Mission Focus: AI First Approaches

At USGIF’s July Mission Focus event, government and industry leaders discussed how AI is shifting from an add-on tool to the foundation of daily intelligence work.

Artificial intelligence has often been treated as an add-on: a new tool bolted onto existing intelligence workflows. That era is giving way to “AI First Approaches,” the subject of USGIF’s July Mission Focus event on July 22.

The standing-room only event was hosted by event sponsor AWS at their offices in Arlington, Virginia, where views of the monuments, the Potomac, and activity at Reagan International provided the backdrop as Government, industry, and academic leaders explored what it will take to design GEOINT and all-source intelligence workflows with AI in mind from the start, rather than retrofitting AI into existing workflows.

USGIF CEO Ronda Schrenk delivered welcome remarks before joining Dr. Chris Parrett, Director of NGA’s Rapid Capabilities Office (RCO), for a fireside chat.

The AI-First Workflow Vision

Dr. Parrett framed the shift with an analogy to e-commerce: a customer places an order, and Amazon’s backend automatically handles inventory, routing, payment, and delivery without the customer ever seeing the complexity behind it. Intelligence operations, he argued, should work the same way. Rather than analysts manually moving between a dozen or more separate systems, searching databases, downloading imagery, transferring files, analyzing in one application, exporting to another, and building briefings by hand, AI should orchestrate the workflow behind the scenes.

Dr. Parrett pointed to a familiar example: imagery analysts spend a large share of their time building PowerPoint slides for dissemination rather than doing analysis. In an AI-first architecture, an analyst would submit a single intelligence request. AI would identify and retrieve the relevant data, condition it, assist with analysis and validation, and generate reporting across multiple formats automatically. The analyst remains responsible for judgment, while AI manages the repetitive technical work, moving the enterprise toward a “Unified Workflow” that consolidates multi-app, multi-task manual transfer process, allowing the technology to seamlessly recede into the background.

Traditional intelligence production scales roughly one analyst to one task. An AI-first model, panelists said, could allow a single analyst to responsibly oversee dozens, hundreds, or even a thousand-plus concurrent processes. That does not mean replacing analysts. It means shifting their role from production to supervision, giving them more time for interpretation and judgment instead of data preparation.

 

Rebuilding the Organization, Not Just the Tech Stack

Dr. Parrett said realizing that vision requires structural and cultural change in equal measure. On the structural side, he called for standardized enterprise data architecture, modular and open systems that can plug in new AI models rapidly, and multidisciplinary mission teams in place of isolated functional ones. Software, in his view, is never finished. It evolves continuously alongside the mission.

The harder obstacle, Dr. Parrett argued, is culture rather than technology. Government organizations have traditionally relied on lengthy requirements development, extensive documentation, and low-risk acquisition. AI development moves far faster, and closing that gap means becoming comfortable with rapid experimentation, incremental improvement, and manageable technical risk: fielding useful capability quickly and refining it in partnership with operational users, rather than waiting for a perfect system before deployment. Mission accomplishment, he said, needs to matter more than ownership of any individual program.

 

A Different Kind of Industry Pitch

As Director of NGA’s Rapid Capabilities Office, Dr. Parrett wants fewer technology demonstrations from industry and more mission understanding. Rather than opening with “here is our AI platform,” he said the more effective pitch starts with the operational problem an analyst faces and works backward to how a given technology reduces that specific burden.

That mission-first posture extends to acquisition strategy, Dr. Parrett said. NGA increasingly uses small prototype efforts and limited operational testing over large enterprise buys, scaling up only after a capability proves itself, a precision-over-bulk approach intended to reduce acquisition risk while getting innovation to analysts faster. RCO is also working to compress contracting timelines and is proactively communicating capability gaps to industry rather than waiting for industry to guess, breaking missions into smaller, more addressable pieces along the way.

