The Algorithmic Tradecraft: Synthesizing Artificial Intelligence and Automation in Modern Geospatial Intelligence Workflows

GeoAI & Data Science
Written By Adam Simmons

The national security apparatus is undergoing a profound structural, cultural, and epistemological paradigm shift, driven by the exponential proliferation of sensor data and the rapid maturation of computational analytics. The intersection of these twin forces necessitates a fundamental restructuring of intelligence tradecraft. This transformation was the central focal point of two critical events hosted by the United States Geospatial Intelligence Foundation (USGIF) in July 2026. The first gathering, titled Mission Focus: AI First Approaches, examined the operationalization of machine learning from isolated, bespoke experiments to foundational, ubiquitous workflow components. This event featured deep perspectives from Sean Batir, Global Head of Mission Innovation at Amazon Web Services (AWS); Brian Bataille, Chief of the National Digital Exploitation & OSINT Center (NDOC); Dr. Amanda Fetch, USGIF Machine Learning and Artificial Intelligence (MLAI) Working Group Co-Chair; Dr. Chris Parrett, Director of the Rapid Capabilities Office at the National Geospatial-Intelligence Agency (NGA); and Gina Hurtado, NGA ASPEN Program Manager.

Days later, the GEOINTeraction Tuesday: It's About Automation, Not AI event featured Dr. Steve Hall, NGA Deputy Director of Analysis, who presented a critical counter-thesis to the artificial intelligence hype cycle. His overarching argument posited that the future of geospatial intelligence (GEOINT) relies less on the isolated brilliance of artificial intelligence models, and entirely on the end-to-end automation of the analytic pipeline. Synthesizing the discourse from these dual gatherings reveals a community in deep transition. The prevailing thought is no longer fixated on proving that computer vision can detect a surface-to-air missile battery or naval vessel in satellite imagery; that capability is now a baseline assumption. Instead, the strategic imperative revolves around overcoming the extreme friction of integrating these algorithmic detections into legacy workflows, managing the sheer velocity of machine-generated observations, and ensuring that computational outputs meet the rigorous legal and ethical thresholds required for positive identification in modern warfare. The contemporary challenge demands a holistic reevaluation of how intelligence professionals partner with commercial industry, rethink legacy procurement strategies, and deploy automated solutions at true mission speed.

The Data Deluge and the Automation Imperative

The core architectural crisis facing the intelligence community is a severe and widening mismatch between sensor collection capacity and human cognitive bandwidth. As imagery collection aggressively expands across both classified national technical means and commercial satellite constellations, analysts face an unprecedented and geometrically expanding volume of data. Projections shared by NGA leadership indicate a near tripling of GEOINT data arriving at analytical workstations over the next half-decade, driven by new satellite deployments reaching full operational capability. In this specific environment, relying on human analysts to manually review raw imagery across multiple screens constitutes a catastrophic operational bottleneck.

Dr. Steve Hall eloquently articulated this challenge during his keynote address by utilizing a powerful metaphor: the dog has finally caught the car. For decades, the intelligence community clamored for more sensors, more coverage, and more data. Now that the enterprise possesses an abundance of highly structured observation data, capturing everything known in finite database rows, the sheer volume has overwhelmed the human capacity to exploit it. Hall noted that the community has been building structured observation management (SOM) frameworks for twenty-five years, resulting in tens of thousands of human-verified observations generated every single week.

However, the introduction of computer vision has radically altered this math. Alongside those human-verified data points, algorithms are now generating millions of computer-generated observations weekly that remain entirely unvalidated. The inability of human analysts to manually verify each machine-generated observation creates a massive validation debt. If a highly accurate computer vision model merely generates tens of thousands of uncontextualized bounding boxes per week, it has not solved the analyst's operational problem; it has merely transformed a data-sorting problem into an alert-fatigue problem. True operational advantage requires intelligent, overarching automation.

