Hybrid AI in defence and aerospace: what goes private, what goes managed

Engineers in defence and aerospace already use AI. The question most organisations are wrestling with is not whether, but where: which work can run on a managed model such as Claude or GPT, and which must never leave infrastructure you control. Treating that as a single decision for the whole company is what makes AI programmes stall.

A hybrid approach answers it task by task. Every request passes a policy check before any model sees it. If it involves controlled data, or if the check is unsure, it goes to private models running in your cloud, data centre or air-gapped network. If it involves only public information, it may go to a managed model, for the task types you have allowed.

What should stay private

The private side is larger than many teams expect, because so much of the daily work in this industry touches controlled or confidential information:

• Engineering data. Drawings, CAD metadata, specifications, design standards and lessons learned, including anything with export-control markings.
• Export-controlled technical data. Material within the scope of ITAR, the EAR or the EU dual-use regulation, where access must follow licence and nationality rules.
• Maintenance and MRO records. Work orders, task cards, defect reports and component histories, which also belong to your customers.
• Quality records. Non-conformance reports, concessions and supplier corrective actions.
• Programme and commercial data. Contracts, pricing, bids and customer requirements.
• Test, sensor and inspection data. Flight-test and qualification results, engine-health data, and borescope or NDT imagery.

What can use managed AI

The managed side is narrower but genuinely useful. Frontier models are very good at reading and summarising large volumes of public text, and a good deal of work in this industry depends on it: new and amended regulations, published airworthiness directives, public tender notices, defence budget documents, standards news and industry reporting.

The key design rule is that the managed model only ever sees the public source. When the result needs to be connected to your own fleet, programmes or capabilities, that mapping runs privately. A managed model can summarise a new airworthiness directive; a private model works out which of your aircraft it affects.

When in doubt, it stays private. The cost of a wrong private route is a slightly slower answer. The cost of a wrong managed route can be an unauthorised export.

Right-sizing the private side

Keeping work private does not mean running the largest possible model for everything. Most requests, such as work-order triage, document checks and procedure look-ups, are handled well by small language models that answer quickly and cheaply. Larger open models are reserved for root-cause analysis, long documents and complex drafting. That is what keeps private AI affordable when work orders, sensor data and documents arrive around the clock.

Making the policy check trustworthy

A routing rule is only as good as its evidence. Three practices matter:

1. Default to private. Uncertain classifications stay inside your boundary.
2. Let your teams tighten it. Security and export-control officers should be able to switch managed AI off for any task type, customer or programme, or entirely.
3. Log every decision. Each request’s route, sources and approvals should be recorded in your infrastructure, so you can answer an auditor’s question months later.

Handled this way, hybrid AI is not a compromise between capability and control. Your engineers get frontier models where public information helps, and your technical data stays where your contracts and export licences say it must.