For MRO and sustainment teams, AI in aerospace has mostly meant a generative chat assistant over technical manuals. Useful for retrieval, but it leaves the harder problem untouched. A maintenance technician spends about 61% of the day searching for tools and parts, on top of admin, by IFS's count, and the people who hold decades of that knowledge are retiring faster than they can be replaced: Boeing's 2025 Pilot and Technician Outlook projects the industry will need roughly 710,000 new aviation maintenance technicians over the next 20 years. Fewer technicians, less usable time from the ones on the floor, and their institutional knowledge walking out the door.
This post is about the shift past the search box: from generative AI that answers questions to AI agents that carry out the work, under human command.
Why is AI in aerospace moving from generative tools to AI agents?
AI in aerospace is moving from generative tools to AI agents because answering the question is only the first step. The work is the chain of actions after it: opening the work order, checking the part, scheduling the job, updating the record. Generative AI is reactive, producing content when prompted, and a human still takes every one of those next steps. Agentic AI is software that plans and carries out multi-step work across the systems where the work lives, the CMMS (Computerized Maintenance Management System), inventory, and technical documentation, under human direction and approval.
The risk is different too. An agent updating a component's time-before-overhaul limit in the CMMS, or a procedure followed on pressurized equipment, carries configuration-control and airworthiness consequences a chat box never does, which is why governance decides whether any of this is usable. Agents are also not a guaranteed win: Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027. Regulated, safety-critical environments raise the bar further, and the rest of this post is about what that takes.
What does agentic AI actually do in aerospace maintenance operations?
On reactive jobs, about one in three work orders sends a technician back for the right part: leave the shop, reach the asset, find the part missing, walk back, and the work order ages another shift. Agentic AI in aerospace maintenance operations is built to close gaps like that, by carrying out the multi-step jobs a maintenance team already does, across the systems they already run, with a human approving the steps.
Legion Intelligence delivers this as the Maintenance Technician Pack, a role-based bundle of apps that follow the maintenance job from inspection to closeout. A supervisor turns an inspection report into structured work orders in the CMMS with Inspection WO Generator, then assigns them by craft, certification, location, and availability with Work Order Scheduler. Before a technician leaves the shop, Kit Builder reads the asset's work-order history, predicts the likely failure mode, and names the parts to bring, while Parts Manager checks availability across storerooms and reserves them against the work order, so the return trip above does not happen.
On the job, Manual Assistant answers procedure questions in natural language from the technical manuals the team already keeps in systems like SharePoint, returning the source with every answer. On the parts side, Reorder Point Optimizer recalculates reorder points from actual usage and lead times, which matters most on long-lead items where a wrong threshold means a stockout.
Two outcomes anchor the Pack. Manual Assistant drove a 25% increase in technician capacity in one deployment. And because it turns a retiring technician's manuals and history into answers any technician can reach, it keeps institutional knowledge on the floor as the workforce turns over. The apps are mobile-first, because a tool that needs a walk back to a desktop does not get used on a shop floor. The same approach extends to test and quality engineers, who can surface prior non-conformance findings and the fixes that resolved them, with linked sources.
Across every role the pattern is the one that separates a Pack from a chatbot: agents draft the work order, reserve the parts, surface the procedure, and update the record, a human approves before anything is final, and every action carries full attribution. Agents move the steps; the human stays in command.
What is prescriptive maintenance, and how is it different from predictive maintenance?
Prescriptive maintenance is the practice of recommending a specific corrective action, and increasingly starting the digital workflow that turns it into a generated, attributed work order. Predictive maintenance only forecasts that a failure is likely. The gap between the two has always been the human handoff: a prediction still has to become a diagnosis, the right parts, an assigned technician, and a closed work order.
Predictive maintenance
- Answers: when is a failure likely?
- Output: a forecast or flag for a human to act on
- Role of AI agents: flags the condition
Prescriptive maintenance
- Answers: what action should be taken, and can the workflow start now?
- Output: a recommended action, and with agents, a generated, attributed work order
- Role of AI agents: carries the flag into a work order, parts check, and assignment, under human approval
That gap is what agents close, on the digital side of the work. Instead of leaving a prediction in a dashboard for someone to act on, an agentic workflow carries the flagged condition into a generated work order, a parts check, and an assignment, all under human approval. The physical repair stays with the technician. The payoff lands in the metrics sustainment teams track: fewer return trips, higher first-time fix rates, shorter turnaround. Analyses from McKinsey and Deloitte put the downtime reduction from predictive maintenance at up to roughly 50%.
