AI has proven itself in intelligence analysis, autonomous systems, and knowledge management. But real-world command and control—where minutes or lives are at stake—exposes a critical gap: Most AI breaks the moment systems degrade or disconnect.
That reality defined this year's Air Force Association panel with leaders from Raft, Legion, Oracle, and Virtualitics. Colonel Ryan "Ape" Hayde, commander of the 505th Command and Control Wing, opened with the challenge every C2 leader faces: AI must be tested and trusted in realistic exercises long before it reaches a mission.
"We always try to practice like we fight. AI can't just happen in the fight. It can't be the thing we only unleash when it's time to go."
Watch the full panel discussion
The panel revealed a simple truth: AI's promise and its operational reality in C2 are still far apart. Most defense organizations deploy AI as isolated tools; think chatbots, computer vision models, and NLP pipelines. But command and control isn't a collection of point solutions. It's a system of systems. And when those systems can't coordinate, AI fails.
The breakthrough isn't better models. It's solving how AI works across every system.
Why AI Pilot Programs Fail
Organizations test whether models can generate good answers, rather than testing whether AI can actually complete work across their systems. That's because certain barriers make piloting AI across systems difficult, in both military and commercial operations:
- Siloed systems that don't communicate
- Data locked in legacy formats
- Security requirements that eliminate most commercial solutions
- The last-mile problem: AI works in demos but breaks under real complexity
These aren't technical challenges—they're architectural ones.
Peter Guerra, Group Vice President of AI & ML at Oracle Government, Defense & Intelligence, described what production actually looks like: Oracle processes 85% of its purchase orders autonomously, moving "from knowledge to action."
That's the goal: getting from pilot to production. And it isn't about better models—it's about solving how AI connects across systems.
Why Orchestration Is The Answer
Trey Coleman, Chief Product Officer at Raft, framed the current state of AI within C2 bluntly: "Most of how we're using AI today is autonomy, is if-then statements."
In other words, AI is being used to automate simple or isolated tasks. But complex operations, like those within C2, require something fundamentally different.
Consider what Col. Hayde described: You plan 100 aircraft to launch. Seventy actually launch. Fifty return. Meanwhile, three runways are destroyed, assets have moved, and aircraft need coordinated landing instructions across multiple bases—right now.
In this moment, humans shouldn't be manually:
- Checking runway statuses across systems
- Calculating fuel availability and routes for returning aircraft
- Tracking asset locations in real-time
- Coordinating landing instructions across bases
As Coleman put it, "What workload can I offload to a machine so I only have to think about the hard things?" In this case, the “hard thing” is the critical decision of how to safely recover the mission. The above scenario proves that the bottleneck isn't human analysis. It's friction between systems that don't communicate.
Ben Van Roo, CEO of Legion, explained where AI is headed: "Artificial intelligence is getting more proficient in orchestrating all these different systems together—giving you the best picture given the information available."
This isn't theoretical. Legion's work with U.S. Special Operations Command demonstrates what this looks like in production:
- 94% faster intelligence summary generation
- 24X faster surveillance log analysis
- 216 hours per week reclaimed from manual workflows
These are production results from AI that connects to existing systems, treating models as interchangeable components rather than dependencies.
Why Trust Requires Testing Under Pressure
The orchestration architecture is necessary but not sufficient. Col. Hayde posed the harder question: How do you ensure operators trust AI outputs when lives are on the line? The answer is in testing out in the field.
"As soon as you start getting real reps in exercises, you can see where we can trust it and where things fall over,” Van Roo explained.
Field testing reveals failure modes that are impossible to detect in laboratories. In one Joint Readiness Training Center rotation, Legion's team discovered a workflow that broke when systems desynced—a problem that only emerged under realistic operational conditions.
Guerra added the technical reality that makes testing essential: "These are statistical systems with a degree of error—10, 15, 20%. [They] require testing to make sure you're okay with those margins."
Trust isn't built through technical specifications or vendor promises. It requires:
- Extensive practice in realistic conditions
- Repeatable outcomes across multiple scenarios
- Complete attribution of every action taken
When Everything Breaks: AI in Degraded Environments
Most commercial AI platforms are built on assumptions of constant connectivity, high bandwidth, and centralized infrastructure. When communications are contested and the cloud is unavailable, they simply stop working.
The military equivalents are obvious: contested networks, jammed communications, tactical edge operations. The critical question is then: Does your AI work when conditions degrade? And if it breaks, what do you do then? It's an issue on leaders' minds.
"I'm always worrying that we might really start trusting all of this stuff, and then if we don't have that connection—what's our decision cycle look like?" Van Roo said.
One way to protect against this is by using an AI platform like Legion, which is built for operations (military or commercial) where favorable conditions cannot be assumed and does not require the cloud to function.
Other key foundational features that will stop your AI from failing include:
- Model flexibility: Treating models as interchangeable components
- Data sovereignty: Processing within your environment, not external clouds
- Deployment options: Supporting cloud, on-prem, air-gapped, and edge
Building Governance Into the Architecture
As soon as AI moves beyond just generating answers and starts executing across platforms, governance becomes an architectural requirement, not a policy decision.
Coleman posed the ethical question underlying all AI deployment: "We could build a machine to make really complex decisions. We can build anything, but should we? I personally draw the line at decisions about taking human life."
It's the right line to draw. But it highlights a broader governance challenge.
Why Procedural Governance Isn't Enough
As AI completes more work automatically across operational systems, how do you maintain accountability? The answer cannot be procedural, with policies and review processes bolted onto systems after deployment. Instead, governance must be architectural:
- Attribution: Showing not just what was done, but on whose behalf
- Audit logs: Capturing every action in real-time
- Granular permissions: Governing access at the workflow and model level
- Intelligent observability: Providing visibility into resource usage and performance
The question shifts from "Did a human approve this?" to "Is every action traceable, auditable, and operating within defined policies—even when executed by autonomous agents?"
Moving Forward
Van Roo's closing was direct: "It's gonna move as fast as we want it to."
AI in command and control is moving beyond chatbots. The future is AI that executes over multiple platforms—systems coordinating across applications, completing work automatically, and operating under human-defined policies, even when networks are contested.
The next decade of advantage won't go to those who adopt AI first, but rather those who operationalize it best.
Ready to move past demos? Learn how Legion helps enterprises and government agencies deploy AI with security, flexibility, and control built in. Contact our team.