AI/ML for DoD: What Actually Works When Deploying AI in Defense Environments

As the Department of Defense accelerates its adoption of artificial intelligence and machine learning, defense contractors face unique challenges that commercial AI companies rarely encounter: classified data handling, air-gapped networks, stringent accreditation requirements, and the imperative that AI-driven decisions in military contexts must be explainable, auditable, and trustworthy.

At Zapata Technology, we’ve spent over 17 years deploying AI/ML solutions in some of the most demanding defense environments — from tactical edge deployments with LIGHTNER to enterprise intelligence platforms like CASCADE. Here’s what we’ve learned about what actually works when bringing AI to the warfighter.

The DoD AI Challenge Is Different

Commercial AI operates in environments with abundant data, cloud connectivity, and fast iteration cycles. Defense AI operates under fundamentally different constraints:

• Classified data environments — Models must be trained and deployed on air-gapped networks with strict data handling requirements. You can’t just spin up a GPU cluster on AWS.
• Explainability requirements — When an AI system informs a targeting decision or intelligence assessment, operators need to understand why the model reached that conclusion. Black-box models are unacceptable.
• Edge deployment — Tactical AI must run on constrained hardware in bandwidth-limited environments, often disconnected from enterprise infrastructure.
• Accreditation burden — Every AI system deployed on a DoD network must go through the Risk Management Framework (RMF) process, adding months to deployment timelines.
• Human-in-the-loop — DoD policy requires meaningful human oversight of AI-informed decisions, especially in lethal engagements. AI augments the warfighter; it doesn’t replace judgment.

Practical Approaches That Work

1. Build for the Classification Level From Day One

The most common mistake we see is developing AI models in unclassified environments and then trying to migrate them to classified networks. The data pipeline, training infrastructure, and deployment architecture need to be designed for the target classification level from the start. Our experience on programs like DCGS-A has shown that architectures designed for classified environments from inception deploy faster and perform better than retrofitted solutions.

2. Prioritize Explainability Over Raw Accuracy

A model that achieves 98% accuracy but can’t explain its reasoning is less useful to an intelligence analyst than one at 93% that shows its work. Our CASCADE platform was designed around this principle — every output includes confidence scores and the data sources that contributed to the assessment, enabling analysts to validate AI-generated insights against their own expertise.

3. Design for Degraded Operations

Tactical AI systems will lose connectivity. They will operate on hardware that overheats. They will receive incomplete data. Systems must degrade gracefully — providing reduced capability rather than failing completely. LIGHTNER was designed to run on edge hardware precisely for this reason, delivering object recognition capability even in bandwidth-constrained tactical environments.

4. Invest in Data Engineering Before Model Development

The most sophisticated ML model is useless if it can’t ingest the data it needs. In defense environments, data comes from dozens of disparate sources with different formats, classification markings, and update frequencies. We built ZIngest specifically to solve this problem — automating data ingestion, CAPCO classification extraction, and format normalization so data scientists can focus on model development rather than data wrangling.

The Road Ahead: DoD AI in 2025 and Beyond

The DoD’s AI adoption is accelerating, driven by the Chief Digital and AI Office (CDAO) and programs like the Replicator initiative. We expect to see increased demand for:

• Autonomous sensor processing at the tactical edge
• Large language models adapted for intelligence analysis and report generation
• Predictive maintenance for weapons systems and logistics
• Computer vision for ISR and targeting support
• Multi-domain data fusion across joint force operations

Small businesses like Zapata Technology have a critical role to play in this ecosystem. Our agility, cleared workforce, and mission focus allow us to develop and iterate on AI solutions faster than large primes, while our contract vehicles (OASIS+, STARS III, SeaPort-NxG) make procurement straightforward for contracting officers.