Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide.
If there is a revolution brewing in the automation of white-collar work, you wouldn’t know it at the Pentagon. Here, the typical action officers, staff who tackle the daily work of the bureaucracy, perform essentially the same duties as their predecessors in 1996 or 1956: creating charts, authoring papers, delivering briefings, and coordinating across offices. Given the promise of AI tools, allowing this staff ecosystem to remain stuck in the past represents a dangerous failure to adapt, especially as senior leaders criticize the workforce for lacking organizational speed. Having served as an action officer across the White House, Pentagon, and intelligence community, I am confident that national security agencies can automate much of this work for greater efficiency and better decisions.
However, automation is not a binary choice. Too little, and an agency may congratulate itself even as its processes remain archaic. Too much, and an agency will atrophy the critical thinking skills of its workforce while losing the chain of human accountability that is foundational to good government. Too inconsistent, and the staff will grow desynchronized as interdependent teams work at vastly different speeds. The question for staff automation is not, “should we,” but “where should we?”
Answering that question requires deliberate thinking about the role of the action officer, the work they perform, and their operating environment. To build a flexible, responsible system for automating action officer duties, agencies must first answer the classic question posed in the film Office Space: “What would you say you do here?”
The action officer’s ecosystem is a series of interactions between unpredictable external problems and rigid internal procedures. Modern national security agencies are classic bureaucracies: hierarchical, specialized, process-driven, and mostly merit-based, with tasks rigorously documented by email, software, or paperwork. Bureaucratic processes, like automated ones, perform best where workflows are predictable and data-driven. In contrast, many strategic assignments handed to action officers are complex, ambiguous, and demand high degrees of qualitative analysis or abstract reasoning. The action officer lives between these fluid external inputs and static internal processes, where frustration at the pace of change often builds. Navigating this tension requires mastering the end-to-end workflows of strategic action.
A typical workflow comes in five parts, beginning with a requirement, or task. While some requirements are predictable, annual, or codified, others emerge spontaneously, verbally, or partially formed. Once a requirement emerges, agencies make an assignment, determining who will perform or coordinate the work. The assigned work then undergoes analysis, a process of framing a problem, breaking it down into manageable parts, and tying those parts to deliverables. The deliverable is typically a paper, course of action, or decision brief. Finally, implementation involves follow-through or continuous monitoring until completion. Together, these five stages represent the target areas for staff automation.
While technical solutions to optimize this process already exist, like staffing dashboards and decision support software, the concepts and guidance needed to effectively scale do not. Establishing shared concepts is critical, given AI’s risks: degraded team performance during complex decision-making, discarded human contextual insight, the lack of trust in valid machine outputs, or a dangerous surrender to automation bias during times of high stress.
To illustrate the informed thought that’s needed, consider a task that once crossed my desk: a well-intentioned effort to merge disparate operational and intelligence reporting streams into a new decision support tool before a fixed date. Remarkably, this action managed to exhibit fatal flaws across all five stages of the staffing process, ending a complete failure.
It failed in its requirements: The action came in a memo from the department deputy, signed just days before they departed office, and without identified points of contact, severing the chain of accountability.
There is no automated substitute for human responsibility. AI tools can augment incomplete guidance or prompt executives for needed clarity, but task ownership should remain human, preserving a chain of accountability from action officers back to executives. Absent this linkage, actions easily lose legitimacy. Conversely, there is little reason an agency should fail to meet a timeline for a requirement that is predictable or statutory, though they routinely do. Automating the initiation of such tasks is one area of quick gain with little downside.
It was misassigned: By the time the action landed on my desk, it was weeks overdue. Misassigned several times, it spends days languishing at each office.
This is the use case where an AI tool will consistently outperform a human. Work is usually assigned based on trust, proximity, or, more formally, official areas of responsibility. While routing is often left to an executive secretariat office, an automated tool with access to internal agency organizational charts, regulations, and terms of reference documents will outperform a human staff officer both in speed and accuracy. Companies have also found success by using AI tools to assign tasks based on employee availability or past performance. For tasks that do not neatly align to specific portfolios, automated tools can flag them for human review.
The analysis was insufficient: The task was, in some ways, proposing something genuinely new. Lacking a clear precedent, it demanded original thought and healthy skepticism.
This is the riskiest step to automate, as research has shown that AI tools can compromise human critical thinking skills. Action officers should begin with their own method of problem framing before turning to automation. The action officer should conceptualize the core questions — who, what, when, where, and why — and only then use AI tools to dive deeper, break the problem into parts, and tie them to deliverables. Overly automating analysis breaks the chain of accountability and leads to a staff that can’t explain how they reached a conclusion beyond copying and pasting model text. Moreover, critics have already pointed to a “yes men” problem in hierarchical organizations. Outsourcing more thinking to AI models that are designed to please and flatter their user will only compound this echo chamber effect. More useful is an AI tool to serve as a dedicated devil’s advocate to critique the action or plan. The originators of our example action could have well used such a tool, and indeed, research indicates that gains in efficiency from AI are concentrated among employees with weaker skills and less experience
Deliverables were infeasible: The deliverables depended on non-existent technical infrastructure and levied burdensome reporting requirements on senior executives.
