At a recent United States Geospatial Intelligence Foundation (USGIF) Mission Focus event, I participated in a working-group discussion about where artificial intelligence should be applied to intelligence pre-processing. It was a reasonable question, and the answers were largely predictable. Participants pointed toward the cloud, data conditioning, edge computing, sensor processing, and the infrastructure needed to move information more efficiently. Each is a legitimate place to invest. Yet the discussion seemed to assume that whatever improvements we made upstream would ultimately feed the same human-intensive intelligence production model.
I argued for a different starting point: the analyst. That answer was deliberately uncomfortable. The imagery analyst has become one of the most consequential constraints in the intelligence lifecycle, partly because the enterprise has organized so much of its production around what a human can observe, verify, format, and disseminate. The criticism was directed at the architecture surrounding the profession, including one I served in, rather than at the people performing the mission. As technology advances, preserving every intermediate human task is increasingly difficult to justify.
The argument deserves to go further. If AI can continuously interpret requirements, recognize changes in activity, correlate observations, recommend additional collection, and deliver relevant intelligence to authorized users, then the Intelligence Community (IC) faces more than a software modernization challenge. It faces a reckoning with the workforce, contracting structures, measures of productivity, and institutional habits built around yesterday's processing limitations. The mission must determine which roles remain necessary. The existence of a role cannot itself be the justification for preserving it.
The Workforce We Built to Absorb the Data Deluge
During my Air Force service from 2005 to 2012, I worked across imagery and ISR missions involving Global Hawk, U-2, Predator, and satellite collection. Those years belonged to an extraordinary expansion in the demand for persistent surveillance. Operations in Iraq and Afghanistan increased the appetite for imagery, full-motion video, and rapidly updated observations. Collection capacity grew, but the data still needed to be exploited. The practical response was to build and sustain a substantial human processing apparatus.
The Air Force Distributed Common Ground System (DCGS) embodied the advantages and compromises of that period. Its distributed, reachback-oriented approach made it possible to bring specialists, exploitation tools, and multiple intelligence sources to bear on operations thousands of miles away. It provided real operational value. At the same time, it reinforced a model in which substantial numbers of people in rear locations supported a smaller number of forward intelligence consumers and decision-makers. A great deal of labor existed between the sensor and the operational question.
Many tasks were essential but repetitive: scanning imagery for objects, counting vehicles or aircraft, comparing successive observations, maintaining tracks, producing annotated graphics, moving coordinates between systems, updating databases, and preparing recurring reports. These activities can contain analytical judgment, especially when the environment is ambiguous. But their routine execution frequently consumes the attention of professionals trained for much more sophisticated work. From my experience, the balance could feel overwhelmingly weighted toward what analysts call 'bean counting.' I have used the figure of 80 percent to describe that experience, but it should be understood as an operational observation, not a validated estimate for the entire imagery analyst corps.
From my experience inside that reachback architecture, a powerful way to understand its human footprint is through a 30:1 illustration: dozens of intelligence personnel across functions and shifts supporting an intelligence requirement ultimately serving a far smaller number of forward operational consumers. The exact staffing relationship changes with the mission, but the larger pattern is unmistakable. Intelligence analysts often produce information for other analysts, who in turn serve additional intelligence customers, before the result reaches a commander or operational decision-maker. Each layer may require exploitation, validation, correlation, reporting, and coordination. The 30:1 illustration exposes the magnitude of the supporting enterprise that can accumulate behind a seemingly simple intelligence question. If AI removes substantial labor at several of those layers simultaneously, the workforce implications could be dramatic. A process that once drew on dozens of personnel may eventually require only a handful, depending on mission risk and required human judgment. The ratio is illustrative, but the structural challenge is real: we must question how much of that intermediate human apparatus will remain necessary. Even that 30:1 relationship represents only one segment of a much larger intelligence architecture. From initial collection through exploitation, fusion, assessment, and dissemination, successive teams and organizations support other intelligence customers, multiplying the human roles and handoffs between a sensor observation and the ultimate operational decision.
The Air Force is already moving beyond some of the old platform-centered practices. In 2025, its DCGS enterprise completed a transition to Analysis and Exploitation Teams, emphasizing problem sets, multi-source information, and tailored products rather than isolated platform outputs. That is meaningful progress. It also reveals the next question: how much of the problem-solving process must still be assembled and sustained through manual human coordination when machines can increasingly maintain context across systems and collections?
