help a team list options quickly. It cannot replace expert judgment. It must be used at the right time, reviewed by people, and recorded. Public sources show steady steps toward safer use: joint doctrine for planning, NIST risk rules for logging and review, CDAO checklists, and research on how to avoid over-trust and under-trust. See JP 5-0, MCDP 5, NIST AI RMF 1.0, NIST AI 600-1, CDAO toolkit, RAND One Team, One Fight (2025), and RAND IPB bias report (2024).

Two closing facts show the wider context. Public OECD reports confirm that adoption outside defense is real but uneven across sectors and countries. See “The Adoption of Artificial Intelligence in Firms” (May 2, 2025) and “Emerging divides in the transition to artificial intelligence” (June 23, 2025). Current conflict reporting shows sustained use of drones and counter-drone actions, which increases the need for fast but reviewable planning. See NATO’s support for Ukraine and NATO Secretary General Annual Report 2024 (April 26, 2025), and the open lessons paper “Tactical Developments in the Third Year of the Russo-Ukrainian War” (2025).

In plain terms, the public record supports a simple, balanced view. AI can help planning teams think broadly at the start, but people must stay in charge at the end. Good process makes that possible: use the tool at fixed times, give people specific roles, collect short written reasons for accept or reject, and save the records. Official guides from NIST, the CDAO, and joint doctrine show how to do this. Independent research from RAND confirms what works inside teams. Real-world cases show why speed and trace both matter. When these parts are in place, planning is faster, clearer, and more reviewable. When they are missing, planning can be slower, more biased, or harder to audit. The public can judge progress by asking for the basics: standard prompts with version numbers, completed checkpoint forms, logs from recent cycles, and short training plans tied to those prompts. These are straightforward items that any responsible program should be ready to share, within security limits, today.

---

Doctrinal Baselines and Governance Anchors for Generative AI in Joint Planning

The United States Joint Planning architecture—anchored in JP 5-0: Joint Planning (current version) and interpreted through component service doctrine—remains the canonical scaffolding upon which operational planning teams (OPTs) operate. The planning cycle sequentially moves through mission analysis, course of action (COA) development, COA comparison / wargaming, orders development, and transition to execution. Doctrine treats mission analysis as a “design-oriented” phase in which planners frame the operational problem, define objectives, develop initial hypotheses, identify constraints and assumptions, and perform staff estimations of risk, grounds for judgments, and decision criteria. This design locus is doctrinally suited to generating hypotheses and framing solution spaces.

Within the Department of Defense, the 2023 Data, Analytics, and Artificial Intelligence Adoption Strategy—issued November 2, 2023—established a strategic commitment to reshape organizational enablers so that data, analytics, and AI can deliver decision advantage across military and business domains. The strategy prescribes a cascading “AI Hierarchy of Needs” comprising quality data, governance, analytics/metrics, assurance, and responsible AI. It also articulates five decision-advantage outcomes: superior battlespace awareness, adaptive force planning, fast-resilient kill chains, resilient sustainment, and efficient enterprise operations. (Verified via DoD press release and official factsheet) (U.S. Department of War)

That strategy is operationalized through the Chief Digital and Artificial Intelligence Office (CDAO), which in turn deploys oversight via the CDAO Council coordinating across DoD components, linking to the Deputy Secretary’s management bodies, innovation steering groups, and workforce councils. (Verified via DoD and implementation commentary) (govCDOiq.org)

Foundational to trustworthy adoption, the National Institute of Standards and Technology (NIST) published AI Risk Management Framework 1.0, formalizing core functions (Govern, Map, Measure, Manage) and providing “profiles” enabling domain-specific tailoring. (Verified via NIST) (nvlpubs.nist.gov) In July 2024, NIST released NIST AI 600-1, a generative-AI profile companion document expanding risk considerations applicable to large language models and generative systems. (Verified via NIST) (nvlpubs.nist.gov)

These strategic and risk frameworks provide doctrinal guardrails: any adoption of generative AI within joint planning must align to DoD’s governance, data, and assurance expectations—and must survive scrutiny under NIST’s trustworthiness dimensions (validity, robustness, interpretability, privacy, fairness, accountability). The planning cell is thus not free to “bolt on” a generative model; it must inherit constraints from this layered governance architecture.

Operationally, doctrine does not yet prescribe how to embed AI into joint planning steps; generative AI is not a doctrinal actor. That gap opens an institutional responsibility: planning organizations must interpret how to apply generative systems under the existing architecture, while bridging the demands of assurance, auditability, trace logging, and chain-of-custody for AI artifacts. Thus the planning cell becomes a node in a broader governance and compliance topology.

To ensure compliance, any AI-assisted planning cell must integrate at least four structural overlays.

• First, prompt logs, versioning, and rationale tracing must be maintained in a controlled collaborative environment conforming to DoD security, auditing, and data classification policies.Each prompt → response → user revision path must be tracked to satisfy accountability.
• Second, the cell’s workflows must align with DoD’s “adopt-buy-create” model, which the 2023 Strategy uses to determine whether to reuse existing shared AI-aware capabilities, to acquire commercial AI services, or to develop new ones in-house—subject to data rights, commonality, and governance constraints. (Crowell & Moring – Home)
• Third, cells must incorporate risk-management checkpoints mapped to NIST’s four functions. For example, before accepting any AI-generated alternative COA, the cell must execute Map (identify risks, model assumptions), Measure (evaluate the model output’s confidence bounds or validation metrics), and Manage (apply mitigations, adversarial review, fallback to human judgment).
• Fourth, senior leadership must incorporate a govern function into the planning governance—i.e., oversight roles or councils must have visibility into when, why, and how a generative AI was used in plan derivation, enabling ex post review or red-teaming if necessary.

Given this environment, the planning cell’s adoption must also honor DoD’s interoperability and federated infrastructure mandates. The 2023 Strategy emphasizes federated data architectures, shared data services, and common analytics interoperability—designed to reduce silos. (U.S. Department of War) Planning artifacts, AI model outputs, intermediate knowledge, and rationale logs must be stored, shared, and versioned in a federated but secure infrastructure—simultaneously accessible across joint staff echelons yet conformant to classification and compartmentalization constraints.

Further, planning cell designs must respect DoD’s expectation of continuous learning and feedback loops. The 2023 Strategy emphasizes agile development and iterative deployment, under which AI capabilities are matured through deployment, feedback, and refinement cycles. (U.S. Department of War) That implies that any AI-enabled planning cell must not treat the model as static; instead, cells should schedule retrospective reviews of prompt effectiveness, error modes, and modification of prompt templates across planning cycles.

In the absence of public, verified open documentation citing the internal success rates or usage percentages from Marine Corps Command and Staff College experiments or specific planning cell pilots, there is no validated numeric anchor on how many planning cells currently use generative AI or their performance delta. No verified public source available.

Doctrinally, adopting generative AI in planning must also contend with the ethical and liability principles already endorsed by Department of Defense, e.g. the DoD’s Ethical Principles for Artificial Intelligence (responsible, equitable, traceable, reliable, governable), originally promulgated in 2020, which remain relevant guardrails. (Verified via DoD AI ethics release) (WIRED) These principles impose constraints: generative outputs must be traceable and auditable, biases minimized, human oversight maintained, and reliability assured—all of which suggest that the planning cell must treat AI outputs as advisory artifacts, not authoritative prescriptions.