Automation, in this precise context, refers to the systematic orchestration of data flows and the elimination of administrative friction. Hall pointedly observed that analysts currently spend an exorbitant amount of their cognitive bandwidth acting as manual data routers, swiveling between a dozen disparate applications, copying and pasting coordinates, and ultimately dedicating the largest portion of their day to formatting Microsoft PowerPoint slides for senior briefings. This manual overhead destroys mission speed. The objective is to automate the entirety of the intelligence cycle up to the exact moment where human judgment, critical thinking, and ethical consideration are legally and strategically required. The constant, comforting narrative of keeping a 'human-in-the-loop' is rapidly approaching an expiration date. Within the next five years, the intelligence community is facing a very real operational reality where human operators will be completely removed from the bottom layers of the intelligence cycle. The necessary shift to 'human-on-the-loop' means human intervention will be consolidated entirely at the highest, decision-critical nodes, leaving machines to handle the foundational exploitation and formatting.

Architecting the Automated Workflow: MAVEN and ASPEN

The transition from theoretical artificial intelligence to deployed, enterprise-scale automation is most visibly represented by the evolution of Project Maven and the implementation of the Analytic Services Production Environment for the National System for GEOINT (ASPEN). These two programs, operating in tight tandem, represent the structural foundation of the NGA's modernization strategy, internally championed by NGA Director Vice Admiral Frank Whitworth as the "Year of NGAI".

Project Maven, originally established in 2017 as the Department of Defense's Algorithmic Warfare Cross-Functional Team, officially transferred its GEOINT operational control to the NGA in 2023. Maven's core function is object detection at the scale of global conflict. It utilizes advanced convolutional neural networks, heavily descended from the YOLO (You Only Look Once) architecture, to process overhead imagery. The mathematical objective of these networks involves minimizing a multipart loss function that severely penalizes localization errors, objectness confidence deviations, and classification inaccuracies. For overhead satellite imagery, where critical targets may only occupy tens of pixels, the localization loss function is highly sensitive to minor spatial deviations, ensuring high-fidelity bounding box predictions over vast geographic expanses.

While Maven provides the raw detection capability, ASPEN, managed by Gina Hurtado, provides the necessary orchestration, data pipelining, and analytical environment. ASPEN is a multi-year program of record designed to integrate these AI capabilities directly into the analytical workforce, addressing the looming tripling of GEOINT data. If Maven acts as the visual cortex of the intelligence apparatus, ASPEN serves as the central nervous system. ASPEN leverages the millions of detections generated by Maven to conduct deep, time-series pattern of life analysis. By establishing strict mathematical baselines of known behaviors at known geographic locations, ASPEN can autonomously flag unknown behaviors or anomalous deviations.

This architecture fundamentally alters the temporality of geospatial intelligence. Legacy workflows were inherently reactive; analysts would review imagery of an event only after it had occurred or after receiving a tip-off from another intelligence discipline. The combination of Maven's computer vision and ASPEN's automated pattern analysis pushes the intelligence cycle "left of launch". By continuously monitoring adversary tactics, techniques, and procedures via automated statistical analysis, the system can anticipate events, providing advanced predictive warning to warfighters and policymakers.

System Architecture Primary Function Core Technology Operational Impact
Project MAVEN Scalable Object Detection Convolutional Neural Networks, Spatial-Spectral Matched Filters Automated feature extraction, positive identification of military assets from multimodal sensor inputs.
ASPEN Analytic Services Production Automated Workflow Orchestration, Time-Series Analysis Pattern of life baselining, left-of-launch predictive warning, structured observation management.
NDOC OSINT Multi-Source Fusion Natural Language Processing, Graph Neural Networks Cross-referencing visual detections with unclassified signals, establishing holistic intelligence pictures.

Deconstructing the Three-Legged Stool: AI, Automation, and Tradecraft

During the USGIF GEOINTeraction Tuesday event, Dr. Steve Hall introduced a critical conceptual framework, describing the modernization of the intelligence enterprise as a "three-legged stool" consisting of Artificial Intelligence, Automation, and Tradecraft. His explicit warning to the audience of government and industry professionals was that focusing solely on AI development while neglecting the other two legs will cause the entire structure to collapse.

