Table of Contents
Table of Contents
AI in Defence Has Left the Research Phase
AI in defence is no longer a hypothesis under evaluation. It is already funded, in use, and scaling across the world’s major defence institutions. NATO’s Defence Innovation Accelerator for the North Atlantic (DIANA) is running its largest-ever cohort — 150 innovators, selected from a record field of applicants — with access to a growing network of accelerator sites and test centres across the Alliance.
NATO and the EU are pursuing this from different institutional angles: alongside DIANA, the European Union has committed more than €4 billion to defence research and development through the European Defence Fund since 2021, including, in 2026, a dedicated technological challenge on AI-based tactical situational awareness under its EU Defence Innovation Scheme (EUDIS).
AI is already changing how militaries plan, sense, maintain, defend, and move – and is running into real technical and institutional limits as it scales. This article maps where that impact is measurable today, what’s slowing broader adoption, and where the field is heading next.
Where AI Is Already Operational in Defence
AI’s footprint in defence extends well beyond autonomous weapons and headline-grabbing platforms. In practice, it shows up in six distinct areas of military operations — some highly visible, others running quietly in the background. These aren’t roadmap items: each is grounded in real, funded defence AI activity today, though maturity varies — from capability already fielded at scale to systems only now operational.
Decision Support and Command Speed
AI systems increasingly pull together sensor feeds, intelligence reports, and logistics data into a single operational picture commanders can act on faster — compressing the gap between data arriving and a decision getting made, particularly in fast-moving or contested environments where manual analysis can’t keep pace. These systems are meant to augment human judgment, not replace it.
Situational Awareness: Sensor Fusion and ISR
A closely related application: AI models that combine satellite imagery, radar, full-motion video, and signals data into a coherent, real-time intelligence picture. This is one of the more mature areas of defence AI, and one where computer vision and multi-sensor fusion do most of the heavy lifting.
Predictive Maintenance and Fleet Readiness
AI-driven predictive maintenance applies pattern recognition to equipment health data to flag likely failures before they happen, across vehicles, aircraft, vessels, and their subsystems. In practice, that means less unscheduled downtime, better parts planning, and higher fleet availability in both peacetime readiness and active operations.
Autonomous and Uncrewed Systems
Uncrewed platforms — aerial, maritime, and ground-based — are taking on tasks that are dull, dangerous, or hard to staff at scale, from persistent surveillance and resupply to minesweeping and reconnaissance in contested areas, with a clear trend toward greater autonomy in navigation and task execution — generally within human-defined mission parameters, not fully independent decision-making.
AI-Enabled Cyber Defence
Militaries and defence contractors use AI to detect anomalous network activity and respond to intrusions faster than manual security operations can — driven by necessity as much as choice, since increasingly networked, software-defined systems generate more to monitor and defend.
Logistics and Sustainment Optimisation
Less visible than autonomous weapons but arguably more consequential day-to-day: AI applied to demand forecasting, transport routing, and spares management — the unglamorous work of keeping forces supplied, and one of NATO DIANA’s own priority challenge areas for 2026.
Each of these six areas is being actively funded, coordinated, and accelerated by defence institutions on both sides of the Atlantic — which is what makes this shift durable rather than experimental.
The Institutional Push Behind Defence AI
That momentum doesn’t happen by accident. Three institutional efforts — one Alliance-wide, one EU-wide, one national — are funding, accelerating, and regulating this shift in parallel.
Governments and Alliances Are Funding the Shift
Different structures, the same underlying logic: governments are no longer waiting for AI capability to arrive — they’re building the infrastructure to pull it in.
- NATO (DIANA) — Selected 150 innovators from 24 NATO countries for its 2026 cohort, chosen from a record 3,680 applications, with access to a network of 16 accelerator sites, 200+ test centres, and 600+ mentors across all 32 member nations, aimed at moving dual-use technology into military use faster.
- European Union (EDF / EUDIS) — The European Defence Fund’s total budget runs to nearly €7.3 billion through 2027. Within that, its EU Defence Innovation Scheme (EUDIS), which specifically targets startups and smaller players, has €231 million allocated for 2026 alone — 23% of that year’s EDF budget
- United States (DIU / Replicator, and CDAO) — Two distinct efforts: the Chief Digital and AI Office (CDAO) coordinates AI adoption across the US defence enterprise, while the Replicator initiative – announced in August 2023 and delivered through the Defense Innovation Unit (DIU) — targeted multiple thousands of attritable autonomous systems within 18–24 months. Replicator fell well short of that target by August 2025, with technical failures in testing and procurement friction slowing delivery, and its portfolio was transferred that year to the newly formed Defense Autonomous Warfare Group (DAWG) under U.S. Special Operations Command.