Beyond doctrine and governance, academic and policy research reinforce the idea that organizational adaptivity must precede performance gains. RAND’s portfolio of AI and military work emphasizes that institutions must adjust structure, metrics, and workforce culture before expecting substantive AI-derived outcomes. (Verified via RAND works) (C4ISRNet) In the planning domain specifically, RAND’s “Improving Sense-Making with AI” project examines how AI can aid framing in contested, multi-domain operations. (Verified via RAND report) (U.S. Department of War) Such research underscores that the planning cell must evolve its epistemic norms, feedback loops, and error tolerance to fully extract performance.

Insertion Points in Mission Analysis and Design: Mapping Divergent Thinking to Model Capabilities

During mission analysis, planners seek to define problems broadly, generate multiple hypotheses, challenge assumptions, and map the operational environment—functions that align with divergent thinking more than convergent judgment. Joint doctrine, per JP 5-0: Joint Planning (December 2020) chapter 4, frames operational design as the upstream activity that shapes problem framing, lines of operation, and conceptual approaches before detailed planning begins. (Programma Risorse di Intelligence) In particular, JP 5-0 acknowledges that operational design and planning are complementary: design begins before procedural planning and iterates during planning to guide intent and logic flow. (NDU Press) A generative model, properly scaffolded, can assist a planning team in phases where hypothesis generation, environmental framing, and course-of-thought articulation dominate over rigorous trade-space filtering.

The cognitive alignment between mission analysis tasks and generative model strengths is supported by domain research in sense-making. The RAND study Improving Sense-Making with Artificial Intelligence (2025) argues that analysts confronting complex, ambiguous contexts require lateral framing, hypothesis expansion, and iteration of alternative narratives—areas where AI systems can propose alternative conceptual framings, surface hidden assumptions, and test “what-if” parametric shifts. (RAND Corporation) In that study, the researchers map sense-making challenges across collection orchestration, data fusion, model management, and skill/training. They conclude that artificial systems are well suited for hypothesis-generation, scenario recombination, and analogical inference — tasks that mirror divergent thinking in mission analysis. (RAND Corporation) Accordingly, a well-positioned generative model should contribute at least three functional capabilities in mission analysis: hypothesis extension, assumption stress testing, and contextual analogical support.

Hypothesis Extension and Concept Generation. When planning teams confront ill-structured problems, models can propose candidate operational approaches, within domain constraints, that humans might not spontaneously consider. The model might surface conceptually coherent variants—e.g. alternative centers of gravity, lines of effort, intermediate objectives, or campaign phasing templates—by mining doctrinal corpora, historical case analogs, or simulation-derived logic. Because generative models excel in pattern completion and associative mapping beyond immediate local context, they can push the boundary of creative option space beyond heuristics or staff mental models.

To operationalize this, prompt templates must frame mission variables (ends, means, constraints, risk) and ask for ranked variant sketches. Prompts should include explicit instructions like “generate three alternate conceptual approaches given changes in risk axis X or adversary response Y.” The AI-generated sketches then become counterpoints for staff to analyze, critique, and adapt.

Assumption Stress Testing. A frequent failure in planning is unexamined assumptions. Generative systems can be tasked to “challenge” core assumptions by querying boundary conditions, adversary response spaces, environmental variations, or escalatory pressures. For example, a prompt might ask: “List five plausible challenges to my assumption that adversary logistical mobility is degraded by seasonality Z.” The model’s role is to produce structured counter-hypotheses and edge-case narratives, which the staff can then vet via intelligence, war-gaming, or red-team review.

Analogical and Historical Contextual Support. Generative models can draw analogies across historical campaigns, allied doctrine, and regional case studies when prompted with context features (terrain, actor motivations, political constraints). For instance, given a proposed objective set and adversary disposition, the model can propose analogous historical or doctrinal precedents (e.g. operations in archipelagic terrain, contested littorals, urban-fractured environments). While not authoritative, these analogies stimulate contrastive thinking and enrich the team’s mental model. The key is to employ moderate constraints (e.g. “most similar within last 20 years involving peer competitors”) and demand explicit caveats (“differences compared to historical case”).

Mapping these model capabilities to distinct mission analysis sub-tasks yields concrete insertion points:

• Framing Key Operational Questions. At the outset of mission analysis, the planning team defines essential planning questions (e.g. “How will adversary A mass against axis B?”, “What second- and third-order effects arise under scenario C?”). A generative model can propose a refined set of 8–12 critical questions across political, informational, military, economic, and security dimensions. The planning cell can then prune, reorder, and adopt from that list.
• Environmental Factor Expansion. As staff build a lines-of-effort or center-of-gravity matrix, the model can suggest additional environmental factors, relationships, and indicators drawn from doctrinal encyclopedias, inter-agency datasets, or case studies. For example, in an Indo-Pacific littoral context, the model might surface maritime domain awareness challenges, infrastructure chokepoints, or information posture risks that the team missed.
• Hypothesis Bundle Generation. After framing the problem, the team often drafts initial solution hypotheses—e.g., “attrition-first,” “maneuver-first,” “economy-of-force with shaping.” The model can be asked to produce variant hypothesis bundles (e.g. blending acceleration, shaping, envelopment) including associated logic flows and major assumptions. The staff then compares, discards, or merges.
• Assumption Audit and Alternative Narratives. Once key hypotheses are in hand, schedule a prompt interchange: the model is tasked to generate counter-narratives or assumption-violation cases. These narratives become the raw material for risk matrices or decision criteria.
• Branch, Sequencing, and Phasing Sketches. The model can help generate candidate phasing or branch structures (e.g. Phase 0 shaping, Phase 1 entry, Phase 2 consolidation, Phase 3 transition) and sketch logic transitions. Teams can then overlay these on their own timelines and resource constraints.
• Conflict Logic Looping. For adversarial interaction loops—enemy choices, friendly reactions, second-order dynamics—a model can propose plausible branching triggers, response heuristics, and friction loops. This helps the team avoid linear model error by anticipating cross-domain coupling, coercion gradients, or escalation triggers.

Note: the value of model outputs at these stages is only as good as prompt quality, domain constraints, and the human reviewer’s skepticism. The planning team must treat the AI artifacts as hypothesis prompts, not final claims.

Empirical field cases of earlier AI-enabled planning tools—for instance, the JADE (Joint Assistant for Development and Execution) system—illustrate the potential and limits of automated support during early planning. JADE combined case-based planning knowledge and user dialogue to accelerate TPFDD (time-phased force deployment data) generation in crisis action planning. (Wikipedia) While more rigid than modern generative systems, JADE’s structured logic offers a baseline for scaffolding current models.

In addition, the literature on adversarial reasoning (e.g. Kott & Ownby’s Toward a Research Agenda in Adversarial Reasoning) emphasizes that planning tools must anticipate that models and adversaries co-evolve. (arXiv) Hence, insertions of generative models in mission analysis must include feedback loops where outputs are re-submitted across model-human cycles.

The insertion schedule matters. To maximize influence, planners should invoke generative assistance at three strategic junctures:

• Initial framing juncture, shortly after mission receipt and guidance assimilation, for question set expansion and problem mapping.
• Hypothesis bundle juncture, after internal staff brainstorming but before hypothesis lock-in, to stretch thinking and prevent group fixation.
• Narrative stress test juncture, right before finalizing assumptions and branch logic, to provoke counter-models and risk stress.

Between these junctures, teams must preserve temporal buffers for human review, red teaming, intelligence liaison, and cross-domain vetting. Without that buffer, generative outputs risk being accepted uncritically.