The first leg, Artificial Intelligence, provides the high-speed pattern recognition and computer vision necessary to process the data deluge. However, Hall noted a deep frustration with the AI community's fixation on technical metrics rather than operational outcomes. He challenged the audience to stop citing "F1 scores" or isolated algorithmic accuracy benchmarks, as these metrics rarely translate linearly into decision advantage for a military commander. If an algorithm is mathematically perfect but takes days to deploy or requires an analyst to navigate clunky software interfaces, its operational value is zero.

The second leg, Automation, represents the deterministic logic, application programming interfaces (APIs), and data pipelines that move information from the sensor to the algorithm and ultimately to the human interface. Hall emphasized that he does not need artificial intelligence to pull data out of a database; he needs robust, high-functioning APIs and databases that are inherently interoperable rather than siloed. Automation handles the tedious, repetitive formatting, routing, and administrative tasks that currently plague the workforce. By automating the classical mathematics and predictive analytics, the system can autonomously sift through the millions of unvalidated observations to pull the true signal from the noise, presenting the analyst with only the most statistically significant anomalies.

The third and most critical leg is Tradecraft. Tradecraft encompasses the human reasoning, contextual understanding, and geopolitical knowledge that transforms a digital bounding box into actionable intelligence. The intelligence community must aggressively update its tradecraft to match the speed of its tools. Analysts are traditionally trained to operate with absolute certainty, building intelligence products from the ground up by manually verifying every pixel. In an automated, AI-first environment, the analyst must evolve from a primary detector of objects into a sophisticated auditor of algorithmic confidence. This requires a profound cultural shift. Analysts must learn to manage probabilistic outcomes, understanding when to trust the machine's baseline analysis and when to inject human skepticism.

Consequently, a deep understanding of tradecraft is no longer required for the sake of executing the raw analysis itself; rather, it is essential for possessing the foundational context needed to make critical, highly informed decisions based on machine outputs. This structural shift will inevitably place an immense, concentrated burden of pressure on the select few human operators positioned at these decision-critical nodes, fundamentally changing the psychological profile of the modern analyst.

Multimodal Fusion, Semantic Search, and the Agentic Horizon

The modern geopolitical threat landscape is characterized by deliberate adversary deception, camouflage, and multi-domain operations, making single-source intelligence inherently brittle. The deep integration of Open Source Intelligence (OSINT) with highly classified GEOINT is a critical requirement for maintaining decision advantage. Brian Bataille's representation of the National Digital Exploitation & OSINT Center (NDOC) at the USGIF event underscores the absolute necessity of fusing disparate data streams to corroborate findings, provide context to visual observations, and mitigate algorithmic hallucinations.

Sean Batir, Global Head of Mission Innovation at AWS and the former Chief Technology Officer of the NGA Maven program, highlighted the rapid, industry-driven transition toward multimodal foundation models. Traditional computer vision models process pixel arrays, while traditional large language models (LLMs) process text. The operational frontier currently being explored involves models capable of simultaneously reasoning over satellite imagery, translated text, and signals intelligence. In a multimodal Retrieval-Augmented Generation (MRAG) framework, an analyst can query a highly classified system using natural language, and the system can autonomously retrieve relevant synthetic-aperture radar data, cross-reference it against intercepted logistics documents, and generate a synthesized, fully cited intelligence product.

This capability is actively moving from theoretical whitepapers into classified cloud environments. Recent integrations of advanced models, such as Anthropic's Claude running on AWS within Defense Information Systems Agency (DISA) Impact Level 6 (IL6) accredited boundaries, demonstrate the commercial sector's vital role in delivering secure, high-performance computing to the intelligence community. The operational pattern relies entirely on strict data ontologies and bounded knowledge graphs, ensuring that the generative models are mathematically grounded in verified, authoritative intelligence rather than probabilistic guesswork that could lead to dangerous hallucinations.

A profound third-order insight emerges from this shift toward multimodal, agentic workflows. As systems like ASPEN integrate agentic AI, the intelligence architecture transitions from being a passive repository of information to an active participant in collection orchestration. An agentic system, upon autonomously detecting an anomaly in a specific geographic grid via low-resolution commercial radar, could immediately cross-reference local OSINT feeds for corroborating social media chatter regarding military movements. If a predefined confidence threshold is met, the software agent could automatically task a high-resolution national technical asset to capture optical imagery of the target on its next orbital pass. This concept of "informed collection orchestration" operates twenty-four hours a day, presenting optimized collection options based on constellation constraints and standing intelligence needs, effectively closing the sensor-to-shooter loop at machine speed.