My Air Force experience provides a particularly visible example, but the institutional problem extends across the National System for Geospatial Intelligence and the wider IC. Army GEOINT personnel support tactical formations with imagery exploitation, targeting support, collection coordination, and geospatial production, often under severe time constraints. Recent Army writing identifies imagery PED as one of the most time-intensive intelligence demands. NGA operates at a different scale, sustaining national-level imagery analysis, structured observations, activity-oriented intelligence, and foundational geospatial knowledge. The specific functions and analyst-to-customer ratios differ across organizations. What repeats is the layered dependence on people to create, validate, maintain, reconcile, and pass along intelligence that other analysts and customers must then interpret.
NGA is especially important to this discussion because so much of the broader intelligence enterprise depends on the geospatial information, observations, and analytic services it produces. Structured Observation Management (SOM) captures observations in standardized forms so they can be stored, shared, compared, and reused. Activity-Based Intelligence (ABI) connects observations across time and context to identify patterns and meaningful changes. These approaches are vital to machine-usable intelligence, yet they can also reveal the enormous human effort associated with populating records, validating attributes, reconciling contradictory observations, and ensuring that downstream systems receive usable information. Foundation GEOINT adds another enterprise-scale workload: terrain, elevation, transportation, infrastructure, feature, reference, and other geospatial information must be created, refreshed, and quality-controlled across vast areas. NGA's AI modernization efforts already acknowledge the relevance of SOM, computer vision, and automated reporting. The next question is whether future workflows should continue allocating people to every intermediate data-production and verification step, or use automation to preserve the quality and utility of those products with a fundamentally different workforce model.
The Difference Between Exploiting Data and Answering a Question
An intelligence system can be extraordinarily productive without being particularly effective. More hours of video watched, more images exploited, and more reports issued do not necessarily produce a better decision. Much of the enterprise measures activity because activity is observable and contractually manageable. The harder question is whether the activity reduced uncertainty about a mission-relevant problem.
This is not a hypothetical criticism. A 2021 RAND study of Air Force airborne ISR found that the service lacked a consistent, empirically grounded method for assessing the value its ISR provided in operational contexts. The report described limited feedback from end users, inconsistent assessment approaches, and difficulty connecting ISR activity to overarching goals. Those findings should be central to any discussion about AI. Automating a workflow whose operational value we struggle to measure may make the enterprise faster without making it wiser.
Consider a facility where analysts are tasked to count aircraft each morning. A recurring count may establish a useful baseline. But the deeper intelligence requirement is rarely the number itself. A commander may need to know whether a unit is generating combat power, preparing for deployment, changing its readiness posture, or attempting deception. The count is evidence in support of those questions. When we organize the workflow around producing the count, the supporting task can become detached from its purpose.
Computer vision can identify and count visible objects, with limitations that must be measured. Change detection can highlight unusual activity. Time-series methods can distinguish persistent patterns from new ones. Language models may help retrieve and organize reporting. None of these capabilities alone establishes adversary intent. Together, under an appropriately validated architecture, they can reduce the amount of routine exploitation a human performs before reaching the actual intelligence question.
That distinction changes how we should view the analyst bottleneck. The analyst is the point where disconnected requirements, data, tools, approval chains, and reporting obligations often converge. Replacing individual tasks while preserving all those dependencies may merely move the queue from one workstation to another. The more valuable transformation would remove unnecessary steps and maintain the connection between the operational requirement and the evidence used to satisfy it.
Start with the Requirement, Not the Sensor
The most underexplored opportunity for AI lies at the beginning of the intelligence lifecycle. Intelligence requirements are intended to articulate what commanders and decision-makers need to know. Collection plans, tasking, processing, exploitation, analysis, and dissemination should follow from those needs. In practice, requirements can fragment across organizations and systems. They may persist long after their usefulness has changed, overlap with other requests, or be translated into a succession of tasks that lose the original operational context.
Imagine an operational requirement to determine whether an adversary is preparing to reposition a formation. Existing reporting, satellite imagery, radar observations, signals indicators, weather conditions, and historical behavior may all bear on the question. A future AI-supported requirements system could maintain a record of what is known, what remains uncertain, which observations would be most informative, and whether an additional collection is justified. It could continuously monitor changes in the evidence, recommend the next collection opportunity, and notify the responsible intelligence team when confidence materially changes.
