Artificial Intelligence‭: ‬ 
When AI Becomes An Actor

Militaries are moving artificial intelligence from experimentation into operations. Uncrewed aerial and surface vehicles can now navigate, identify targets and execute the kill chain on the battlefield. Recent examples from the war in Ukraine show that inexpensive autonomous platforms, such as long-range drones, are dominating warfare and forcing both sides to compete through innovation. The changing character of warfare points towards a future in which autonomy becomes central to military power. As autonomous capabilities expand, militaries that achieve decision advantage will outcompete others.

Yet decision advantage in modern warfare lies not only in autonomous capabilities but also in the integration of AI into command structures. Modern militaries are seeking to achieve faster decision cycles in crises in a way that allows them to determine the operational tempo and manage escalation. Recent examples such as the US’ Combined Joint All-Domain Command and Control (CJADC2) aim to achieve this by connecting sensors and command systems across land, air, sea, space and cyber.

The US military aims to use AI and machine-learning tools to process large volumes of information and accelerate the commander’s decision cycle. The Pentagon’s Project Maven is another example. The project began with the need to automate the application of machine learning to imagery and full-motion video to handle the increase in drone footage and images. However, what started as a standard machine-learning task is expanding into targeting support, logistics planning and predicting supply requirements.

With the introduction of agentic AI, these goals appear more realistic. Systems can now break a problem into smaller tasks, use digital tools, communicate with other agents, and take actions across a workflow. A military planning agent could retrieve intelligence from multiple sources, from field reports to signals intelligence, generate possible courses of action, compare logistical requirements, identify legal constraints, and update its recommendations as new information arrives. 

However, military AI efforts are not possible without sovereignty. If militaries cannot achieve sufficient ownership and control of these systems, decision advantage will remain limited. 

Sovereign military AI therefore means more than possessing advanced technology. It requires the organisational capacity to use AI, the ability to rewire staff structures, military doctrines adapted to human-machine teaming, sufficient computing power to operate AI systems, specialised national data, and benchmarks that determine whether these systems improve military judgement.

Sovereignty in military AI has many pillars. The first is organisational adaptation. To achieve sovereignty in modern command structures, militaries need to decide how much authority can be delegated to AI, which decisions require human approval and who remains responsible when an agent acts incorrectly. Autonomy is best understood through the authority given to a system. Future staff structures must allow new forms of information movement between units and redefine the relationship between commanders and machines. Staff officers will increasingly need to supervise AI workflows, question machine recommendations and intervene when systems move outside their assigned authority. Military staff structures should enable human-AI teaming, while future concepts and doctrines should clearly define the authority granted to AI systems.

Secondly, testing and evaluation become a key component of this adaptation process. The restructuring of military staff and doctrines should incorporate proper test and evaluation frameworks. This will help to ensure that the knowledge foundations and world models of AI systems remain aligned with institutional priorities. These systems must also be updated regularly to address changing operational environments. Benchmarking is already common in AI and can be used for military AI. Yet existing benchmarks measure capabilities such as coding, mathematics or general reasoning. They reveal little about whether a model can interpret a commander’s intent, detect deception, revise an assessment, follow the law of armed conflict or connect tactical actions to strategic objectives. Military benchmarks should place systems inside realistic scenarios and measure decision quality, speed, uncertainty, legal compliance and adaptability.

The same principle applies to data. Military AI must learn from information that reflects national doctrine, geography, force structure, intelligence requirements and likely missions. AI models trained mainly on general internet data cannot be expected to understand country-specific operational needs. Similarly, models should be tested and benchmarked against customised datasets rather than only against general global test sets. Each nation and military institution will need to develop datasets for training and evaluating its AI systems. Sovereign data infrastructure must allow militaries to collect, label, secure and update mission-relevant information while controlling who can access it.

Finally, sovereignty is possible only with stable access to compute for AI inference. Modern agentic systems rely on significant compute infrastructure for large-scale data processing. A military that depends entirely on external cloud providers, foreign data centres or remote technical support may lose access to critical capabilities during a crisis. Modern militaries must therefore ensure stable access to computing power for extended periods. Sovereignty does not require every country to build a frontier AI model. It does, however, require sufficient national capacity to operate and adapt mission-critical systems and to secure sensitive data and logs when external connections are degraded or denied.

The future of military AI relies on sovereignty. This requires deliberate efforts to restructure military staffs and command-and-control systems, integrate autonomous capabilities into decision-making workflows, establish sovereign data and test and evaluation structures, and secure stable access to compute infrastructure. With sovereignty, decision advantage will shape both peacetime deterrence and wartime crisis management.