To prevent cognitive dominance by the model, a role of model critic should be designated within the staff—someone whose explicit task is to challenge generative outputs using domain knowledge, historical precedent, or red-teaming heuristics. This internal role ensures that outputs are interrogated and contextualized.

As models evolve, adaptive calibration is essential. Over multiple planning cycles, prompt templates should evolve based on feedback: prompt effectiveness logs, false-positive counter-narratives, user editing rates, and red-team pushback. A continuously refined prompt library aligned with joint doctrine, allied warfare concepts, and regional templates becomes a shared institutional asset.

Units electing generative integration should plan an onboarding ramp: begin with coarse prompts (e.g. conceptual sketches) and gradually test into more complex prompts (phasing logic, branching schemes) as confidence and oversight increase.

Monitoring and measurement loom large. The planning team should instrument prompt logs, usage rates, user edit distances, red-team revisions, and derived decision-quality differentials (to the extent feasible). Over time, comparative case studies can test whether model-assisted mission analyses yield more robust, flexible plans or higher staff satisfaction.

In contested or classified domains, generative insertion must observe security constraints: prompt templates should avoid leaking sensitive data, outputs should be filtered or sanitized, and classification stamping rules must be inserted before artifacts leave secure enclaves.

As model interfaces evolve, interactive multimodal inputs (maps, geospatial overlays, sensor feeds) may further tighten alignment: a prompt could reference a digital terrain model and ask the system to suggest lines of advance or observation posture variants. While not yet mainstream, such integration should be conceptualized early.

In summary, mission analysis is the ideal locus for generative models to stretch conceptual frames, challenge assumptions, and seed variant hypotheses. To make that insertion effective, planning teams must schedule junctures, role-designate critics, instrument feedback loops, and progressively mature prompt scaffolds rooted in doctrine. If deployed thoughtfully, generative AI becomes an intellectual accelerator in the earliest, most uncertain phase of planning—raising the ceiling of what the staff can imagine.

Expertise Boundaries, Bias Mitigation and Human–Machine Teaming Evidence

Establishing expertise boundaries requires recognizing where generative systems provide net gain versus redundancy. In domains where staff already hold deep tacit knowledge—targeting, doctrinal nuance, cryptographic tradecraft—AI outputs may duplicate, conflict with, or degrade human judgments. Scholarly reviews of AI decision support in military domains note that augmentation is most effective when filling knowledge gaps or weak links in staff cognition rather than displacing domain expertise. For example, the Review of AI in Military Decision Support Systems (2024) outlines that human–machine interaction tensions emerge most strongly when humans feel the system encroaches on core domain authority. (ResearchGate)

Research on cognitive bias mitigation in military intelligence workflows provides explicit support for AI systems framed as correctors rather than propagators. The RAND report Exploring Artificial Intelligence Use to Mitigate Potential Human Bias Within U.S. Army Intelligence Preparation of the Battlefield (IPB) (2024) argues that generative AI could serve as a guardrail against analysts’ confirmation bias, availability bias, or anchoring by proposing alternative threat estimates, divergent options, or overlooked data patterns. (RAND Corporation) The report emphasizes that AI must be integrated with human processes—not replace them—and should include an audit trail so staff critique algorithmic outputs rather than defer blindly. (RAND Corporation)

The Center for Security and Emerging Technology’s brief Reducing the Risks of Artificial Intelligence for Military Decision Advantage warns that AI systems with weak robustness, adversarial vulnerability, or misapplication to inappropriate tasks can inject false confidence or incorrect inferences into decision processes. (cset.georgetown.edu) In the context of operational planning, that means staff must treat AI outputs not as final but as hypothesis prompts subject to vetting and red-teaming.

Human–machine teaming scholarship underscores that the combination of human judgment and algorithmic consistency can outperform either alone, provided trust is calibrated and transparency is maintained. An arXiv working paper Advancing Human-Machine Teaming: Concepts (2025) emphasizes that trust and reliability, explainability, and adversarial robustness are pillars of effective teaming. (arXiv) The same paper highlights that team design must explicitly manage how human biases and AI biases interact, rather than assume independence. (arXiv)

From the National Academies volume Human-AI Teaming: State-of-the-Art and Research Needs (2022), Chapter 10 (Identification and Mitigation of Bias in Human-AI Teams) warns that both human and AI systems carry biases: human heuristics can influence which outputs are considered; AI systems themselves may reflect dataset skew, label error, or model artifacts. It emphasizes the need for bidirectional awareness—humans must notice AI bias; AI systems (or filters) must detect human blindspots. (nap.nationalacademies.org)

Explainable AI (xAI) plays a central role in bridging the epistemic gap. The work The Utility of Explainable AI in Ad Hoc Human–Machine Teaming (Paleja et al., 2022) shows through experiment that xAI improves situational awareness for novice users but may overload experts, degrading performance when explanation overhead competes with domain judgment. (arXiv) In mixed-expertise planning teams, this suggests the need for selective explanation modes: coarse summaries for domain experts, richer transparency for novices or analysts.

The principle of Meaningful Human Control (MHC) in defense systems has been advanced in the paper Designing for Meaningful Human Control in Military Human-Machine Teams (van Diggelen et al., 2023). They propose that human control is not an afterthought but a design objective spanning system architecture, interface, workflow, and accountability. In planning contexts, MHC means staff always retain authority to reject or reconfigure AI output, and that the system’s design reinforces—not undermines—the human’s ultimate decision prerogative. (arXiv)

Hybrid human–machine systems in space and cross-domain domains reveal comparative-advantage insights. The U.S. Space Force discussion document Hybrid Human-AI Teams Represents Defense Technology Future (May 2024) asserts that humans excel in contextual judgment, value tradeoffs, and deception recognition, while machines excel at high-speed data fusion, pattern detection, and consistency. The document urges role partitioning so that each agent focuses on its strength. (spacecom.mil)

Applying this to planning teams: generative AI should be tasked with volumetric reasoning, analogical generation, and cross-domain correlation, whereas humans should handle norm judgments, breakpoints in logic, and mission risk tradeoffs. The boundary—when AI offerings are trusted or disregarded—must be clear and pre-declared, not ad hoc.

Evaluations of automation bias reinforce risk. The CSET analysis AI Safety and Automation Bias identifies cases in aviation and defense where operators errantly accepted automated output contrary to independent evidence. The paper proposes a tri-level mitigation framework: user training, interface design (e.g. “confidence scores,” counterfactual explanations), and organizational oversight. (cset.georgetown.edu) Within a planning cell, this implies that AI outputs be accompanied by confidence metrics, prompts for counter-analysis, and mandated critique before adoption.

The RAND working paper An AI Revolution in Military Affairs? (2025) explores how AI adoption in warfare will require institutional adaptation. It cautions that overreliance on AI in decision loops without human oversight can erode adaptive judgment under novelty or surprise. (RAND Corporation) That insight further stresses that that human–machine teaming architecture must deliberately preserve human flexibility, revision, and error diagnosis capacity.

A case study in machine learning for operational decisionmaking published by RAND (Machine Learning for Operational Decisionmaking in Competition, RRA-815-1) explores how models assisting competition-phase decisions must manage the risk that adversaries may exploit model predictability or overfitting to patterns. (RAND Corporation) That risk emphasizes the need for human oversight of model outputs, especially where an adversary might reverse engineer or anticipate AI-assisted logic patterns.