The opportunity begins before a new sensor is tasked. The system might discover that existing imagery and reporting already answer part of the requirement. It might identify duplicative requests from different commands, determine that a scheduled revisit is unlikely to add value, or flag an assumption that has not been tested. When a new observation contradicts the current assessment, it could recommend cross-cueing another sensor or raising the issue for human review. The objective is to conserve attention, collection capacity, and communications bandwidth while improving what the mission actually knows.
This should not be mistaken for granting a language model unrestricted authority to direct scarce collection assets. Sensor availability, orbital geometry, airspace restrictions, tasking priorities, classification boundaries, data rights, communications, and command authorities remain real constraints. A credible architecture would combine structured requirements, rules-based controls, statistical models, machine learning, and auditable recommendations. Some actions may be automated within established authorities; others would require collection-manager or commander approval. The distinction is an engineering and governance decision, not a slogan about autonomy.
This is also why edge processing and cloud modernization remain important but insufficient. Processing an image onboard a satellite can reduce data movement or alert latency under the right conditions. It does not, by itself, establish whether the image was worth collecting. Likewise, an exquisitely curated cloud repository is of limited mission value if no system connects its contents to an active intelligence requirement. The strategic advantage comes when requirements management, collection, exploitation, and assessment reinforce one another continuously. RAND's research on AI-supported sense-making provides useful context for this broader analytical challenge.
A Requirement-Driven Workflow in Practice
Consider a hypothetical, unclassified scenario in which an intelligence team needs to assess whether unusual activity at a major logistics hub signals a sustained change in throughput. In a familiar production model, collection requests generate images, imagery analysts perform repeat counts of vehicles and storage areas, observations are entered into reports, and another intelligence cell attempts to assemble a pattern from successive products. Additional analysts may be asked to recheck the same imagery because the first report is incomplete, lacks context, or conflicts with another source. None of those handoffs is inherently illegitimate. Together, they can turn a straightforward operational question into a large recurring labor commitment.
In a requirement-centered system, the question remains the organizing object. Approved software retrieves existing observations, compares activity against the historical baseline, documents image quality and collection gaps, and evaluates whether new collection would resolve meaningful uncertainty. Computer vision may perform the routine counts while independent checks examine temporal consistency, geolocation, and unexpected objects outside the original categories. A structured record preserves each observation's source, confidence, and relationship to prior reporting. If a genuine anomaly emerges, the system can recommend a follow-up observation under established collection authorities and elevate the discrepancy to a qualified analyst.
The analyst then spends less time reconstructing a spreadsheet of counts and more time asking what changed, whether an alternative explanation fits, what the infrastructure can actually support, and how much confidence decision-makers should place in the result. Where errors, novelty, or consequences warrant it, people intervene earlier and more deeply. Otherwise, established rules permit routine processing to continue without requiring another person to retype, route, and endorse every intermediate product. The exercise is hypothetical, and the actual mix of automation and approvals would depend on the mission. Its value is to reveal what changes when the intelligence question, rather than the production queue, controls the workflow.
The Coming Reckoning with Manpower
Here is the uncomfortable implication that deserves to be stated plainly: a successful transition will eliminate some intelligence production positions. It is implausible that every role displaced by automation will be replaced by an equivalent position supervising algorithms. The IC may need more specialists in model evaluation, data engineering, adversarial testing, mission integration, and complex analysis. It does not follow that it will need one such specialist for every person previously assigned to routine exploitation or reporting.
The reassuring language of 'human-machine teaming' can obscure the economics. It describes a valuable operating method, but it is not a workforce plan. If a task that once required dozens of shifts, handoffs, or recurring manual actions can be performed with a smaller oversight team, the labor requirement changes. The relevant questions are how much labor is actually removed, what new duties emerge, how reliability changes, and whether the mission benefits. Those answers will vary across tasks and agencies. They cannot be resolved by an invented universal replacement ratio.
There will be resistance, and not all of it is irrational. Intelligence units must maintain continuity during conflict, preserve specialized knowledge, sustain surge capacity, and meet security obligations. Automation systems introduce new failure modes and may require human fallback capacity. Leaders who have experienced brittle technologies or unreliable procurement promises are right to demand evidence. Yet these constraints should inform deliberate transition plans rather than become a standing argument for preserving each existing billet and workflow.