Compressing the Prototype-to-Deployment Timeline
Speed distinguishes this wave of institutional effort from earlier defence innovation programmes. DIANA’s Rapid Adoption Service (RAS) exists specifically to shorten the path from tested technology to procurement contract — it produced its first NATO Ally–industry R&D contract under this mechanism in April 2026 (a UK undersea-robotics company, HonuWorx, contracted on behalf of Defence Research and Development Canada).
One distinction worth being precise about, since it’s often misstated:
Programme track | Typical maturity requirement |
TRL 4–7 (not near-finished products) | |
TRL7+ specifically, for deployment within months |
The second is a named exception, not DIANA’s general posture — a distinction worth knowing before assuming a deployment timeline.
Where Defence AI Regulation and Governance Stand Today
Regulation hasn’t kept pace with funding, but it hasn’t been entirely absent either. It’s fragmented across institutions rather than unified under one framework:
EU AI Act — Excludes only systems used exclusively for military, defence, or national-security purposes. A dual-use system, or a military system later repurposed for civilian, humanitarian, or law-enforcement use, falls back into the Act’s scope. That word — exclusively — is doing a lot of work, and it’s the detail most worth understanding for anyone building dual-use technology.
NATO’s own framework — its AI Strategy, revised in July 2024, is built around six Principles of Responsible Use (lawfulness, accountability, explainability, reliability, governability, bias mitigation). The strategy states directly that Alliance-wide AI test, evaluation, validation, and verification (TEV&V) will utilise the network of DIANA-affiliated test centres — a firm, explicit link between NATO’s AI governance framework and its accelerator infrastructure, not an implied one.
The overall picture: institutional momentum is running ahead of a governance landscape still being assembled piece by piece — and that gap is precisely where the real barriers to scaling defence AI begin.
The Barriers to Scaling AI in Defence
The institutional momentum described earlier is real, but it runs into a consistent set of technical and operational obstacles once programmes move from pilot to fielded capability. Four barriers show up repeatedly, explaining much of the gap between funding announced and capability deployed.
Data Scarcity
Many of the scenarios AI systems most need to handle reliably are also the hardest to collect data for. Combat damage states, sensor-failure conditions, and adversarial signatures are rare by definition, and often classified or dangerous to capture in the real world — waiting for enough real-world examples to accumulate isn’t always practical, especially for low-frequency, high-consequence events. Synthetic or simulated data is one of several tools defence programmes are exploring to help close this gap, alongside approaches like data-sharing agreements between allied forces and improved sensor instrumentation on existing platforms.
Validation and Test-and-Evaluation Complexity (TEV&V)
Proving an AI system performs reliably — especially in contested or degraded conditions — is harder than building it in the first place. Performance in a lab or simulation doesn’t always transfer cleanly to contested electromagnetic environments or adversarial conditions, which is precisely why DIANA’s test-centre network exists and why NATO’s AI Strategy leans on it for Alliance-wide validation. That infrastructure is being built in parallel with the technology itself, not after it.
Telling in this respect: NATO’s first Rapid Adoption Service award, the HonuWorx contract is itself built around developing a high-fidelity simulation suite to demonstrate mission potential at operating depths that are difficult to test against directly, specifically to reduce technical risk ahead of a deployable system.
Operational Trust and Human-Machine Teaming
Even a technically validated system still has to earn the confidence of the people using it. Commanders and operators need to understand why an AI system is recommending a given action, not just what it recommends, and accountability questions — who is responsible when an AI-assisted decision goes wrong — remain largely unresolved in practice, especially in the high-stakes, time-pressured situations where AI support is supposed to help most.
Legacy Interoperability and Expanding Attack Surface
New AI capability rarely gets deployed on a blank slate. Most militaries are integrating it into command and control infrastructure that is decades old, and every new AI-enabled system is also a new potential target — more networked software means a larger attack surface, not a smaller one. Interoperability is a recurring reason deployment timelines slip.
Replicator is arguably the sharpest illustration of these barriers compounding: a well-funded, senior-sponsored programme whose shortfall traced back to testing and procurement friction rather than to any lack of ambition or money. Its transfer to DAWG reads as a structural correction, not proof the barriers don’t matter.