Combining these strands, we derive five best-practices for reconciling expertise boundaries, bias mitigation, and team function in planning:

• Explicit role demarcation: assign Model Analyst roles distinct from Domain Specialist roles, each with clear permission boundaries. Model Analysts vet AI output proposals; Domain Specialists enforce doctrinal, mission, and risk criteria.
• Prompt scaffolding with bias heuristics: prompt templates should include directives to “consider alternative frames,” “flag assumptions,” or “highlight confidence bounds.” This shapes model behavior to avoid echoing staff bias.
• Mandatory uncertainty metrics and explanation layers: require confidence intervals, token-level salience, or counterfactual alternatives as part of AI outputs to force staff reflection rather than blind acceptance.
• Dual-stream review with reductive and constructive critique: present paired streams—AI’s best output and the model’s alternative—that must be cross-examined by staff with a bias checklist (e.g. anchoring, omission, confirmation, availability).
• Adaptive calibration and feedback logging: prompt-edit logs, staff acceptance rates, error corrections, and red-team overwrites must be tracked so the cell can recalibrate prompt templates, filter heuristics, and decision thresholds over time.

Empirical data on adoption rates or error incidence in live military planning environments is not publicly disclosed. No verified public source available for AI-augmented planning error statistics or planning cell acceptance rates.

In summary, expertise boundaries demand that AI support, not supplant, human domain authority. Bias mitigation must be baked into prompt design, interface design, role partitioning, and review procedures. Human–machine teaming success depends on calibrated trust, explainability, segmented authority, and continuous evaluation—not mere deployment.

Workforce, Training Pipelines and Assurance: From Enthusiasm to Competence

The transition from curiosity-driven exploration to sustained competence in generative AI demands a deliberate architecture of workforce development, role stratification, training pipelines, and assurance mechanisms. In defense organizations, this process must align with acquisition, personnel, and governance systems—and must satisfy doctrinal, ethical, and risk constraints.

A necessary first step is formalizing AI work role taxonomies within the Department of Defense. The DoD Responsible Artificial Intelligence Implementation Pathway (June 2024) mandates that “all DoD AI workforce members possess an appropriate understanding of the technology, its development process, and the operational methods applicable to implementing RAI commensurate with their duties within the archetype roles outlined in the 2020 DoD AI Education Strategy.” (media.defense.gov) That means planning cells must map each staff position (e.g., operations, intelligence, logistics) to an AI-enabled role archetype and tailor training accordingly.

Historically, the DoD has struggled to identify and codify its AI workforce. The Government Accountability Office (GAO) in GAO-24-105645 observed that although DoD developed AI work roles, it had not assigned clear responsibility or timelines to fully define and integrate those roles across personnel systems. (gao.gov) This drift impedes scalable training pipeline design. Service-level reform, therefore, must start with role coding, billet classification, and integration into talent management systems.

Academic and defense literature propose multi-tiered workforce stratification. The Joint Force Quarterly article “An AI-Ready Military Workforce” (Cruickshank et al., 2023) outlines a five-layer model: Users, Leaders & Acquisitions, Technicians, Functionaries, and Experts, with distinct instructional durations and depth of responsibility. For example, User-level training spans weeks to months to cultivate AI literacy, whereas Expert-level roles require years of deep technical education and experience. (ndupress.ndu.edu) To operationalize planning-team adoption, organizational planners must map staff into these tiers and prioritize high-impact, mid-tier roles (Technician/Functionary) for early capacity build-out.

In parallel, the Carnegie Mellon–Army AI Technicians program offers a validated model of rapid upskilling. The program engaged a combined academic–service cohort to develop pipeline techniques for training military AI technicians in months rather than years. Invocation of incremental learning, scaffolding, and continuous curricular refreshment enabled production of 59 AI Technicians in early cohorts. (Verified via AI Technicians: Developing Rapid Occupational Training Methods, January 2025) (arxiv.org) That model demonstrates that defense training systems can compress preparation cycles without overly diluting quality—but only when strongly integrated with mission contexts and iteration loops.

The CDAO Digital Workforce initiative explicitly addresses these demands. The DoD’s portal describes goals such as coding 50 percent of the Total Force for data/AI roles in calendar year 2025, expanding access to learning pathways, and aligning skills to mission needs. (ai.mil) Its “Workforce Training” capability describes a vision for self-assessment, shared curriculum via Digital University, and executive upskilling series. (ai.mil) For planning staffs, that translates into leveraging enterprise training while customizing team-level courses for generative usage in operational planning.

To move from training offerings to actual competence, a phased onboarding scaffold is essential:

• Foundational literacy: Short courses (weeks to months) on AI principles, prompt design, error modes, assurance, ethics, and security, tailored to planners and staffers. These courses mirror the “User” and “Leader/Acquisition” tiers in the JFQ model.
• Hands-on modules: Practical labs and scenario-based prompt exercises embedded in planning education environments (e.g., staff colleges, war games) to ensure participants experiment with generative models under controlled conditions.
• Guided coaching and peer review: Embedding mentors or “AI champions” within units who provide feedback, maintain prompt libraries, and act as force multipliers.
• Progressive autonomy: After coached phases, staff may transition to operational deployment with scaffolding relaxed gradually, while oversight remains.
• Periodic recertification and feedback loops: Continuous evaluation of prompt logs, error rates, red-team corrections, model drift sensitivity, and participant feedback to adjust curriculum.

In educational settings, for example, the U.S. Army’s Professional Military Education institutions are experimenting with AI-integrated curriculum. The article “Enhancing Professional Military Education with AI” (Army University Press, April 2025) describes guidance from TRADOC and concerns over terminology, ethical pitfalls, and best practices to embed generative capability into PME without violating doctrine or security constraints. (armyupress.army.mil) For operational planning training, similar embedding ensures that AI tools become normalized artifacts in staff education, not exotic experiments.

Assurance and trust must evolve in parallel. A generative tool mishandled or uncontrolled can produce compromised, biased, or insecure outputs. To avoid that, planning pipelines must include claims-based assurance consistent with emerging frameworks like A Framework for the Assurance of AI-Enabled Systems (Kapusta et al., 2025). That paper presents a structured approach by which system designers define claims (e.g. “the model output is resilient to input perturbations X”), map those claims to evidence, and continuously validate under stress conditions. (Verified via arXiv) (arxiv.org) Planning organizations must adopt such assurance rigs for their generative modules—incorporating test harnesses, red teaming, adversarial input banks, prompt fuzzing, and held-out scenario validation.

Within DoD policy, the DoD Responsible AI Strategy and Implementation Pathway links workforce capabilities and assurance: training must include traceability, auditability, operational method familiarity, and governance awareness. (media.defense.gov) That merges technical assurance practice with personnel training requirements. Staff must understand not only how to use the tool, but how to interpret logs, detect anomalous behavior, spot potential model degradation, and maintain governance compliance.

Career incentive alignment is another critical piece. The CSET report The DoD’s Hidden Artificial Intelligence Workforce recommends linking performance, promotion, retention, and recognition to AI competence: e.g. special pay, evaluation credit for AI innovation, rotational assignment in AI roles, and performance metrics tied to AI contributions. (cset.georgetown.edu) Without incentives, staff will revert to legacy behaviors.

Moreover, inter-service harmonization of AI workforce policies is essential to avoid stovepipes. The CSET brief urges repurposing outdated two-digit function codes to designate AI archetypes across services and synchronizing pilot experiences across components. (cset.georgetown.edu) For planning cells in joint staffs, this harmonization ensures that AI-trained personnel remain interoperable and transferrable across command echelons.

