National Security AI Requires More Than Better Models

National Security AI Requires More Than Better Models

Mission advantage depends on how effectively AI is integrated with data, systems, workflows, and operators.

By Joe Kelleher • August 11, 2026

Years before Russia’s full-scale invasion of Ukraine, I had the opportunity to spend significant time in country with some of the people who would later help defend it. When Russia began gathering forces near Ukraine’s borders in early 2022, I watched with particular interest as U.S. intelligence analysts faced an all-too-familiar challenge. The warning signs were there, but they were buried within hundreds of millions of publicly available data points, images, and videos. Analysts needed a way to sift through that volume fast enough to see the pattern taking shape. While all of this was unfolding, advances in AI were giving analysts new ways to work through the volume. Automated collection brought the information together, while machine learning and computer vision surfaced the observations that mattered, including troop movements, equipment deployments, unit composition, and sustained staging activity near the border. No single indicator proved an invasion was imminent. Taken together, they revealed a pattern that distinguished preparations for war from another large-scale exercise, giving policymakers earlier warning and greater confidence. Generative AI has since expanded what these systems can do. The underlying lesson still holds: mission advantage depends on how well technology, data, tradecraft, and human judgment work together.

The U.S. Intelligence Community (IC) and Department of War (DoW) are no longer asking whether artificial intelligence and machine learning (AI/ML) belong in national security operations; that question has been answered. AI/ML is already being applied across multiple long-standing fields such as intelligence collection, geospatial analysis, cyber defense, open-source exploitation, and pattern recognition. AI/ML tools are also increasingly being applied to targeting support, decision intelligence, predictive modeling, and operational planning. The more important question now is whether the national security enterprise can move fast enough to turn AI from a set of promising tools into durable mission advantage.

The threat environment demands it. State actors are using tried-and-true tactics involving offensive cyber operations, information manipulation, and economic coercion to challenge U.S. interests globally. All the while, state-sponsored or sanctioned proxy forces demand a high level of U.S. attention during a period of rapid military modernization. Non-state actors continue to exploit increasingly available commercial technology, encrypted communications, unmanned systems, illicit finance, and ransomware to move faster than traditional intelligence workflows were designed to handle. The ODNI’s 2026 Annual Threat Assessment of the U.S. Intelligence Community warned that China, Russia, Iran, North Korea, and non-state ransomware groups continue to threaten U.S. government, private-sector, and critical-infrastructure networks, while AI is expected to accelerate both offensive cyber operations and defensive countermeasures.¹

That reality changes the role of AI/ML in national security. AI should not be treated as a laboratory capability, a dashboard feature, or a narrow automation tool. It should be treated as operational infrastructure for decision advantage. BigBear.ai’s direct support for principal IC programs such as MIDB and MARS speaks to our focus on this distinction.

Today, the IC and DoW are already applying AI/ML in important ways. In GEOINT, AI/ML and computer vision are being used to help produce trusted intelligence at speed and scale. NGA has publicly emphasized GEOINT AI as a way to apply machine learning and computer vision to geospatial missions, and it has also launched accreditation efforts to evaluate the methodology, robustness, development, and testing procedures behind GEOINT AI models.^2,3 In operational environments, AI/ML is also moving closer to the edge: NGA’s GAMBLER effort demonstrated an integrated AI capability designed to operate on small unmanned aerial systems in austere Army environments.⁴ At BigBear.ai, we’ve taken disparate and seemingly unrelated datasets and, using machine learning and predictive modeling, turned mountains of data into actionable insights for military planners and commanders.

These are the right vectors, but the next stage is not simply “more AI.” The next stage is better-integrated AI: systems that can ingest and curate data while operating across classification and mission boundaries and ultimately provide support for analyst tradecraft and decision-maker requirements. Additionally, the IC and DoW will need these systems to maintain data provenance, expose confidence and uncertainty in the decision process, adapt to operational feedback, and remain secure, auditable, and accredited throughout their lifecycle. The national security enterprise must move faster in these areas.

The DoW’s 2026 AI Strategy is clear: speed wins. The strategy calls for the defense enterprise to “weaponize learning speed,” measure deployment velocity and operational cycle time, reduce blockers around data sharing, ATOs, test and evaluation, certification, contracting, and cross-domain access, and sustain rapid model updates so warfighters are not operating on stale capabilities.⁵ This set of requirements is not a wish list of technology preferences; it’s an entirely new operating model for the enterprise.

This is where BigBear.ai is investing significant resources and attention.

The future of national security AI will not be won by model performance alone. It will be won by integration performance: how quickly a team can move from data to insight, from insight to decision, from decision to action, and from operational feedback into model and system improvement. This is also why the IC and DoW should broaden how they think about prime contractors for AI/ML programs. Traditional primes will remain important across major defense platforms, enterprise services, and large-scale integration efforts. But AI/ML programs are different. They require continuous experimentation, rapid iteration, data engineering depth, mission intimacy, cybersecurity discipline, and the willingness to challenge slow acquisition and delivery patterns.

The DoW’s AI Strategy identifies “robust competition by small teams, with transparent metrics for results” as a driver of commercial AI leadership. It also calls for small, accountable teams, continuous field experimentation, operator feedback within days rather than years, and updates that move faster than adversaries can adapt.⁶ That language should change how agencies evaluate prime readiness.

Our technology teams are built around the kind of mission-ready AI integration that programs now require: close alignment with analysts and operators, applied AI/ML, data-centric engineering, open and extensible architectures, secure delivery, operational sustainment, and measurable decision advantage.

As the IC and DoW confront a faster, more complex, more data-saturated threat landscape, the next generation of AI programs must move beyond pilots and point solutions. They must become trusted, scalable, operational capabilities. That will require partners who understand both the technology and the mission environment: partners who can deliver innovation without creating fragility, speed without sacrificing governance, and automation without diminishing human judgment.

The national security enterprise does not need AI for AI’s sake. It needs AI that helps analysts see what they would otherwise miss, helps operators act before windows of opportunity close, helps commanders understand risk and consequence, and helps policymakers make decisions with greater speed, confidence, and precision.

That is the standard. And BigBear.ai is ready to help lead as a prime mission partner.