Application Area | Institutional Driver | Primary Barrier |
Decision support & command | DIANA Decision Superiority challenge | Operational trust & human-machine teaming |
ISR / situational awareness | EDF 2026 AI tactical situational awareness challenge (EUDIS) | Data scarcity |
Predictive maintenance & logistics | DIANA Critical Infrastructure & Logistics | Legacy interoperability |
Autonomous / uncrewed systems | DIANA Autonomy & Unmanned Systems; US DAWG | Validation & TEV&V |
Cyber defence | National cyber commands; EDF cyber R&D | Expanding attack surface |
These barriers rarely appear one at a time — most systems face two or three of them at once. That’s the real reason adoption is slower than funding levels alone would suggest, and it’s also where the next five years are likely to see the most meaningful progress.
What the Next Five Years Look Like
The barriers previously discussed aren’t likely to be solved all at once, and they aren’t likely to stop adoption either. Based on where institutional priorities and funding are already pointing, three trends — and the shift underneath them — look set to define the next five years of AI in defence.
Convergence With Adjacent Technologies
AI won’t scale in isolation — it’s already being developed alongside a specific set of adjacent priorities. DIANA’s ten 2026 challenge areas cluster into four groupings: resilience and infrastructure (energy, critical infrastructure and logistics, extreme environments); sensing and decision-making (advanced communications, contested electromagnetic environments, data-assisted decisions); operational domains (maritime, space); and autonomy and human factors (autonomous systems, human resilience and biotechnology).
The pattern that matters isn’t the list itself: institutions aren’t funding AI as a standalone capability, but as one layer across nearly every other defence priority, from power grids to space operations.
Faster Commercial-to-Military Pipeline
The accelerator model is deliberately built to shorten the distance between a commercial AI breakthrough and a fielded military capability: DIANA and EUDIS lower the barrier to entry for non-traditional suppliers, while US efforts like DIU engage commercial and non-traditional vendors directly rather than relying solely on established primes. Companies without a defence pedigree are increasingly viable entrants, provided they meet the maturity and validation bar required — and that pipeline is likely to keep compressing.
Toward Defence-Specific AI Assurance
Governance is catching up institution by institution, not through one unified framework. The European Defence Agency’s “Trustworthiness for AI in Defence” white paper (May 2025) proposes four principles for military AI — lawfulness, explicability, robustness, controllability — alongside NATO’s own six-principle framework: two frameworks converging on similar ground, developed independently, not yet merged into one standard.
Evaluation Shifting From “Can It Work” to “Can It Be Trusted at Scale”
This is the underlying shift beneath the three trends above. Early-stage adoption was largely a feasibility question — could a given model or system perform the task at all — and that bar has mostly been cleared across the application areas we covered.
What’s replacing it is harder: whether performance holds up reliably under real operational conditions, at the scale a military actually needs, under exactly the conditions hardest to source data for or test against in the first place.
From "Can It Work" to "Can It Be Trusted"
The picture across this article is consistent: AI in defence has moved past the pilot stage. It’s operational across command support, ISR, maintenance, autonomy, cyber defence, and logistics. It’s backed by serious institutional funding from NATO, the EU, and the US. And it’s running into a predictable, well-understood set of barriers — data scarcity, validation complexity, operational trust, and legacy integration — that no amount of funding alone will resolve.
The next five years will be defined less by whether AI can perform these tasks — that question is largely settled — than by whether it can be proven trustworthy enough, under contested and degraded conditions, to be relied on at scale.
For anyone evaluating a defence AI claim or vendor, a few questions cut through most of the noise:
- Maturity — What TRL is the solution actually at, and does that match the deployment timeline being promised?
- Validation — How has performance been tested under degraded or contested conditions, not just controlled ones?
- Data provenance — Where does the training and validation data come from, particularly for rare or high-consequence scenarios?
- Governance fit — Does the system fall inside or outside frameworks like the EU AI Act’s military exclusion, and has that been assessed deliberately or assumed?
- Integration path — How does this fit into existing, often decades-old command and control infrastructure — not a greenfield system?
The industry doesn’t have easy answers to any of these yet — which is precisely why they’re worth asking early.
Anyverse’s own work applying synthetic data to safety-critical perception systems sits adjacent to exactly this problem — though closing the defence data gap at scale will take more than one company’s effort.
