Five principles for bringing physical AI into the national airspace

The national airspace is entering a new era of complexity. Legacy systems are being modernized, uncrewed aircraft and advanced air mobility are entering shared airspace, and the volume of data flowing through air traffic operations is growing faster than traditional maintenance and staffing models can keep pace with. At the center of this shift is a class of technology known as cyber-physical systems, sometimes called “physical AI,” which are systems that not only crunch data but also sense, interpret and act on the physical world in real time, much like a pilot or controller does.

The temptation in moments like this is to treat AI as a switch to flip — deploy the model, capture the efficiency, move on. That instinct undersells an environment where precision and safety are nonnegotiable. Having spent years inside programs that modernize mission-critical aviation infrastructure, the more useful question isn’t whether physical AI belongs in the national airspace. It’s where it earns its keep and what separates a well-executed deployment from one that falls short. Five things stand out.

1. Physical AI mirrors how pilots already operate, which is why it deserves the same rigor pilots earn through training.

Pilots synthesize sensor readings, radar and instrument data to make split-second decisions across thousands of hours of training. Physical AI is built on the same premise: continuously ingesting sensor and operational inputs and then reasoning through them fast enough to act. The parallel is useful, and it is also a reminder. A system modeled on expert judgment still has to earn that judgment through rigorous validation. It requires real testing against operational edge cases before it ever touches live infrastructure. That groundwork is what turns a promising capability into one worth trusting.

2. The biggest near-term payoff is in ground-based infrastructure, and it is a data problem before it is an AI problem.

Radars, runway lighting and other ground-based assets fail in ways that ripple through entire operations. Most maintenance today still relies on checklists built around assumed failure rates that do not account for how environmental exposure or manufacturing variability change a part’s real lifespan over time. The organizations that get value from AI-enabled predictive maintenance first will be the ones that treat their sensor and asset data as an actual asset. They will clean it, structure it and feed it continuously to the model rather than bolt AI onto data pipelines that were never built for it.

3. New entrants are straining a system built for a different era, and the fix is architectural rather than purely algorithmic.

Uncrewed aircraft, advanced air mobility and commercial space are pushing into airspace that has historically operated along fixed, high-capacity corridors, much like a rail network. As more operators seek access to the full three-dimensional airspace, congestion, delays and safety risks all increase. Applying physical AI to flight planning and traffic prediction helps, but only if it is layered onto a modern, cloud-native backbone.

We have seen firsthand that migrating decades-old infrastructure to a micro-serviced, CI/CD-backed architecture with minimal operational disruption is possible. This modernization serves as the foundation that makes advanced AI capability viable at scale, rather than a fragile add-on to legacy systems.

4. Automation should extend human oversight, and the deployment approach should prove that before go live, not after.

The strongest use cases pair automated sensing and decision-making with human supervision and quality control. These efforts automate routine flight planning, ground control functions and logistics while keeping people in the loop for judgment calls. The organizations that do this well build and stress test working prototypes against real operational scenarios before deployment. This validates that a solution is safe, will not interrupt operations and is fully understood by the team running it. That prototype-first discipline is what separates a defensible modernization roadmap from a leap of faith.

5. Reliability of the AI itself is now a mission requirement, and safety-centered engineering has to be the design constraint rather than an afterthought.

These systems need to continuously observe, orient, decide and act on fast-changing conditions. That requires a robust, redundant and often local compute stack, not just a capable model. Any physical AI system touching mission-critical operations should be engineered, tested and vetted for functionality and security with the same rigor applied to any system where a mistake carries real physical consequences. Agencies should pressure-test vendors on this discipline as closely as they evaluate the underlying algorithm.

Ultimately, none of this is about replacing the expertise that keeps the national airspace safe. It is about giving that expertise better tools, built and proven the right way. The organizations that get ahead of this shift will be the ones that pair a clear-eyed appetite for physical AI with partners who can demonstrate — not just promise — that they know how to implement it responsibly at mission scale.