Industrial AI attaches to the EU AI Act by a different route from software, and the difference matters more than the content of the obligations. If your AI is a safety component of a product already regulated under EU harmonisation law - machinery, lifts, pressure equipment, radio equipment, medical devices - you are in Annex I, not Annex III. That changes your deadline by eight months and folds the AI Act into a conformity assessment you already run.
Annex I obligations were deferred to August 2028 by the June 2026 omnibus. Annex III is 2 December 2027. Manufacturers reading generic AI Act guidance are usually reading the wrong date for themselves.
Which route applies
| System | Route | Deadline |
|---|---|---|
| AI as a safety component of machinery, lifts, pressure equipment, RED products | Annex I - folds into existing CE conformity assessment | Aug 2028 |
| AI as a safety component of digital infrastructure, road traffic, water, gas, heating, electricity supply | Annex III point 2 - critical infrastructure | Dec 2027 |
| Predictive maintenance on your own plant | Minimal - machines, not people | Art 4 only |
| Visual quality inspection | Minimal - inspecting output, not persons | Art 4 only |
| Demand forecasting, scheduling, supply-chain optimisation | Minimal | Art 4 only |
| Collaborative robot safety systems | Annex I via the Machinery Regulation | Aug 2028 |
| Worker productivity or behaviour monitoring | Annex III point 4 - employment | Dec 2027 |
| Emotion or fatigue inference on workers | Prohibited - Article 5 | Now |
The safety-component test
A safety component is one whose failure endangers health and safety of persons or property. A vision model that stops a press when a hand enters the guarded zone is a safety component. The same model counting defective parts on the same line is not. Identical technology, different intended purpose, different regulatory route - which is why the Act keeps insisting classification follows intended purpose.
Where you are in Annex I, Article 8(2) is designed so the AI Act's requirements are met through the existing sectoral conformity assessment rather than a parallel one. Your quality management system, technical file, risk assessment and post-market surveillance largely absorb the equivalents. The genuinely new work is usually Article 10 data governance, Article 12 logging, and Article 14 human oversight design.
The exposure most plants actually have
Not the robots. The workforce monitoring.
Productivity scoring, behaviour analytics, camera-based "safety compliance" monitoring that scores individual workers, fatigue detection - these are Annex III point 4 employment systems, and any of them that infer emotional or psychological state cross into Article 5 prohibition, which is live now and cannot be cured by consent, works-council agreement or documentation.
Fatigue and drowsiness detection is the sharp case. Where it infers an operator's internal state in a workplace context, it is emotion inference and prohibited. Where it detects a physical event - eyes closed, head position, hands off controls - the analysis is different. Do not assume; the vendor's marketing language is not the classification.
Deployer duties if you buy rather than build
A manufacturer buying an AI-equipped machine is a deployer under Article 26: operate it per instructions, assign competent human oversight, monitor, keep logs, and inform workers where an Annex III system is used in the workplace. You do not inherit the machine builder's Chapter III duties. The line - and note that substantial modification, including retraining a vision model on your own line data, can make you the provider.
Live today, whatever your route
- Article 4 - recorded AI literacy for operators, maintenance and EHS staff
- Article 5 - audit any workforce monitoring for emotion or state inference
- Article 50 - if any interface converses with people, it discloses
- A dated inventory of AI systems on the plant, their intended purpose, and which route each takes
Annex I or Annex III?
The route decides your deadline and your assessment path. The free check settles which one each system is on.
Start the free scanNo card, no account. Full gap report from €59. See a real report first.
Self-assessment, not legal advice. This page and the Reglynn report are a structured self-assessment based on the EU AI Act (Regulation 2024/1689, as amended by the June 2026 Omnibus). They indicate where you likely stand and what to verify with a qualified advisor. Reglynn is not a notified body and issues no certification.
How this was made. Reglynn reports are generated by an AI system built on Anthropic's Claude and reviewed before delivery. Reglynn is a Limited Risk AI system under Article 50; we disclose AI involvement wherever it occurs, including in our chat assistant. Our own self-assessment.