Contracting incentives also matter. A service contract built around providing cleared labor rewards staffing capacity in a way that may differ from a contract built around validated mission outcomes. Vendors selling faster exploitation tools may have little incentive to question whether the downstream reporting requirement should exist. Government program offices may be evaluated on delivery, utilization, and compliance measures that are easier to count than improved operational decisions. An AI transformation that ignores these incentives could spend billions modernizing the machinery of yesterday's workflow.
The answer is neither abrupt wholesale cuts nor permanent protection from change. Organizations should establish baselines for labor by task, validate machine performance against real operational conditions, redesign the process, and then determine the necessary workforce. Retraining and reassignment should be pursued where new mission demands justify them. Some billets can migrate to higher-value work; others may disappear through attrition, contract redesign, or explicit force-structure decisions. Pretending that all can be retained risks diverting resources from capabilities the mission will need.
Where Human Judgment Earns Its Place
The argument for reducing routine human processing should not be confused with the claim that AI can independently resolve the most consequential intelligence questions. Object recognition is not the same as interpreting intent. A model may accurately detect a vehicle and still misunderstand its significance. A change in adversary behavior may reflect deception, political signaling, weather, logistics, or an entirely new tactic outside the model's training experience. Faster automated outputs can also propagate errors across systems before anyone recognizes the original mistake.
The Human Quality-Control Double Standard
Cultural resistance to AI often begins with a valid observation: automated systems make mistakes. A detector can misclassify a vehicle, miss a concealed object, or carry forward an erroneous geolocation. The conclusion is then that a human analyst must inspect the output because humans are the trusted quality-control mechanism. What receives less scrutiny is the fallibility of the human process itself. Analysts misidentify equipment, transpose coordinates, overlook objects under workload pressure, inherit incorrect assumptions, and sometimes repeat earlier reporting without independently checking its basis. Supervisors and subsequent consumers may catch those errors, but they can also institutionalize them by treating an apparently reviewed product as authoritative.
A single observation can pass through exploitation, correlation, database entry, reporting, fusion, command briefings, and follow-on collection decisions. Each handoff creates another opportunity for omission, distortion, mistaken transcription, or unwarranted confidence. Once a faulty observation enters a structured record, later analysts may unknowingly use it as corroboration, creating a circular evidentiary chain. The downstream influence can compound as more products and decisions depend on the same initial error. This is not evidence that human errors necessarily grow exponentially or that AI is automatically safer. It is a reason to compare complete systems, including the probability of errors escaping detection and the scale of their consequences, rather than compare an AI model's isolated error rate with an idealized human reviewer.
Human quality control is not free, instantaneous, or perfectly independent. In an environment where sensors can generate more observations than personnel can inspect, mandatory review of every item may lower real coverage while creating a comforting appearance of assurance. A different institutional bias can operate in the opposite direction: an organization can become so committed to human sign-off that it fails to ask whether repeated manual intervention is adding information or merely delay. RAND research on potential human bias in Army intelligence preparation of the battlefield reinforces why human cognition belongs in the same quality-control discussion as algorithmic limitations. The appropriate comparison is among measured alternatives, including human-only, AI-assisted, and carefully bounded automated workflows. Each should be tested for accuracy, omissions, error propagation, reviewer workload, and mission consequence.
Human judgment should therefore be allocated according to the consequence and uncertainty of the task. Routine, well-characterized observations may support automated handling with periodic sampling and audit. Ambiguous or novel patterns may require trained analysts to challenge assumptions and consider alternative explanations. Assessments with substantial strategic or operational consequences deserve stronger review, clear source pedigree, confidence statements, and an accountable decision process. The burden of review should be proportional to risk rather than automatically attached to every intermediate step.
Layered Assurance Without a Human at Every Step
This is also where specialized AI agents could add an important layer of quality control. One automated process may maintain routine counts and structured observations while another deliberately searches the same imagery for outliers that fall outside predefined object lists or established activity patterns. Separate agents could check geospatial alignment, temporal consistency, duplicate observations, source provenance, and contradictions with previous assessments. In an ABI workflow, the most valuable alert may concern something the original collection requirement never anticipated. That possibility came up briefly in our USGIF discussion and deserves more attention than another round of object-counting demonstrations.