Governance, Accreditation, and Ethical Tradecraft: AGAIM and GREAT

The most profound challenge discussed across the USGIF events is not purely technological or bureaucratic, but deeply epistemological. How does an intelligence professional trust a machine with life-or-death decisions? The aggressive transition to automation surfaces a deep, natural cultural resistance within the analytical workforce, rooted in the inherent "black box" nature of deep neural networks. Analysts are rigorously trained to trace their conclusions back to independently verifiable evidence. When a complex neural network outputs a highly confident detection without human-readable reasoning or a transparent causal chain, it clashes directly with traditional tradecraft standards.

This friction is incredibly acute in targeting scenarios. The Law of Armed Conflict requires the strict differentiation of combatants from non-combatants, and military assets from civilian infrastructure. The concept of Positive Identification (PID) is the unyielding legal and operational threshold that must be met before lethal action is authorized. As NGA Director Vice Adm. Frank Whitworth explicitly noted, ensuring PID is the ultimate prerogative of the Commander-in-Chief, the Secretary of Defense, and their delegated engagement authorities; it is a profound moral responsibility that cannot be blindly outsourced to an algorithm.

To bridge the dangerous gap between machine speed and human accountability, the NGA launched the Accreditation of GEOINT AI Models (AGAIM) initiative. AGAIM is a standardized, highly rigorous evaluation framework designed to test the methodology, robustness, and reliability of computer vision models deployed for national security. Accreditation is not merely a technical benchmark or a software testing phase; it is a vital legal and operational mechanism for risk management. AGAIM ensures that AI models are rigorously evaluated against adversarial perturbations, severe distribution shifts, and varying spatial-spectral conditions before they are allowed to inject data into the automated ASPEN pipelines. It is the technical guarantee that the machine's baseline performance is worthy of the analyst's trust.

A direct, human-focused corollary to AGAIM is the GEOINT Responsible AI Training (GREAT) program. While AGAIM audits the machine, GREAT audits the human operator. The GREAT program is a mandatory certification designed to teach coders, engineers, and analysts how to dynamically assess risk, algorithmic bias, and ethical responsibility across the entire AI lifecycle. By mandating that developers and users sign formal pledges to mitigate collateral damage and adhere to ethical guidelines, the NGA is deliberately attempting to hardcode human morality into the deployment of automated systems. The interplay between AGAIM and GREAT reveals a deep recognition that technology alone cannot solve national security challenges; it must be bound by rigorous ethical guardrails.

Bureaucracy, Acquisition Friction, and "Tokenomics"

The technological capability to execute AI-first strategies is advancing significantly faster than the federal government's ability to procure it. Dr. Chris Parrett, Director of the NGA's Rapid Capabilities Office (RCO), emphasized the critical, existential need to rethink legacy acquisition processes during the Mission Focus panel. The traditional Federal Acquisition Regulation (FAR) framework, originally designed for procuring massive physical hardware like aircraft carriers or fighter jets over decade-long timelines, is fundamentally incompatible with generative AI and computer vision software that undergoes iterative, fundamental updates on a weekly basis.

The structural latency of government procurement poses a direct, immediate threat to national security. If a near-peer adversary updates its camouflage techniques, alters its signal emissions, or changes its dispersal tactics, the computer vision models and electronic warfare algorithms monitoring them must be retrained, tested, and redeployed immediately. Waiting eighteen months for a contract modification effectively blinds the intelligence apparatus. To counter this inertia, Parrett's RCO is aggressively leveraging Other Transaction Authorities (OTAs) and commercial data indefinite delivery, indefinite quantity (IDIQ) vehicles to bypass bureaucratic roadblocks and inject capabilities directly into the operational environment. The underlying philosophy, championed by NGA leadership, is that the agency must accept a substantially higher degree of acquisition risk in order to effectively lower its operational risk on the battlefield.

