There is a dominant assumption built into how most defence AI systems are designed and procured. It holds that capability can be defined at the point of deployment and will remain sufficient. That assumption is wrong, and the consequences are already visible.

Defence AI procurement is largely structured around the idea of a finished product. A system is specified, developed, validated, and fielded. The expectation is that what was good enough at delivery remains adequate in use. This was a reasonable model when the pace of change was in years. It no longer holds.

Security environments do not pause at the pace of procurement cycles. Adversaries adapt. Tactics shift. Connectivity degrades. Operating conditions in the field diverge from the assumptions baked into system design. A static system — AI trained on a fixed dataset and deployed without a mechanism for continued learning — begins falling behind from the moment it leaves the lab.

Why static autonomy breaks in practice

In controlled environments, fixed capability is manageable. In real operational conditions, it becomes a critical vulnerability. Connectivity degrades at the worst possible moment. Adversaries adapt faster than update cycles allow. Systems that cannot respond begin to fail, and the reason is rarely poor engineering. The problem is structural: they were deployed once, on closed platforms, where capability evolution remains with the vendor.

Capability can no longer be defined by what a system does at deployment. It must be defined by how a system performs as conditions change, and how quickly it can be updated while in use.

The gap is architectural, not operational

‍Most deployed AI systems are built on closed architectures. Capability is locked at the point of delivery. When conditions change, updating the system might even mean going back to the vendor, renegotiating a contract, and waiting for the next programme cycle. For a system that must perform under real-world pressure, these timelines are not acceptable.

The consequences run deeper than slow update cycles. Closed architectures prevent the continuous feedback loop that keeps AI systems current. Operational data generated in the field cannot be used to improve deployed AI models. Every mission produces data that could make the next one better. In most systems today, that data is lost.

Sovereignty is an engineering problem

‍The conversation about European technological sovereignty often centres on ownership: where systems are built, who funds them. This framing misses what matters most. Sovereignty is defined by who has the authority to update a system, inspect it, and determine how it behaves after deployment. Even a system built in Europe, but dependent on a closed external platform for its updates, its AI models and its data pipeline, is not a sovereign system. The dependency simply moves.

Buying European is not enough. Closed systems do not become more sovereign over time. They become more dependent. The only path to genuine control is open architecture from the start, not as a procurement preference, but as a design principle embedded before the first line of code is written.

The decisions being made now

‍Europe is rebuilding its defence capability at speed. The risk is not a lack of ambition. It is building on foundations that will compound the same structural problems at larger scale.The next generation of critical systems will be defined by three things: how quickly they can be updated when conditions change, how reliably they operate under real-world constraints, and who retains control over how they evolve.

That last question is not technical. It is political. And right now, the answer depends on architectural decisions being made today, by the people building these systems, and by the organisations choosing which ones to deploy.

Static autonomy is obsolete. What replaces it is not more capability at the point of deployment. It is the ability to keep building capability after deployment. That is what adaptive intelligence means: systems that do not stop learning when they leave the lab.

NestAI is a European AI lab for defence building the adaptive operating system for modern battlefield operations. We develop adaptive intelligence for unmanned systems that continuously learns from operational data and adapts to changing conditions, on an open and modular platform built for European control.

‍