Additional layers could screen products against applicable policy constraints, identify possible law-of-armed-conflict concerns for qualified reviewers, and verify that an intelligence conclusion does not exceed the available evidence. These would be triage and assurance functions, not substitutes for legal advice, command authority, or human determinations about the use of force. ODNI's analytic standards provide a useful benchmark for source credibility, uncertainty, alternative explanations, and distinguishing evidence from judgment. Agents can make those checks more systematic and scalable, but they need traceable rules, testable thresholds, escalation paths, and auditable records.
Multiple models agreeing with each other is not proof of correctness. Agents trained on similar data, relying on the same flawed map, or sharing one underlying model can reproduce correlated errors and create false confidence. A robust design should use genuinely independent evidence where available, cross-modal checks, adversarial tests, human sampling, and explicit monitoring for distribution shifts. It should also preserve negative findings and uncertainty rather than discard observations that do not match the model's expectations. Quality control should be judged by the faults it catches and the harmful errors it prevents, not the number of automated reviewers added to a workflow.
DoD Directive 3000.09 addresses autonomy in weapon systems and the requirement for appropriate human judgment over the use of force. It should not be casually treated as a blanket rule governing every intelligence-processing activity, nor as permission to automate all earlier stages of target development. Different intelligence, targeting, and weapons functions operate under different authorities and obligations. The Department's Responsible AI framework likewise emphasizes accountability, traceability, testing, and governance. An AI-first intelligence enterprise must be designed around those requirements, including the ability to contest outputs, audit model behavior, and stop an automated process when its assumptions fail.
The remaining human analysts may become more valuable precisely because their work is harder to automate. They will need deeper domain expertise, strong analytic tradecraft, an understanding of model limitations, and the ability to interrogate machine-generated assessments. The goal should be to give these professionals genuinely analytical responsibilities, not merely turn them into an endless queue of people clicking 'approve' on outputs they lack the time or context to scrutinize. Human oversight without the capacity for independent judgment is procedural theater.
The Tradecraft We Cannot Afford to Lose
One serious cost of a largely automated intelligence lifecycle could be the erosion of tradecraft itself. If the next generation of GEOINT professionals rarely performs manual exploitation, interprets raw sensor signatures, works through uncertain geolocation, or traces how a conclusion was derived from original observations, they may struggle to recognize when a machine is wrong. A profession that no longer understands the evidence beneath its automated products becomes dependent on systems it cannot meaningfully challenge. That would be a strategic vulnerability, especially when adversaries deliberately exploit assumptions, sensor limitations, and model behavior.
Preserving tradecraft requires defining it carefully. Some of what we call tradecraft is knowledge of sensors, geometry, uncertainty, source reliability, adversary behavior, collection limitations, structured analytic reasoning, and the evidentiary chain connecting an observation to an assessment. Those competencies remain essential. Other elements are learned procedures for navigating particular software menus, manually copying coordinates, formatting reports, or reconciling incompatible databases. Such workflows have historically consumed enormous training time and daily attention, but their continued manual performance should not be confused with professional mastery. We should preserve the ability to reconstruct and independently test intelligence judgments without requiring every analyst to spend a career repeating tasks better handled by automation.
The potential upside is substantial. An analyst development program freed from excessive workflow instruction could invest more deeply in how the world operates: military organizations and doctrine, critical infrastructure, logistics networks, industrial production, economic geography, political incentives, historical patterns, and the distinctive signatures of human activity. GEOINT specialists would still learn what electro-optical, radar, infrared, and other sensors can and cannot establish. They would also spend more time understanding why a particular observation matters, what alternative explanations exist, and which additional evidence would change their assessment. In that sense, AI could strengthen the intellectual core of the profession even as it reduces the workforce needed for routine exploitation.
That outcome will not happen automatically. Organizations should preserve foundational manual skills through deliberate training, periodic proficiency checks, red-team exercises, and scenarios in which automated tools fail or are unavailable. Analysts must be able to return to raw evidence, identify poor data and flawed model assumptions, explain confidence levels, and challenge an apparently authoritative output. A smaller cadre cannot merely know how to operate AI systems; it must retain enough domain and sensor expertise to audit them independently and teach the next generation. The purpose is not nostalgia for older workflows. It is operational resilience and the preservation of knowledge that technology may otherwise conceal.
The deeper shift is from teaching analysts primarily how to produce intelligence products toward teaching them how to understand intelligence problems, evaluate the inputs and outputs of increasingly automated systems, and recognize their mission significance. Today's training and operations can allow software procedures and data munging to compete with, and sometimes crowd out, contextual understanding. Over the next five to ten years, that balance should reverse. If we accomplish it, the profession may employ fewer analysts while demanding more from each one. Tradecraft could become stronger, provided the community consciously preserves the technical foundations required to question its machines.