Industry representatives, including Al Peguero of ECS Federal and academic liaisons like Dr. Amanda Fetch of USGIF, highlight the extreme importance of deeply integrated, highly transparent partnerships. The commercial sector, heavily driven by venture capital and hyperscaler cloud providers, vastly outpaces the government in foundational AI research. Consequently, the intelligence community's role is rapidly shifting from building bespoke, proprietary algorithms in-house to acting as a sophisticated multi-source integrator. The NGA is currently investing hundreds of millions of dollars in commercial analytics, such as the massive $708 million Sequoia data labeling effort, while deliberately maintaining strict government ownership of the underlying data ontologies, security architectures, and integration pipelines.

However, during audience interactions at these events, a critical financial constraint emerged: the concept of "Tokenomics." Operating massive, 300-billion parameter foundation models requires an exorbitant amount of computational power, which translates directly into massive recurring cloud costs. While commercial entities can justify these costs by tying them directly to revenue-generating products, government agencies operate on fixed, congressionally appropriated budgets. The audience raised deep concerns about the financial sustainability of running natural language processing and computer vision models over petabytes of continuously arriving satellite imagery. The consensus indicates that the intelligence community cannot simply apply a massive foundation model to every single query; it must develop a tiered architecture where cheap, highly optimized, specialized algorithms handle the bulk of the automated processing, reserving expensive, compute-heavy generative models only for the most complex analytical reasoning tasks.

Unanswered Questions and the Operational Horizon

Despite the aggressive, optimistic push toward AI-first automated workflows presented by government leaders, several critical questions remain unanswered by the representatives mapping out this space, highlighting the gap between strategic vision and operational reality.

  1. The challenge of multi-spatial and multi-temporal resolution fusion remains structurally complex. An audience member specifically challenged the panels on how to fuse moderate-resolution persistent data (like Landsat), high-resolution commercial satellite tasking, and hyper-local drone full-motion video into a single analytical workflow. While the theoretical answer involves sensor-agnostic platforms, the practical reality of synchronizing disparate pixel sizes, temporal revisit rates, and spectral bands into a unified algorithm without overwhelming the analyst with conflicting data points has not been fully resolved at scale.

  2. The integration of legacy formats remains an unresolved, severe friction point. While advanced machine-to-machine APIs represent the future, massive amounts of highly classified intelligence are currently communicated and stored via PDFs, PowerPoint presentations, and unstructured text documents. Can agentic AI workflows seamlessly parse decades of poorly structured legacy reports without introducing massive hallucination errors into the knowledge graph? The technical challenge of translating highly specialized, domain-specific military jargon into standardized vector embeddings for retrieval-augmented generation systems is immense, and the community has yet to outline a concrete timeline for this data conditioning phase.

  3. The cybersecurity architecture for true multi-classification AI remains brittle. For an agentic workflow to operate as envisioned, it must autonomously pull data from unclassified OSINT feeds, synthesize it with SECRET-level telemetry, and cross-reference it against TOP SECRET//SCI national technical means. Implementing high-assurance data diodes and Cross-Domain Solutions (CDS) that allow data to flow upward without exposing the highly classified networks to adversarial steganography, data poisoning, or malware triggers is a profound mathematical and engineering bottleneck. Until these cross-domain solutions are automated at machine speed, true multi-INT AI fusion will remain siloed by security classifications.

  4. There is the unresolved psychological tension regarding "automation complacency." As pipelines like ASPEN become highly accurate, human analysts faced with intense time constraints may naturally begin to defer to the machine's judgment, eroding their own analytical rigor and tradecraft over time. How does the intelligence community maintain the critical thinking skills of its workforce when the machine is right the vast majority of the time, to ensure the human catches the catastrophic, highly consequential error? The community has not yet answered how to design the human-computer interface to deliberately introduce productive friction, requiring human articulation of reasoning to prevent cognitive atrophy.