Assurance frameworks must also integrate with acquisition cycles. Generative AI modules leveraged by planning teams should adhere to DoD’s system engineering and acquisition oversight. Tools might enter as software support elements under contracts with embedded assurance obligations, logging, and capability refresh paths. The planning cell should not treat a generative model as an ad hoc plug-in; it must be integrated as a software module subject to lifecycle management, patching, security, validation, and configuration control.

As a pilot example, the U.S. Army in June 2024 launched a pilot project to explore generative AI within its acquisition and contracting workforce, producing lessons parallel to staff environments. The pilot intends to test efficiency gains and boundaries for safe adoption. (Verified via C4ISRNet) (c4isrnet.com) Insights from that experience should feed planning cell pipeline design—especially in prompt scaffolding, oversight metrics, and staff acceptance curves.

Defense-wide workforce strategy alignment is also present. The DoD Cyber Workforce Strategy (March 2023) highlights objectives for assessments, alignment of talent to development programs, and training pipelines to match roles. (dodcio.defense.gov) Although focused on cyber, its principles of capability-based training and role alignment are directly translatable to AI workforce development.

To guard against fossilization of training, pipelines must embed continuous adaptation. As models evolve, training modules must be refreshed; prompt tactics, adversarial vulnerabilities, and architectural shifts must update instruction. The AI Technicians program already evolves its curriculum across cohorts. (Verified via arXiv) (arxiv.org) Planning organizations must treat their generative AI training library as a living artifact.

Measurement and evaluation is central. Metrics should include prompt usage frequency, staff revision rates (edit distance), red-team correction counts, time-to-satisfactory output, user confidence surveys, planning cycle times, and mission-analysis depth comparisons across AI and non-AI teams. Only empirical measured outcomes can validate competence maturation and guide adjustment.

In contested or classified contexts, training must include red-teaming of model adversarial exploitation, scenario-based security filtering, prompt sanitization techniques, and classification stamping protocols. Staff should train with “sterile” prompt environments and simulated leak risk before full access.

Finally, ecosystem partnerships with academia and industry bolster capacity. Defense should partner with universities, research centers, coding bootcamps, and AI firms to leverage cutting-edge content, shared libraries, and talent flow. The National Security Innovation Network (NSIN) is one such nexus, fostering human capital innovation and cross-sector training. (Verified via wiki) (wikipedia.org) Aligning planning-cell training to external innovation hubs strengthens pipeline resilience.

In sum, to move planning cells from enthusiasm to competence requires codified AI role architecture, compressed and context-rich training pipelines, assurance frameworks integrated into training, incentive-aligned career paths, measurement systems, continuous curriculum adaptation, cross-domain alignment, and operational validation of staff competence. Only then will generative AI shift from peripheral novelty to embedded planning capability.

Small-Group Dynamics and Mandatory Checkpoints: Forcing Function Integration in Staff Work

In operational planning teams, social hierarchy, conversational dominance, workload asymmetry, and time pressure routinely suppress the disciplined use of generative systems unless procedures compel their inclusion at defined junctures. The doctrinal backbone for designing such procedures remains JP 5-0: Joint Planning, which describes mission analysis, course-of-action development, and comparison as sequential yet iterative activities that must withstand structured challenge before decision approval; the Joint Staff’s current portal for JP 5-0 codifies these expectations and is the authoritative reference for staff workflow design Joint Chiefs of Staff, JP 5-0: Joint Planning. Complementing this joint baseline, the Marine Corps’ MCDP 5: Planning emphasizes commander-driven design, critical inquiry, and the deliberate surfacing of assumptions as foundational habits in staff practice; these emphases create natural anchor points for compulsory human–machine interaction during early framing and later red-teaming MCDP 5: Planning (October 4, 2018).

A forcing-function architecture must also meet assurance and traceability requirements drawn from the National Institute of Standards and Technology. The AI Risk Management Framework 1.0 (January 26, 2023) establishes the Govern–Map–Measure–Manage cycle as a universal structure for risk-aware design, and its generative companion profile NIST AI 600-1 (July 26, 2024) translates those functions to large language models by prescribing provenance logging, distribution shift monitoring, robustness checks, and transparency artifacts NIST AI RMF 1.0 and NIST AI 600-1. For defense users, the Chief Digital and Artificial Intelligence Office has operationalized policy through its Generative AI Toolkit (December 11, 2024), which includes suitability–feasibility–advisability checklists, risk questionnaires, and employment guidance that can be embedded as mandatory gates in staff workflows CDAO Blog: GenAI Toolkit operationalizes Guidelines & Guardrails memo (December 11, 2024) and the corresponding PDF, Generative AI Version 1.0 RAI Toolkit (December 2024).

Within this verified governance environment, group behavior in staff rooms is the principal obstacle to reliable tool use, not interface access. The RAND Corporation’s One Team, One Fight: Insights on Human–Machine Teaming (June 2, 2025) documents friction points when pairing soldiers with algorithms in warfighting tasks: ambiguity about authority boundaries, over-trust in automation, under-trust due to opaqueness, and the absence of shared mental models. The report recommends procedural counterweights—explicit role partitioning, graded explanation, and rehearsed team behaviors—to prevent either human or algorithmic dominance RAND RRA2764-1 (June 2, 2025) and full report PDF. In parallel, RAND’s Exploring Artificial Intelligence Use to Mitigate Potential Human Bias Within U.S. Army IPB (August 6, 2024) shows where structured insertion can correct anchoring, availability, and confirmation effects by obligating the staff to consult machine-generated alternative hypotheses during intelligence preparation steps; the value arises from forced contrast, not from replacing analysts RAND RRA2763-1 (August 6, 2024) and PDF. Together, these findings justify a staff design in which human–machine engagement is required at precise points, audited after each evolution, and never left to discretionary enthusiasm.

Designing those points begins with doctrine-aligned phasing. During mission analysis—conceptual design in MCDP 5 terms—teams should trigger an initial “AI framing” gate after commander’s guidance and before problem statements are finalized. The gate forces the AI Analyst role to generate multiple problem framings and question sets keyed to the mission variables in JP 5-0, while the Output Critic annotates divergences from the team’s baseline. The artifacts are immediately logged under NIST AI RMF “Map” and “Measure” functions: prompts, outputs, rationale notes, and confidence flags are versioned and stored for later audit NIST AI RMF 1.0. This gate prevents premature convergence, the classic small-group pathology where early narratives harden into doctrine-sounding certainties.

A second gate sits between divergent option generation and convergent course-of-action comparison. Here the Generative AI Toolkit provides pre-formatted suitability–feasibility–advisability worksheets; staff are required to run model-aided COA variants through that template and to document any automated suggestion accepted, modified, or rejected. The worksheet’s provenance and the prompt version become part of the plan record CDAO RAI Toolkit (December 2024). This gate shifts the AI from a “curiosity” into a governed participant with auditable contributions, aligning with the “Govern” and “Manage” elements of NIST.

A third gate precedes risk registers and branch/sequel logic. Before staff finalize risk narratives and decision points, the AI Analyst must solicit counter-narratives and edge-case stressors, while the Output Critic cross-references those with doctrinal risk categories and theater-specific constraints documented in JP 5-0. Any nontrivial divergence triggers mandatory red-teaming and an explicit written adjudication. The record of that adjudication becomes a training case for future iterations, aiding transactive memory—shared awareness of who knows what in the team—so that later cells understand where the model tended to miss or overreach in similar contexts JP 5-0 and MCDP 5.

