The EU AI Act and Defence: What the Article 2(3) Exclusion Does — and Doesn't — Cover

Partly excluded. That is the honest one-line answer for a defence company, and the word "partly" is where the whole problem lives. Article 2(3) of the EU AI Act, Regulation (EU) 2024/1689, removes from scope any AI system placed on the market, put into service, or used exclusively for military, defence or national-security purposes — whoever operates it. But a defence contractor is not a military unit. Its recruitment AI, its corporate chatbot, its civil and dual-use products live entirely outside that carve-out and are fully in scope. Drawing that line correctly is the single most important compliance task for a defence firm.

What Article 2(3) actually says

Article 2(3) is a purpose-based exclusion, not a sector-based one. The Regulation does not apply to AI systems where they are placed on the market, put into service, or used — with or without modification — exclusively for military, defence or national-security purposes. The exclusion attaches to the use, and it applies regardless of who operates the system: a private contractor running an AI tool exclusively for a defence purpose is as excluded as the armed forces themselves.

Two features of the text do the work. First, it is purpose-bound — the trigger is the purpose the AI serves, not the identity of the company. Second, and decisively, it requires that purpose to be exclusive. An AI system used for military and civilian ends does not get the exclusion. That single word is the dividing line a defence firm has to manage.

The "exclusively" test

The exclusion is binary on a per-system, per-use basis. Ask of each AI system: is it used only for a military, defence or national-security purpose? If yes, Article 2(3) takes it out of scope entirely — no classification, no high-risk stack, no transparency duty. If it has any substantive civilian or commercial use, the exclusion does not apply and the system is assessed under the Act like any other.

The recital framing reinforces this: where a system is developed or used for both excluded and non-excluded purposes, the dual nature means the Act applies. So a firm cannot label an entire AI estate "defence" and walk away. The test runs system by system, use by use.

Where a defence contractor is fully in scope

The systems that catch defence firms out are the ordinary corporate ones — the AI that has nothing to do with a weapon.

HR and recruitment AI — Annex III point 4. A defence contractor screening engineers, analysts or production staff with an AI hiring tool is running a high-risk Annex III point 4 employment system. The product the company builds may be a missile guidance subsystem; the CV-screener that hired the team is plain high-risk AI, fully in scope.

Corporate chatbots and generative content — Article 50. A customer-service or website chatbot, or AI-generated marketing content, falls under the Article 50 transparency rules, which have applied since 2 August 2026: users must be told they are interacting with AI and synthetic content must be marked — design duties on the tool's provider, which is the firm itself if it builds or white-labels the tool — and the firm must disclose any deepfake it publishes. Nothing military about it, fully in scope.

Civil and dual-use products. A defence firm with a civil aviation line, a commercial cybersecurity product, or a dual-use sensor sold into both military and civilian markets cannot claim Article 2(3) for the civil or dual-use part. Those products are assessed on their use like any other manufacturer's.

Article 5 prohibitions and Article 4 literacy apply to in-scope systems as normal — in force since 2 February 2025 — and the AI literacy duty reaches the staff operating any in-scope AI.

The line, drawn

The clean mental model: the exclusion follows the purpose of the system, not the nature of the business. A defence company is a normal EU business that happens to also operate some exclusively-defence AI. Its exclusively-military systems are out under Article 2(3). Everything else — HR, marketing AI, civil products, dual-use anything — is in. National-security work commissioned and used exclusively for that purpose is excluded; the corporate back office that supports it is not.

A note on related law: AI used exclusively for national security sits with member-state competence, and dedicated EU defence and dual-use export regimes continue to apply on their own terms. Article 2(3) only governs whether the AI Act applies — it does not switch off other regulation.

Worked example: Falke Defence Systems OÜ

Falke Defence Systems OÜ, an Estonian-registered contractor with about 210 employees and roughly €58M turnover, builds a targeting-support subsystem sold exclusively to NATO-member defence ministries. That subsystem is out of scope under Article 2(3) — purpose exclusive, operator irrelevant.

But Falke's AI estate is not one system. It uses a vendor CV-screening tool to hire engineers: Annex III point 4 high-risk, fully in scope, Falke as deployer (Article 26). Its website runs a generative-AI sales chatbot: in scope under Article 50, which requires that visitors be told they are talking to AI. It also sells a dual-use perimeter-sensing product to civilian critical-infrastructure operators and to military bases — the dual use loses the exclusion, so the civil-market version is assessed under the Act. Falke's compliant position is precise: one excluded military system, three in-scope systems, and a documented per-system Article 2(3) test for each. "We're a defence company" is not a compliance answer; the per-use test is.

How Confir helps

Confir applies the Article 2(3) exclusion at the system level, not the company level: each AI system is tested for whether its use is exclusively military, defence or national-security — and if not, it is classified under Articles 5, 6 and 50 like any other. The deterministic, rule-based engine records why each system was excluded or kept in scope, which is exactly the documented per-use rationale a defence firm needs when an auditor asks how it drew the line. Same intake, same finding, the rule visible.

Related guides

• The scope of the EU AI Act
• Article 2: scope and exclusions
• Extraterritorial scope
• Recruitment screening AI: high-risk?
• Article 50 transparency duties
• Provider vs deployer