  5. The community has not established an honest comparative baseline for acceptable margins of error. For decades, human-sourced intelligence has relied on rigorous Quality Control (QC) reviews, implicitly accepting that human reporting is inherently fallible. As models are evaluated through accreditation frameworks like AGAIM, the government must address a hard question: has it compared the QC margin of error of machines directly against the historical QC error rates of humans? Holding algorithms to a standard of absolute perfection while forgiving human error creates a false paradigm. The intelligence apparatus must define what constitutes an acceptable, operational margin of error for automated systems rather than comparing them to an imaginary standard of flawless human analysis.

Finally, there is the uncomfortable reality of workforce displacement. Agency leadership frequently emphasizes that AI will 'augment' rather than replace humans, but this dances around a fundamental economic and operational truth: if end-to-end automation truly scales, specific analysis roles will be eliminated. AI will not just augment existing workflows if analysis becomes entirely automated and machine-driven intelligence integrates deeply into the ecosystem. The community must stop avoiding the reality of job replacement and begin answering exactly how many analysts will remain, what highly specialized roles those survivors will actually occupy, and how the agency plans to manage the structural downsizing of its legacy workforce.

Strategic Conclusions

Synthesizing the discourse from Mission Focus: AI First Approaches and GEOINTeraction Tuesday reveals that AI and automation are not competing operational paradigms, but interdependent phases of a single tradecraft shift. The "AI-First" movement addresses the fundamental problem of perception, converting massive, unstructured multi-INT sensor streams into structured, machine-readable observations. However, as the "Automation-First" counter-thesis correctly identifies, model proliferation in isolation creates an operational paradox. An algorithm capable of detecting thousands of targets in seconds becomes an analytic bottleneck if those detections stall in legacy networks, generate unmanaged alert noise, or require manual validation at every hop. AI generates the perceptual signal, but end-to-end pipeline automation provides the velocity required to make that signal operational.

Resolving this tension requires the geospatial intelligence community to redefine its metrics of success. True capability in the algorithmic era cannot be measured by model accuracy in a vacuum or by the raw volume of automated data pipelines. Instead, advantage belongs to the enterprise that seamlessly integrates both into a unified architecture, using automated pipelines to ingest, normalize, and route sensor data, while embedding model intelligence directly within those pipelines to enrich data dynamically. When model development and workflow engineering are treated as unified priorities, the technical friction between raw data collection and actionable intelligence delivery begins to dissolve.

Ultimately, this structural synthesis redefines the human element within modern GEOINT tradecraft. The objective of automating the analytic pipeline and operationalizing artificial intelligence is not to displace human judgment, but to elevate it. By entrusting machines with the velocity of routine observation, data transport, and initial triage, the enterprise frees analysts from the fatigue of manual exploitation. This shift enables practitioners to transition from reactive data processors to strategic evaluators, focusing their expertise where computational systems remain inherently limited: contextual synthesis, adversary intent assessment, and decision support under uncertainty.

Yet, as the defense intelligence enterprise struggles to strike this delicate equilibrium between perceptual AI models and workflow automation, a larger structural challenge looms. The velocity of global commercial innovation—from edge-deployed foundation models to autonomous sensor orchestration, is moving at an exponential pace, while bureaucratic acquisition, policy compliance, and legacy system integration continue to move at an incremental one. The ultimate risk to national security is not whether the government can conceptually reconcile AI with automation, but whether it can modernize its operational infrastructure fast enough to absorb innovation before the tradecraft itself becomes obsolete.

Adam Simmons

Geospatial Industry Consultant | Founder, Project Geospatial

Adam Simmons is a geospatial technology liaison and strategic advisor with over 20 years of experience across the defense and commercial sectors. A veteran of the U.S. Air Force, he specialized in imagery analysis and order of battle before transitioning to executive leadership as the CEO of Midgard Raven, LLC and the founder of Project Geospatial, a 501(c)(3) dedicated to highlighting innovation within the geospatial ecosystem. Adam bridges the gap between technical development and market storytelling, leveraging his extensive background as a journalist and industry consultant to help companies navigate complex technology landscapes.

https://www.linkedin.com/in/adamsimmonsgeo
Next

USGIF Mission Focus: It’s About Automation, Not AI