Public debate about military AI concentrates almost entirely on one application — autonomous weapons — which, judging by open-source documentation, is neither the most widespread nor the most consequential use of AI in defence today. This overview maps what publicly available sources actually document, category by category, and flags where marketing outruns evidence.
Intelligence, surveillance and reconnaissance
The best-documented military AI application is the least cinematic: processing more sensor data than human analysts can. Machine-learning systems classify objects in satellite and drone imagery, transcribe and translate intercepted communications, and fuse feeds from disparate sensors into a common picture. This is where military AI began at scale — the US Project Maven, publicly acknowledged in 2017, was an imagery-analysis programme — and open reporting suggests it remains the largest category by actual deployment.
Logistics, maintenance and personnel
The most cost-effective applications, according to think-tank analyses such as CSET's work on AI for military decision-making, are back-office ones: predictive maintenance of aircraft and vehicles, supply-chain optimisation, medical triage support, and administrative automation. They attract little public attention precisely because they resemble civilian enterprise software — which is also why they raise the fewest novel legal questions.
Decision support and command and control
A more consequential middle ground: systems that recommend rather than act. AI-assisted targeting support — systems that nominate candidate targets for human review — sits here, as do broader command-and-control programmes such as the US JADC2 initiative (see also the UN Office for Disarmament Affairs' overview of AI in the military domain), which aims to link sensors across services into one network. The legal and ethical weight of this category comes from a subtle mechanism: formally, a human decides; practically, the tempo and volume of machine recommendations can reduce that decision to ratification. We examine that mechanism in our analysis of IHL and AI-assisted targeting and in the essay on meaningful human control.
Autonomous and semi-autonomous platforms
Uncrewed aerial, ground and naval vehicles increasingly navigate, patrol and identify objects with varying degrees of autonomy. Open sources document loitering munitions with autonomous terminal guidance and experimental uncrewed wingman aircraft. The category that dominates headlines — weapons that select and engage targets without human intervention — remains, per the public record, narrow in confirmed deployment; the definitional and regulatory debate around it is covered in our summary of the UN LAWS discussions and the distinction between autonomous and AI-enabled weapons.
Cyber and information operations
Open literature documents AI use in network defence (anomaly detection at machine speed) and flags its offensive potential — vulnerability discovery, tailored phishing, synthetic media for influence operations. This category has the widest gap between what is claimed and what is verifiable: operational cyber capabilities are precisely the ones states do not document publicly, so honest analysis must mark most specific claims here as unconfirmed.
Reading the field honestly
Three cautions for anyone consuming reporting on this subject. First, procurement announcements are not deployments — a contract award tells you about budgets, not capabilities. Second, vendor demonstrations are not operational evidence — the environments are curated. Third, the loudest category is not the largest — by the documentation available, analysis and logistics dwarf autonomous weapons in actual use, while receiving a fraction of the scrutiny. An accurate picture of military AI is mostly a picture of software that processes, predicts and recommends — with humans still, formally, deciding. Whether that formal role stays meaningful is the question this site exists to track.
Agent AI Army is an editorial publication. We analyse publicly available sources — SIPRI, CSET, UNODA, official procurement records and peer-reviewed literature — and neither provide nor broker any technology or services.