 

Two Misconceptions Worth Retiring

Dr. Parrett pushed back on two assumptions that continue to shape how the community talks about AI. The first, he said, is that the path forward requires more models. Most organizations already have access to many capable models. The harder and more valuable work is coordinating them through intelligent workflows and autonomous agents, rather than developing yet more isolated capabilities.

The second, he said, is that the path forward requires more data. Most agencies already hold enormous volumes of it. The real constraint is organizing, conditioning, and making that data AI-ready: quality, accessibility, and context matter more than raw volume, and running large-scale analysis over unstructured, unconditioned datasets is often simply inefficient.

 

Redesigning Legacy Workflows

USGIF Vice President of Programs Sue Kalweit then moderated a panel featuring:

• Brian Bataille, Chief, National Digital Exploitation and OSINT Center (NDOC), Directorate for Science and Technology, DIA
• Sean Batir, Global Head of Mission Innovation, AWS, and former first CTO of NGA’s Maven program
• Dr. Amanda Fetch, Co-Chair, USGIF Machine Learning and Artificial Intelligence (MLAI) Working Group
• Gina Hurtado, ASPEN Program Manager, NGA
• Al Peguero, Senior Program Manager, GEOINT Operations Segment Lead, ECS Federal

Two speakers offered concrete examples of what workflow redesign looks like in practice. Gina Hurtado, ASPEN Program Manager, NGA, described a detection-versus-adjudication bottleneck in computer vision: modern models are now highly effective at detecting objects in imagery, but organizations are generating far more detections than analysts can validate. Her proposed fix is a hierarchical model structure, in which one model detects, a second validates the first, and human analysts review only the higher-confidence outputs, mirroring how senior analysts already supervise junior ones today.

Brian Bataille, Chief, National Digital Exploitation and OSINT Center (NDOC), Directorate for Science and Technology, DIA, made a parallel case for search and discovery. Traditional intelligence search still relies heavily on keywords, which assumes analysts already know what they’re looking for. He argued that AI-first search should instead surface emerging patterns and previously overlooked connections across datasets, turning search into an exploratory reasoning process rather than simple retrieval.

 

What Stands in the Way

Speakers were candid about the barriers to getting there. Governance maturity remains uneven, with different agencies and IC components interpreting acceptable AI risk differently, though panelists noted a positive shift toward accepting greater technical risk to meet operational needs, with policy expected to catch up over time.

Computing infrastructure is a second constraint: demand for AI compute is growing quickly, and matching model size to task is becoming its own discipline. Panelists were clear that not every task requires the largest, most expensive model — second-tier models are often sufficient, and appropriate, for lower-stakes work like adjudication.

Data readiness drew the most attention as the most critical gap. A decade ago, the challenge was access to data. Today, most organizations have the data but not in a form AI systems can use, and substantial time goes into cleaning, standardizing, and labeling it before it’s usable. Panelists called for commercial datasets to arrive AI-ready from the start, treating data engineering as equal in importance to model development. Finally, several speakers pointed to infrastructure silos, with requirements and innovation teams operating separately from infrastructure teams, as a quieter but persistent drag on progress, and pointed to AWS’s practice of forcing regular cross-functional meetings between those groups as a model government organizations could adopt.

 

The Analyst of the Next Five Years

Skills panelists expect to grow in value include critical thinking, prompt engineering and problem formulation, human-AI collaboration, tradecraft, and judgment under uncertainty, in other words, a “trust but verify” posture in which analysts review AI outputs with the same rigor team leads apply to human work today. Panelists pushed back directly on the fear that AI erodes critical thinking, arguing the opposite: critical thinking becomes more valuable, not less, once outputs can no longer be taken on faith.

Conversely, panelists expect routine, repetitive work, including manual searches, formatting, filtering, and prioritizing raw detections, to be increasingly automated, freeing analysts to serve as supervisors of AI rather than manual processors of information.