Measure the Mission, Then Redesign the Enterprise
A practical transition should begin with a disciplined accounting of intelligence work. For each recurring requirement, organizations should identify the decision it supports, the data it consumes, the labor involved, the time it takes, the confidence of its conclusions, and whether the end user finds it valuable. That exercise will reveal tasks that are essential, tasks that are candidates for automation, and tasks that persist mainly because they have always been performed. RAND's warning about weak ISR value assessment makes this an institutional priority rather than an optional performance exercise.
Pilot programs should be evaluated on how effectively they satisfy requirements, not merely how many objects they detect. Useful measures include time from requirement to decision-relevant insight, the reduction of redundant collection, the proportion of requirements answered with existing information, validated error and miss rates, the burden of human review, and feedback from operational users. An honest test must also count the failures: false reassurance, missed anomalies, data-quality gaps, adversarial manipulation, and the cost of maintaining human fallback paths.
These experiments should include procurement and organizational redesign from the beginning. A model that reduces exploitation labor but leaves every approval layer, reporting obligation, funding line, and contract incentive untouched will produce only a partial return. The intelligence enterprise should be able to retire a requirement, discontinue a product, reallocate a collection asset, and reduce a labor commitment when evidence shows that doing so improves mission performance. Those are institutional decisions, and software alone cannot make them.
The next five to ten years will not bring uniform automation across GEOINT, SIGINT, all-source analysis, and every command. Some activities will change rapidly; others will remain constrained by data quality, authorities, contested communications, specialized tradecraft, and the consequences of error. The pace will vary, but the direction already demands planning across the IC. NGA's work on automated GEOINT production and AI-enabled analytic services, Army investments in expeditionary intelligence capabilities, and the Air Force's shift toward problem-centered exploitation are different expressions of a common transition. The harder challenge for every organization is determining which manual tasks, products, reporting obligations, and staffing structures no longer add enough mission value to justify their cost.
Mission Over Manpower
My argument at USGIF began with the analyst because that is where the limits of the current architecture become most visible. The same challenge extends through NGA's national GEOINT production, Army intelligence organizations, Air Force reachback operations, and the many civilian and contractor teams serving one another across the IC. We built labor-intensive processes to absorb growing collection volumes during an era when machines could do far less, a challenge documented well beyond Air Force ISR in RAND's examination of the Navy's sensor-data flood. Entire organizational functions, contracts, technical systems, and professional specialties emerged around the need to process, transform, and redistribute information. This enterprise served critical missions and developed extraordinary expertise. Its historical value does not guarantee that its present scale or structure is the best way to satisfy tomorrow's intelligence requirements.
AI will undoubtedly help analysts work faster, and some of those improvements will be worth pursuing. The greater opportunity is to organize intelligence around persistent operational questions, intelligent management of evidence, selective collection, and timely delivery of conclusions to those responsible for decisions. When the process is redesigned on those terms, many manual steps will shrink or disappear. So will some positions built to perform them. The responsibility of leadership is to manage that transition rigorously and honestly, protecting essential judgment and operational resilience while refusing to treat legacy manpower as an end in itself.
The quality-control test applies equally to human and machine contributions. If the mission can demonstrate that automated workflows deliver better coverage, earlier detection of anomalies, clearer provenance, and fewer consequential errors under real operating conditions, maintaining the previous human workload becomes difficult to defend. If automation fails those tests, it should be constrained or redesigned. No system will be perfect. The governing obligation is to raise the measurable effectiveness and efficiency of intelligence while preserving accountable human authority where it matters.
The responsibility to taxpayers belongs in this calculation. Government intelligence organizations exist to perform public missions under lawful authority, and the resources entrusted to them should be directed toward the most effective and responsible means of doing so. Employment within an institution is not an independent mission objective. That does not erase the government's obligations to its people, including fair transition planning, readiness, and retention of critical expertise. It does mean that preserving a labor-intensive process simply to preserve the associated positions is a weak justification when evidence demonstrates a better way to deliver the mission. This argument would apply even if the next transformative technology were something other than AI. The standard is public value, not loyalty to a particular tool.