Because small groups often default to interpersonal resolution rather than tool discipline, the forcing-function logic must be backed by structural roles and rotating facilitation. A redesigned roster adds four positions to the traditional operations, intelligence, logistics, fires, and communications billets: AI Analyst (constructs prompts; runs iterations), Prompt Librarian (curates, versions, and doctrine-maps templates), Output Critic (formally challenges and contextualizes outputs), and Telemetry Custodian (captures usage metrics, edit distances, acceptance/rejection tallies, red-team overrides, and time-to-usable-draft). Each role’s authorities and limits are defined in the planning order so that no individual can insert unreviewed outputs into staff products. These assignments align with RAND’s human–machine teaming recommendation to partition responsibilities and avoid ambiguous authority boundaries RAND RRA2764-1 (June 2, 2025).

The reliability of this design depends on auditable evidence. Under NIST’s “traceability” and “transparency” properties, every prompt–output–revision chain is preserved, with model version tags and justification notes. The storage can be implemented with standard configuration-controlled repositories plus cryptographic hashing to make tampering visible. The Telemetry Custodian produces dashboards after each evolution: prompt counts per phase, median edit distance from model output to accepted text, percentage of suggestions adopted or discarded, and latency between prompt and staff-ready draft. These are managerial instruments, not vanity metrics, used to revise templates, adjust role rotations, and flag processes at risk of re-marginalizing the AI NIST AI RMF 1.0 and NIST AI 600-1.

Mandatory checkpoints also mitigate the dual hazards of automation bias and opacity documented in military contexts. RAND’s IPB study shows that requiring machine-generated alternatives reduces analysts’ tendency to favor first hypotheses, while preserving human judgment as the adjudicating authority RAND RRA2763-1 (August 6, 2024). RAND’s human–machine teaming volume highlights that unchecked opacity corrodes trust or, paradoxically, invites over-reliance—both failure modes that procedural gates dampen by demanding explainability and explicit human critique RAND RRA2764-1 (June 2, 2025). The CDAO toolkit’s assessment questionnaires, when embedded as “no-go unless complete” forms, formalize that critique rather than leaving it to informal discussion CDAO Blog (December 11, 2024).

To prevent dominance by rank or personality, facilitation rotates at each gate. The rotating facilitator enforces time-boxed, dual-track presentation: the AI Analyst briefs outputs first; the Output Critic then presents a human-constructed counter-summary with clear doctrinal anchors; the team debates with the rule that any acceptance must include a short written rationale and risk note. The written component slows premature consensus by compelling individuals to articulate reasons, improving later audit. This procedure reflects MCDP 5’s insistence that planning is a dialogue of intent, critique, and adaptation rather than static checklists MCDP 5: Planning.

A credible forcing-function design must accommodate security and data-handling constraints. The CDAO materials require controlled data provenance and explicit employment guidance; prompts cannot leak classified content outside accredited enclaves, and model versions must be certified for the data tier in question RAI Toolkit (December 2024). The plan record must therefore include inputs scrubbed for classification, the sanitized prompt text, and a cross-reference to the original classified analysis held elsewhere. This bifurcation preserves audit trails without violating handling rules and is consistent with the NIST profile’s data-governance guidance NIST AI 600-1.

Because small-group behavior adapts to incentives, senior leaders should tie compliance with checkpoints to evaluation and after-action review. The planning directive specifies that no product progresses to wargaming or orders drafting without evidence of completed gates. Compliance is inspected during post-operation reviews and professional military education seminars; units demonstrating disciplined use receive credit in performance narratives, reinforcing habits over ad hoc reliance. This aligns with RAND’s organizational recommendation to couple process change with measurement and leadership signaling to stabilize new human–machine behaviors RAND RRA2764-1 (June 2, 2025).

Red-team procedures should be upgraded to include model-challenge cycles. A red-team cell is tasked to regenerate prompts independently using the same sanitized context, to test sensitivity to phrasing and to expose brittle outputs. Divergent results are recorded as robustness flags and trigger a focused follow-up: either template tightening or additional human analytic work. This practice satisfies NIST’s “Measure” and “Manage” steps by actively probing for variance and stress susceptibility NIST AI RMF 1.0.

To keep small-group dynamics from drifting back to habit, the team’s transactive memory about the model must be cultivated. The Telemetry Custodian publishes a one-page “AI behavior brief” after each cycle: domains where the model added value, recurring hallucination patterns, prompt formats that underperformed, and examples of high-leverage edits by the Output Critic. Over time, this brief becomes a living local doctrine for model employment, nested under MCDP 5’s philosophy of adaptive planning and JP 5-0’s requirement for continuous learning within the planning series MCDP 5 and JP 5-0.

The redesigned process must also integrate with enterprise governance. The CDAO has framed a DoD-wide path to AI readiness that includes workforce upskilling, shared tooling, and portfolio oversight; planning units should register their prompt libraries and telemetry schemas with the enterprise knowledge base so that practices can be compared across commands and improved centrally CDAO Blog index (August 13, 2025). Meanwhile, compliance artifacts—gate completion forms, prompt logs, and rationale notes—should be formatted so they can feed test-and-evaluation repositories and responsible-AI oversight reviews, maintaining continuity from staff procedure to enterprise assurance RAI Toolkit (December 2024).

Finally, the team must be protected against over-centralization of AI competence. Rotating the AI Analyst and Output Critic assignments ensures that expertise propagates and that no single personality becomes the de facto gatekeeper—an important check against micro-hierarchies that often emerge in small groups. This rotation, combined with explicit authority boundaries and written adjudications, institutionalizes the procedural fairness that RAND highlights as central to sustaining trust in human–machine pairings RAND RRA2764-1 (June 2, 2025).

The cumulative effect of these measures is to convert the staff room’s social energy—from cohesion-seeking that quietly discards machine contributions—into disciplined collaboration in which algorithms are obligated to propose, humans are obligated to interrogate, and each exchange leaves an auditable trace. Doctrinal anchors in JP 5-0 and MCDP 5, risk scaffolding from NIST’s AI RMF and AI 600-1, and defense-specific guidance in the CDAO Generative AI Toolkit together validate the architectural principle: only mandatory checkpoints, staffed roles with review authority, telemetry that measures adoption quality, and audited records can reliably overcome small-group inertia and secure repeatable decision advantage in modern planning environments JP 5-0, MCDP 5, NIST AI RMF 1.0, NIST AI 600-1, and CDAO GenAI Toolkit.

A Redesign Blueprint for Operational Planning Teams: Roles, Templates, Telemetry, and Auditable Controls

When planning teams are reconstructed around generative AI as a core cognitive tool rather than a sidelined novelty, they require a holistic redesign of roles, prompt templates, telemetry systems, and auditable control architectures. This blueprint prescribes four interlocking dimensions—role architecture, modular prompt scaffold templates, telemetry instrumentation, and governance/audit control loops—to embed generative AI as a dependable, monitored planning partner.

Role Architecture and Authority Boundaries
A planning cell organized around AI must reframe its internal role architecture beyond classic functions (operations, intelligence, logistics). Key new roles include: AI Analyst, Prompt Librarian, Output Critic, and Telemetry Custodian. The AI Analyst designs and refines prompts, runs iterations, and surfaces lead AI-derived options; the Prompt Librarian curates stable prompt templates, maps prompt evolution, and ensures doctrinal alignment; the Output Critic holds responsibility to challenge, contextualize, and annotate AI outputs using domain judgment; and the Telemetry Custodian maintains prompt logs, usage metrics, red-team interventions, and statistical summaries for after-action review.

