The harness, not the model: Defence's real AI decision

AI and autonomy spending is set to rise from £380m next year to £1.3bn by the end of the decade. The direction is set. 

The Defence Investment Plan (DIP) 2026 commits £298bn over four years and is unambiguous about where advantage will come from. Not from platforms alone, but from the rapid, sovereign exploitation of data, software, artificial intelligence and autonomy. 

But there is a gap between ambition and capability, and it is not the gap most people assume. Defence does not have a shortage of impressive AI. It has seen years of promising demonstrations, prototypes and laboratory successes. What it has far less of is AI that has made the journey from demonstration to deployed mission capability: assured, secured, governed and continuously improving in environments where communications are contested, adversaries adapt and decisions carry real consequence.

That journey is not completed by building a better model. In mission-critical environments, the model is one component of capability. The decisive advantage comes from the system around it: trusted data foundations, mission-specific assurance, security by design, human oversight, operational monitoring, deployment architecture and the feedback loops that turn operational experience into improvement. We call this the AI Harness. It is what turns a promising algorithm into trusted operational capability.

This is not an argument Thales makes from the sidelines. We have been engineering AI into mission-critical systems for around 40 years. Today that work is carried by more than 800 AI and data engineers and scientists, around 100 of them PhDs, backed by more than 250 patents and, through cortAIx, a focused effort to accelerate trusted AI into products our customers already operate. We sponsor around 50 PhDs at UK universities at any one time.

The harness argument comes from the experience of fielding AI where it has to work, not from a laboratory.

The white paper below draws three conclusions that matter for how Defence spends the next four years.

First, Defence should organise AI delivery around operational problems, not technology demonstrations, and measure investment by mission effect rather than experiment volume. The harness, not the model, should be the common route from prototype to deployed capability.

Second, Defence needs a federated AI ecosystem rather than a single central programme. The operational demand for AI is distributed across every command and domain, and the DIP already reflects that reality. The centre should set the rules of the road; the domains should own mission outcomes; and industry, including SMEs, academia and allies, should provide assured, reusable capability that scales across contexts. Coordinated enough to avoid fragmentation, distributed enough to deliver at the edge of operational need.

Third, and most important, human accountability is the design principle. The next phase of AI will not simply detect and recommend. Agentic systems will plan, use tools and act. That changes the assurance question fundamentally: it is no longer only whether a prediction is accurate, but whether behaviour can be controlled, constrained, monitored and held accountable. The most effective Defence AI will put humans in a better decision loop, reducing cognitive burden and improving the speed and quality of decisions. Our position is simple: AI wherever it accelerates understanding, human judgement wherever accountability matters, integrated always.

None of this is a brake on ambition. Done properly, assurance is what allows Defence to move fast, because capability that is trusted gets used, and capability that is not stays in the laboratory. The DIP has created the investment framework. The task now is to build the harness that lets the UK field AI when it matters most, and the industrial ecosystem that lets Defence apply it everywhere it is needed.