The coming reckoning is ultimately a question of institutional purpose. Is the intelligence enterprise willing to measure its success by what it enables the mission to understand and do, even when that means dismantling familiar workflows, reducing established labor requirements, and changing the economics of intelligence production? If the answer is no, we may acquire remarkable AI capabilities only to make an outdated production system run faster. If the answer is yes, we have an opportunity to build an intelligence lifecycle organized around the decisions it exists to support. Preserving legacy manpower is not an intelligence objective. Delivering the mission is.
-
U.S. Air Combat Command, “AF DCGS AET shift provides tailored analysis to Air Components,” Dec. 18, 2025. https://www.acc.af.mil/News/Article-Display/Article/4363766/af-dcgs-aet-shift-provides-tailored-analysis-to-air-components/
Abbie Tingstad et al., Assessing the Value of Intelligence Collected by U.S. Air Force Airborne Intelligence, Surveillance, and Reconnaissance Platforms, RAND Corporation, 2021. https://www.rand.org/pubs/research_reports/RR2742.html
Department of Defense, Directive 3000.09, Autonomy in Weapon Systems, Jan. 25, 2023. https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf
Department of Defense, Responsible Artificial Intelligence Strategy and Implementation Pathway. https://www.ai.mil/Portals/137/Documents/Resources%20Page/DoD%20Responsible%20AI%20Strategy%20and%20Implementation%20Pathway.pdf
National Geospatial-Intelligence Agency, “VADM Whitworth Delivers Keynote at 2025 GEOINT Symposium,” May 21, 2025. https://www.nga.mil/news/VADM_Whitworth_Delivers_Keynote_at_2025_GEOINT_Sym.html
National Geospatial-Intelligence Agency, GEOINT Basic Doctrine, Publication 1.0, including definitions of Structured Observation Management and Activity-Based Intelligence. https://www.nga.mil/assets/files/170901-038_GEOINT_Basic_Doctrine_Pub_1.pdf
U.S. Army, “Ivy Intelligence (IVI) Large-Scale Combat Operations Targeting,” Military Intelligence Professional Bulletin, 2025. https://mipb.ikn.army.mil/issues/jan-jun-2025/ivy-intelligence
National Geospatial-Intelligence Agency, GEOINT Artificial Intelligence, 2025 capabilities overview. https://www.nga.mil/assets/files/NGA_Capabilities_2025.pdf
Office of the Director of National Intelligence, Intelligence Community Directive 203, Analytic Standards. https://www.odni.gov/files/documents/ICD/ICD-203.pdf
[Isaac R. Porche III et al., Data Flood: Helping the Navy Address the Rising Tide of Sensor Information, RAND Corporation, 2014. https://www.rand.org/content/dam/rand/pubs/research_reports/RR300/RR315/RAND_RR315.pdf
[Kristin Warren et al., Improving Sense-Making with Artificial Intelligence, RAND Corporation, 2025. https://www.rand.org/content/dam/rand/pubs/research_reports/RRA3100/RRA3152-1/RAND_RRA3152-1.pdf
RAND Corporation, Exploring Artificial Intelligence Use to Mitigate Potential Human Bias Within U.S. Army Intelligence Preparation of the Battlefield Processes. https://www.rand.org/pubs/research_reports/RRA2763-1.html
U.S. Army Futures Command, Army Futures Command Concept for Intelligence 2028. https://api.army.mil/e2/c/downloads/2021/01/05/26b729a6/20200918-afc-pam-71-20-3-intelligence-concept-final.pdf
Stockholm International Peace Research Institute, Bias in Military Artificial Intelligence, December 2024. https://www.sipri.org/sites/default/files/2024-12/background_paper_bias_in_military_ai_0.pdf
RAND Corporation, Technology Innovation and the Future of Air Force Intelligence. https://www.rand.org/content/dam/rand/pubs/research_reports/RRA300/RRA341-2/RAND_RRA341-2.pdf
Additional context: United States Geospatial Intelligence Foundation (USGIF), “Mission Focus: Accelerating Sensor Architectures for the Age of AI,” Oct. 20, 2026. https://usgif.org/event/mission-focus-accelerating-sensor-architectures-for-the-age-of-ai/
Author's prior commentary: Adam Simmons, “USGIF Mission Focus: It's About Automation, Not AI,” Geospatial Frontiers, July 30, 2026. https://projectgeospatial.org/geospatial-frontiers/usgif-mission-focus-its-about-automation-not-ai