There is also a measurement gap. The U.S. Government Accountability Office reports that the Department of War spends roughly $90 billion a year on weapon-system maintenance and generally lacks the metrics to evaluate the results of its predictive maintenance. Letting agents act is only safe when there is a record of what was done and on whose behalf.
Why does governance matter for AI in safety-critical aerospace?
Governance matters for AI in safety-critical aerospace because agents act on live systems, and acting on a live system means someone has to stand behind the result. In practice that means a maintenance team can always see who approved an action, what source supported the answer, what system was changed, and what record proves it. Humans stay in command: every step runs under human direction or approval on the policy the team sets, and full attribution records what was done and on whose behalf, including agent actions.
Two of those properties carry the weight. The first is accuracy. A technician tests a new tool with questions they already know the answer to, and one wrong answer on pressurized equipment ends both the trust and the adoption. Answering with the source attached, and saying "I don't know" rather than guessing, is what keeps a hallucinated fix code or a wrong torque value out of a safety-critical maintenance log. The second is the record itself. Recall the GAO finding that the Department of War largely lacks metrics on its predictive-maintenance results. A governed system produces that record as a byproduct of doing the work, which is the evidence a condition-based-maintenance-plus (CBM+) program needs to show it is working.
Speed and resilience follow from the same design: agents compress the time between a flagged condition and a completed, attributed action, and the same Pack and governance run whether the deployment is connected or disconnected. This is where readiness comes back.
Can AI run in on-prem, air-gapped, or classified aerospace environments?
Yes. AI can run in on-prem, air-gapped, and classified aerospace environments, and that is where most tools stop being an option, because they assume constant cloud access. Legion runs where the work runs: cloud, on-prem, air-gapped, and edge, authorized across IL2 through IL6 and FedRAMP High, with ITAR handled through data sovereignty, U.S.-person access controls, and on-prem or air-gapped deployment, so your data stays in your environment and models never train on it.
For genuinely disconnected work, Centurion by Legion Intelligence is the deployable edge AI system, integrated hardware and software, that carries Legion's Packs and the same governance to the edge, the Maintenance Technician Pack among them. It routes between connected and locally hosted models automatically, with no operator action to switch. That was proven at USSOCOM TE 26-02, a disconnected military evaluation where Legion ran planning workflows on government data with the network down and failed over to local models on its own, and at Scarlet Dragon 26-01, where Centurion installed to operational capability in a single day. Those are tactical evaluations, but the property they prove, governed AI that keeps working when the network does not, is what a secure backshop, a classified depot, or an ITAR-restricted line needs. The same platform runs on-prem at the Department of Energy's SLAC National Accelerator Laboratory with more than 100 active users.
The common reaction at a defense systems integrator is that they could build this themselves, and many have built prototypes. The prototype is the straightforward part. The hard part is keeping an agentic system accurate, governed, and integrated with the systems you already run, and current across a multi-year lifecycle: revalidating each model change against airworthiness and accreditation requirements, holding model parity across air-gapped sites that cannot phone home for updates, and re-certifying every integration as the underlying systems change. That sustaining burden, not the demo, is why owning a system built for it usually beats building from scratch, and why Legion layers agents on top of the CMMS you already run, such as IBM Maximo, rather than replacing the system of record.
Legion fits when you have a system of record to connect to, regulated or safety-critical work that needs accreditation and an audit trail, and a need for capability in weeks without a multi-year services program. It is not the shortest path if you want one monolithic platform to replace all of your systems, if you have no system of record to build on, or if a general-purpose chatbot over documents is all you need.
What this means for readiness
Readiness in aerospace comes down to closed work orders and aircraft that are ready when needed. Chat assistants made the underlying information findable. Governed agents carry out the work that turns information into a closed work order, on the systems teams already run, and they leave a record of every action, which is what a readiness review or a contract recompete will ask to see.
To see the Maintenance Technician Pack applied to aerospace maintenance, parts, and supply in your environment, request a demo.