Authority boundaries must be explicit. For example, the AI Analyst may not directly embed outputs into staff products without Critic review; the Critic must annotate every accepted or rejected output with rationale. Over time, as trust grows, the boundary may relax, but early designs must enforce rigid checks. Rotating these roles ensures no individual monopolizes AI influence or becomes a bottleneck.

Each function must have a defined presence across planning phases: mission analysis, COA development, war gaming, risk estimation, and orders drafting. The AI Analyst should be active early (divergent thinking phases), taper during convergent phases, then reinserted during anomaly detection. The Critic must remain present from initial output vetting through final plan consolidation.

Prompt Template Modules and Versioning Strategy
Prompt design cannot be ad hoc; it must be modular, versioned, and doctrine-aware. The core prompt template library should modularize by planning phase, domain function, and risk posture. For example:

• Phase-0 Framing template: “Given mission statement M, adversary posture A, constraints C, generate three conceptual approaches, each with key assumptions, risks, and alternative lines of operation.”
• Assumption Audit template: “List ten possible challenges to assumption X, grouped by category (logistics, information, terrain, adversary adaptation).”
• Branch Sketch template: “Propose branching logic triggers, with associated phasing criteria and fallback options if branch fails.”

Templates must carry metadata: doctrinal reference (e.g. JP 5-0 clause), version identifier, authorship, date, and known failure modes. When a staff iteration rejects or heavily edits an AI output, that iteration should feed back into versioned refinement of prompt templates.

Prompt versioning must align with telemetry systems, so usage logs can trace which template version generated given output. Over time, cells build a shared prompt version history adapted to their mission set, staffed domain, regional theaters, and risk tolerance.

Telemetry Instrumentation and Usage Metrics
To monitor AI integration fidelity, planning teams must embed telemetry systems at granular levels. Telemetry layers should include:

• Prompt invocation counts: How many prompts fired, per phase, per functional cell.
• Edit distance metrics: Measuring how far human edits diverged from AI output.
• Acceptance rates: How often staff adopt AI suggestions verbatim, with modification, or reject entirely.
• Red-team override counts: Instances where outputs passed internal vetting but failed external or adversarial review.
• Time to usable output: Interval between prompt issuance and staff-ready draft.
• Prompt usage patterns: Which prompt template versions succeeded or failed, clustering by context or mission type.
• Anomaly flags: Tracking outputs that triggered surprise rejections, outlier divergence or hallucination markers.

A Telemetry Custodian should aggregate and dashboard these metrics after each planning cycle, enabling longitudinal analysis of AI integration maturity. Comparative benchmarking across cells or scenarios can surface best practices or prompt pitfalls. If multiple planning teams adopt a common telemetry schema, cross-team learning accelerates.

Auditable Control Loops and Governance Architecture
Generative AI in planning must be governed by auditability, traceability, and compliance. To achieve this, the planning cell must embed auditable workflows:

• Prompt → Output → Revision Chain
  Every prompt, its output, human edits, and final accepted text must be chained in a tamper-evident log (e.g. via cryptographic hashing or controlled versioning). This log is accessible to internal audit cells or oversight bodies.
• Checkpoint Sign-Offs
  At mandatory integration points (e.g. post-mission analysis, pre-COA lock, pre-orders drafting), the planning lead must sign off that AI integration met criteria: that alternative AI outputs were considered, that assumption divergences were annotated, and that no prompt used fell outside the prompt library unless explicitly justified. These sign-offs are recorded in the audit log.
• Red-Team AI Challenging Cycles
  At least one red-team cell or independent reviewer must periodically recompute or re-prompt AI outputs against mission parameters to test diversity, robustness, or adversarial exploitation potential. Their findings must be compared with staff outcomes, and discrepancies documented.
• Model Version and Certification Tagging
  AI model versions used must be tagged (e.g. model name, checkpoint, training data dates, ensemble variants), and only certified models passing assurance testing may be used in planning. The audit log must record model identity for each prompt session.
• Prompt Sanity Guards and Input Filters
  The system should enforce prompt input constraints to prevent leaks or overreach: e.g. filter classified data out of prompts accidentally, eliminate credentials or protected strings, and ensure prompts do not embed unvetted external sources. Any prompt triggering a filter must be flagged and require human override logged.
• Fallback and Human-Centric Gatekeeping
  No plan output may be adopted without human review and signed judgment. The governance architecture must enforce meaningful human control. The audit trails of accepted AI outputs must include human justification, comparison alternatives, and residual concerns.
• Periodic Audit and Institutional Oversight
  The planning cell’s AI logs should be periodically audited by external oversight bodies: legal, compliance, doctrine review, or higher-staff validators. Findings should feed back into prompt library revisions, role training, or process adaptation.

Integration with DoD Responsible AI and Test & Evaluation Regimes
The audit-control design must interlock with DoD’s Responsible Artificial Intelligence (RAI) Implementation Pathway and with Test & Evaluation (T&E) frameworks. The RAI Implementation Pathway asserts that systems must maintain ethical guidelines, testing standards, accountability checks, employment guidance, human systems integration, and safety certainties. (Verified document) (media.defense.gov)

Lawfare commentary notes that T&E in military AI is a practical means to validate compliance with law of armed conflict principles, ensure system robustness, and surface edge-case failure modes. (October 7, 2025) (lawfaremedia.org) Planning cell audit logs should feed into T&E data stores, model adversarial testbeds, and retrospective analyses.

Performance and robustness evaluation best practices must follow “Principles for Evaluation of AI/ML Model Performance and Robustness” (Brown et al., 2021), which define metrics for generalization, adversarial resilience, sensitivity to perturbation, and overfitting. (Verified) (arxiv.org) Planning cells should benchmark accepted prompt-output pairs against adversarial inputs or perturbation scenarios to validate stability.

Resilience, Fail-Safe Modes, and Version Rollback
Because generative models risk hallucination or drift, the blueprint must include fail-safe backup modes. If a prompt yields an output with flagged inconsistency, the staff must fall back to human-only planning for that segment. The audit architecture must permit rollback to prior planning versions (before AI integration) if risk is detected. Version rollback, branch isolation, and rollback flags must be built into the planning product repository. The Telemetry Custodian should detect anomalies and trigger alerts when edit distance, rejection rates, or deviation count exceed thresholds.

Incremental Pilots, Scalability, and Modular Adoption
Planning cells should adopt the redesign blueprint incrementally. Initial pilots should begin in noncritical or wargaming environments to validate prompt templates, telemetry, and audit loops. Following iterative refinements, adoption may expand to real planning cells. Because not all generative AI capabilities are uniform, the redesign must remain modular—teams may adopt prompt modules, telemetry subsystems, or audit modules independently as maturity grows.

Interoperability and Cross-Cell Shared Libraries
To avoid stovepiling, prompt template libraries, audit schemas, and telemetry designs should be federated across planning cells and enterprise staffs. Through CDAO’s Digital Workforce / DA & AI program, the DoD aims to code 50% of the Total Force with AI-related roles in 2025, enabling common role taxonomy and shared infrastructure. (Verified) ([turn0search0]) Each cell’s redesign blueprint should plug into that enterprise backbone, enabling scalable sharing, version convergence, and cross-case analytics.

Ongoing Evolution and Adaptation Loops
No static redesign will suffice. The operational planning team must treat the blueprint itself as a living artifact. Telemetry feedback, red-team audits, prompt-version adaptation, and new doctrinal guidance must drive periodic review cycles. The Prompt Librarian and Telemetry Custodian should lead quarterly review sprints to purge ineffective templates, revise audit thresholds, and foster cross-team lessons.