 

What Academia Should Prioritize

Sue Kalweit asked Dr. Amanda Fetch to weigh in on gaps in academic research who shared that she felt academia over-indexes on model development while underinvesting in operationalization. That means understanding how these tools actually function in messy, real-world environments, and in the change-management and systems-thinking work needed to integrate people, data, and technology together, rather than treating each in isolation.

Panelists also flagged a gap in analytic standards. Traditional intelligence products communicate confidence levels, sourcing reliability, and uncertainty. AI tools today generally don’t. Much as intelligence community directives govern sourcing standards for human-produced analysis, panelists suggested a similar framework is needed for AI outputs, including confidence estimates, reasoning explanations, source pedigree, and validation history, so AI-generated products can be held to the same standards as human ones. Ultimately, panelists argued, technological superiority alone won’t determine advantage, since many competitors have access to similar tools. The differentiator is tradecraft: how effectively an organization applies AI within its workflows and decision-making.

 

 

The Economics of AI at Scale

A recurring thread in the Q&A was cost. Every interaction with a large language model consumes computational resources measured in tokens, and panelist Sean Batir, Global Head of Mission Innovation, AWS, and former first CTO of NGA’s Maven program, noted that commercial organizations are far more attentive to “tokenomics” than government currently is. Not every mission requires the largest frontier model. Smaller, cheaper models are often sufficient for routine validation tasks, and organizations need to weigh timeline against cost explicitly: a 30-second answer costs more than one that can wait two hours. Panelists pointed to token-tracking features in tools like AWS Bedrock as a starting point for building that discipline into government AI use.

One Thing That Would Prove AI-First Is Working

Asked to name a single marker of success a year from now, panelists offered a range of answers that, together, sketched out what real progress looks like. Gina Hurtado pointed to a system recently deployed at NGA where analysts produce a report once, and it automatically populates every downstream format and dissemination channel. Batir described a workforce where agentic manages daily priorities; plugged straight into email, chat, and knowledge management. Fetch’s marker was trust: AI summaries that reliably surface the right information without missing nuanced counter-indicators, with humans still in the loop until that bar is consistently met. Peguero framed success as adoption grounded in understanding, where the workforce trusts AI under pressure because it understands the tool’s strengths and limits, not just its interface. And Bataille offered a historical parallel: early typewriters required trained operators for every task. AI reaches full maturity when it’s simply how work gets done, without anyone needing to talk about “adoption” at all.

Looking Ahead

Panelists closed by noting that an AI-first transition requires leadership commitment, trusted data and modern infrastructure, workforce readiness, responsible governance, and a willingness to rethink processes that have been in place for decades. Government, industry, and academia each bring a distinct piece of that puzzle, and forums like Mission Focus exist to bring those perspectives together, accelerating innovation and getting solutions to the mission faster.

USGIF thanks its speakers for sharing their insights with the community, and thanks AWS for supporting and hosting July’s Mission Focus event.

Join Us at an Upcoming USGIF Event

Next Generation GEOINT: Forging Leaders
Tuesday, September 15, 2026, 5:00 pm – 7:00 pm ET
Trajectory Event Center, Herndon, VA
A fireside chat and Q&A on leadership, integrity, and developing the next generation of GEOINT talent. Register Now

GEOGala 2026
Thursday, December 3, 2026, 5:00 pm – 10:00 pm ET
The Ritz-Carlton, McLean, VA
The GEOINT community’s premier black-tie celebration, honoring The Honorable Stacey A. Dixon, 2026 recipient of the Lundahl-Finnie Lifetime Achievement Award. Learn More

GEOINT Symposium 2027
Monday, May 17 – Wednesday, May 19, 2027
Gaylord Palms Resort & Convention Center, Kissimmee, FL
Join 4,000+ attendees from across the GEOINT community to connect mission priorities, emerging capabilities, and the future of geospatial intelligence. Learn More

Want to be the first to know about programs like Mission Focus? Sign up for USGIF news to stay up to date on upcoming events across the GEOINT community.