Delivering finished work is often more art than science. In this example, the burdensome reporting requirements further incentivized agency executives to oppose the program. Automation can’t compensate for a lack of organizational awareness or lack of funding, but it can save time. Once agentic AI can reliably convert natural language instruction into finished products, the work efficiency of the typical national security action officer will certainly increase. Automation can also help tailor products like talking points for a senior official by using speech pattern analysis of their previous correspondence rather than offering the “trendslop” of an AI steeped in the empty jargon of national security. The more explicit the preference, the more suitable the deliverable.
Implementation was disjointed: The action laid out rigorous metrics for implementation tracking, but with no standing oversight body or touchpoints.
Actions require follow-through, and implementation tracking certainly presents opportunities for increased automation. AI agents can prompt action officers to input regular updates for a certain project, and then instantly consolidate collective responses into a polished status report with updated timelines. However, the pressure or urgency to achieve an objective resides with humans and is dependent on understanding organizational priorities, culture, and resources. Automation can aid implementation, but it can’t force people to care.
Ultimately, our example action failed quickly and quietly. Automation would have improved parts such as assignment and analysis, but not the outcome. More likely, excessive automation would have reinforced human failure and digitally legitimized a bad idea. The core challenge for agencies designing staff automation concepts is to preserve human responsibility while augmenting practicable actions, and to guard against ossifying unworkable ones.
Individually, most automation steps will constitute measured change rather than revolutionary transformation. Yet collectively, they have the potential for two fundamental changes to the action officer ecosystem.
First, a transformation from a sequential system to a simultaneous one, where steps can occur out of order. Operational planners and military strategists have already recognized the shift away from phasing, and so should strategic staffs. Departing from the phased process model will be difficult, but using multi-agent orchestration to overcome the basic constraint of “first this, then that” has the potential to maximize decision-making against fixed timelines, especially for those “moderately rare” problems where AI gains are largest. For our earlier example, multi-agent orchestration coordinates specialized models in parallel under human oversight: a requirement agent flags inexecutable tasks while an analysis agent scopes goals and a routing agent alerts potential stakeholders early, collapsing the delays and friction of linear handoffs.
The second change presents higher risk and higher reward. Traditionally, agencies are prone to force external problems into predictable internal processes, even where suboptimal, a dynamic referred to as “grooved” thinking. The potential of an automated system, empowered with appropriate flexibility and bounded ontologically by agency value statements and priorities, is to move to a model where the process adapts to the problem. For example, an automated tool trained in organizational priorities might recognize a case where the cost of a delayed decision outweighs the cost of incomplete coordination or analysis. Cutting specific offices out of the action process in the interest of mission accomplishment requires a lack of ego that an AI tool is well-suited for. However, this entails a significant degree of political danger, potentially reduced stakeholder buy-in, and the risk of homogenized thought. Nonetheless, I can attest to instances where a late deliverable presents the worst consequences, or where a chief executive’s priority needs staff feedback without suffering death by committee. To balance this risk, automated prioritization is best left to “old-fashioned” symbolic AI tools that use explicit logic chains and human-authored rules, preserving at least some form of traceable accountability for risk delegation to machines.
AI tools are beyond nascent stages, but there is still time to build concepts and train personnel prior to widespread adaptation. The need for base concepts that enable efficiency while preserving human responsibility is paramount. A future where agencies learn to offer a variation of “the computer did it” as a deflection for responsibility represents a nightmare scenario for good government and healthy organizations. AI tools are fundamentally not responsible entities, and true accountability depends on more than who authored a prompt.
Because tools that rely upon natural language prompting have a low barrier for use, agencies will be tempted to ignore training, but action officers should possess a baseline conceptual literacy in avoiding automation bias, understanding model function, and recognizing errors. Beyond general staff, dedicated AI quality assurance personnel can ensure that alterations proposed by automated tools make sense in context and are not merely process hallucinations. Specialty personnel who bridge gaps between the technical and general staff will also be key to educating a workforce with an unrealistic understanding of AI tools. Absent this crossover, agencies risk turning over their governance to the type of “scientific-technological elite” that Dwight Eisenhower warned of, as the gap between those with the power to decide and those with knowledge to execute grows wider.
There are also unique government fiscal, regulatory, and legal constraints to overcome. Foremost among these is system classification. Current contracts already demonstrate how AI tools can integrate at scale into unclassified and controlled unclassified systems at all five stages, merging dozens of data sources. However, integrating AI tools into classified systems poses a bigger obstacle due to air-gapped networks, divergent architectures, and different mission requirements. With those barriers, agency concepts should recognize that unclassified workflows are likely to outpace classified or interagency ones.
Bureaucracy can serve as a positive force for complex organizational governance when employed to a certain measure. Yet its inherent self-perpetuation and resistance to change mean that bureaucracy is almost always found in excess rather than an optimal amount. I am optimistic that automation can help make the action officer’s bureaucracy less excessive, but certain that only human judgment can determine to what ends.
Samuel Canter, Ph.D., is an adjunct professor at George Mason University’s Schar School of Policy and Government, a civilian strategic planner at the Pentagon, and a military historian serving in the U.S. Army Reserve. His research focuses primarily on national security organization, bureaucracy, and culture.
The views and opinions presented here are those of the author and do not necessarily represent the views of the U.S. Army, the Department of Defense, or the U.S. government.
Image: Midjourney