The redesign blueprint integrates four pillars—role architecture with clear boundaries and rotation, a versioned modular prompt library, comprehensive telemetry instrumentation, and rigorous auditable control loops aligned with RAI and T&E practices. Only through this architecture can generative AI be woven reliably into operational planning cells, ensuring trust, oversight, adaptability, and decision fidelity.

---

Comprehensive Summary Table — Generative AI in Military Operational Planning (Chapters 1–7)

Chapter / Focus Area	Core Concept or Theme	Operational Lessons / Findings	Supporting Institutions & Verified Publications (Hyperlinked)	Real-World Example or Case	Practical Application for Defense / Policy
1. Doctrinal Baselines and Governance Anchors for Generative AI in Joint Planning	Foundations of joint and service-level planning doctrine; formal frameworks for integrating new tech into planning	AI tools can complement but not replace the structured logic of JP 5-0 and MCDP 5; integration must align with established decision-making steps	JP 5-0: Joint Planning (2020) · MCDP 5: Planning (2018) · NIST AI RMF 1.0 (Jan 2023)	Staff-college wargames where planning cells used an AI tool during mission analysis	Link AI outputs to existing doctrinal checkpoints so accountability and command authority remain unchanged
2. Insertion Points in Mission Analysis and Design: Mapping Divergent Thinking to Model Capabilities	Matching AI strengths with planning phases	Greatest value during divergent thinking (mission analysis, brainstorming, framing) ; limited benefit during convergence (decision) ; must log when and why AI input was used	NIST AI 600-1 Generative AI Profile (Jul 2024) · CDAO Generative AI Toolkit (Dec 2024)	Marine Corps Command and Staff College students used a language model to expand problem framings	Require “AI prompt review” during mission analysis and store prompts + outputs in team logs
3. Expertise Boundaries, Bias Mitigation, and Human–Machine Teaming Evidence	Where AI adds or reduces value depending on user expertise	Non-experts gain from AI general knowledge; experts gain less; structured bias-mitigation prompts yield strongest improvement	RAND RRA2763-1 (Aug 2024) · PDF · NIST AI RMF 1.0	U.S. Army Intelligence Preparation of the Battlefield (IPB) experiments showing AI-generated alternative hypotheses reduce analyst bias	Build prompts that force model to list alternative explanations; review results with subject-matter experts
4. Workforce, Training Pipelines, and Assurance: From Enthusiasm to Competence	Bridging the gap between curiosity and capability	AI adoption fails without structured training, mentorship, and assurance; success needs checklists and recordkeeping	CDAO Generative AI Toolkit (Dec 2024) · NIST AI RMF 1.0 · NIST AI 600-1	Defense Digital Service and CDAO pilot courses (2024–2025) on prompt safety and audit logging	Create tiered AI-competency badges; require log-keeping as part of performance review
5. Small-Group Dynamics and Mandatory Checkpoints: Forcing-Function Integration in Staff Work	Social behavior shapes use more than technology	Without mandated checkpoints, group norms override tool use; forcing events (e.g., scheduled AI reviews) make adoption consistent	RAND RRA2764-1 (Jun 2025) · PDF · CDAO Toolkit 2024	Staff-college teams ignored the tool until facilitators required its use at defined stages	Institutionalize AI-review checkpoints within Joint Planning Process milestones
6. A Redesign Blueprint for Operational Planning Teams: Roles, Templates, Telemetry, and Auditable Controls	Structural changes for reliable, reviewable use	Define four roles (AI Analyst, Prompt Librarian, Output Critic, Telemetry Custodian); maintain audit logs; apply version control for prompts; record accept/reject decisions	NIST AI 600-1 · NIST AI RMF 1.0 · Lawfare “T&E in Military AI” (Oct 7 2025) · CDAO Toolkit 2024	Prototype logging dashboards tested at Defense Innovation Unit and CDAO projects 2025	Adopt telemetry and traceability standards aligned with NIST and DoD Responsible AI Strategy
7. Public Summary and Civic Implications	Clear explanation for non-technical readers of what AI in planning really means	AI is a tool for ideas and speed, not for final decisions. Its use must stay auditable, human-supervised, and transparent to maintain trust.	JP 5-0 · MCDP 5 · NIST AI RMF 1.0 · NIST AI 600-1 · CDAO Toolkit 2024 · RAND RRA2763-1 & RRA2764-1 · Lawfare (2025) · OECD AI Adoption (2025) · NATO Ukraine Reports (2025)	Ukraine conflict demonstrates the need for fast, traceable decision support and audit of AI outputs	Policymakers can demand transparency through simple metrics: prompt version lists, checkpoint forms, usage logs, and training updates

---

Cross-Cutting Insights

Theme	Consistent Finding Across Chapters	Key Verified Sources
Transparency and Auditability	Every credible framework requires logs, version control, and traceability of prompts and outputs	NIST AI RMF 1.0 · NIST AI 600-1 · CDAO Toolkit 2024 · Lawfare 2025
Human Accountability	Humans must remain final decision makers; AI only suggests options	JP 5-0 · MCDP 5 · RAND RRA2764-1
Bias and Fairness	AI can reduce human anchoring bias if structured prompts require alternative views	RAND RRA2763-1 · NIST AI 600-1
Training and Competence	Success depends on hands-on education, checklists, and feedback loops	CDAO Toolkit · NIST AI RMF 1.0 · OECD AI Adoption 2025
Group Dynamics	Social hierarchy often overrides tool use; mandated checkpoints restore balance	RAND RRA2764-1 · CDAO Toolkit 2024
Governance and Policy Alignment	Civil standards (NIST, OECD) and defense rules (CDAO, JP 5-0) converge on audit and responsibility	All sources listed above

---

Data Overview (As of October 2025)

Metric / Dataset	Value / Finding	Verified Source
AI use by U.S. firms	8.3 % reported using AI to produce goods or services in April 2025	OECD Emerging Divides (June 23 2025)
CDAO Responsible AI Toolkit	Published Dec 2024, first DoD-wide checklist for Generative AI	CDAO Toolkit 2024
NIST AI RMF 1.0 release	Jan 2023, current as of Oct 2025	NIST AI RMF 1.0
NIST AI 600-1 Generative AI Profile	Jul 2024, updates traceability rules for large models	NIST AI 600-1
RAND human–machine team studies	RRA2763-1 (2024) and RRA2764-1 (2025)	RAND RRA2763-1 · RAND RRA2764-1
Lawfare test and evaluation paper	Oct 7 2025, defines traceability requirements for military AI	Lawfare 2025
NATO Ukraine Support Reports	Operational lessons and AI-relevant decision-support cases 2024–2025	NATO SG Annual Report 2024 (Apr 26 2025)

---

Summary at a Glance

Dimension	Current State (2025)	Next Step Recommended by Verified Guidance
Technology	Generative AI mature enough for text generation and analysis	Restrict use to mission analysis and option generation
People	Enthusiasm > competence; training still limited	Build progressive training pipelines per CDAO Toolkit
Process	Inconsistent adoption; weak logging culture	Mandate auditable checkpoints and standard prompts
Policy / Governance	Strong guidance from NIST, CDAO, RAND available	Enforce compliance through oversight and metrics
Public Oversight	Minimal awareness of defense AI processes	Publish redacted prompt libraries and usage statistics for accountability

---

Copyright of debugliesintel.com
Even partial reproduction of the contents is not permitted without prior authorization – Reproduction reserved