Article 33 of Machinery Regulation 2023/1230: why machinery with safety-related AI falls into the high-risk category

Direct answer: Article 33 of Regulation (EU) 2023/1230 requires a conformity assessment procedure involving a notified body for any machine whose safety system incorporates AI-based software with evolving behaviour — that is, software that learns or modifies its own decision logic after being placed on the market. This applies even when the equivalent non-AI machine — with the same mechanical function — does not require third-party intervention, because the risk the article regulates is not the machine's physical movement but the unpredictability of a safety system that changes over time. In this article we explain which machines are affected, why the EU legislator treated AI as a risk-category jump, and what this means in practice for manufacturers of autonomous mobile robotics and critical decision systems.

Why is a machine with AI high-risk even when its mechanical version is not?

Direct answer: Regulation 2023/1230 shifts the risk criterion from the machine's physical function to the reliability and predictability of its decision system. Under the former Directive 2006/42/EC, risk classification depended almost exclusively on the mechanical nature of the machine — a saw, a press, a lift — and on whether it appeared on the former Annex IV list. A material-handling machine controlled by a conventional PLC, with fixed, verifiable logic, could in many cases be self-certified. Regulation 2023/1230 introduces a new criterion: when the safety function depends on an AI component that learns, recalibrates or modifies its behaviour after being put into service — without human intervention for each change — the traceability of that decision is no longer static and therefore can no longer be verified using traditional documentary audit methods.

The EU legislator starts from a technical premise: a safety system based on fixed rules can be verified once, at the point of conformity assessment, because its future behaviour is deterministic. An AI system with continuous learning — for example, a neural network that adjusts its obstacle-detection thresholds based on experience accumulated on the shop floor — may behave differently six months after installation, even though the manufacturer has not introduced any conscious design change. That behavioural drift (“model drift”, in technical terminology) is precisely what Article 33 aims to subject to external control, because manufacturer self-certification no longer offers sufficient guarantees that the original risk analysis remains valid over time.

What types of machinery fall within the scope of Article 33?

Annex I, Part A does not require the entire machine to run on AI: it is enough for the safety function to depend on an AI component with evolving behaviour. This extends the scope far beyond pure factory robotics.

How does this affect the relationship between the Machinery Regulation and the AI Act?

Article 33 cannot be read in isolation from Regulation (EU) 2024/1689 on Artificial Intelligence (the AI Act), which classifies as a “high-risk AI system” — Annex III, point 2 — any safety component of a product already covered by the Machinery Regulation. The relationship between the two rules is complementary: the AI Act requires a risk-management system, quality training data and human oversight for the AI component itself, while Article 33 of the Machinery Regulation requires that its integration into the machine be assessed by a notified body before being placed on the market. In practice, a manufacturer of an AMR with safety-related AI must satisfy both regimes in a coordinated manner, and the notified body assessing the machine will also need to verify the AI system's conformity documentation under the AI Act.

→ If your technical department manages conformity files for machines with AI and you need to track which articles of Regulation 2023/1230 apply to each component, try IgeraIndustria free for 14 days and check the exact requirement, with the corresponding article cited, in seconds.

Case study: the AMR manufacturer who underestimated the change in procedure

A logistics robotics company in the Valles Occidental area — which we will call, anonymised, RoboFlux Systems — developed an autonomous mobile robot for warehouses shared with human staff, with a predictive braking system trained through reinforcement learning on real shop-floor data. Under the former Directive 2006/42/EC, a fixed-path AGV with the same transport function could be self-certified without issue. RoboFlux's engineering team initially planned the launch assuming the same procedure, until its regulatory compliance department detected that the predictive braking system, by adjusting its parameters autonomously after deployment, fell squarely within Annex I, Part A of the new Regulation. This forced an EU-type examination with a notified body to be added to the schedule, with an impact of almost four months on the launch date that had not been anticipated in the original project plan. If your engineering team is assessing whether a common mistake is failing to distinguish between “fixed automation” and “evolving behaviour” when classifying a product, this is exactly the point where it pays to consult before setting the certification timeline.

Frequently asked questions about Article 33

Does every machine that uses artificial intelligence fall under Article 33?

No. Article 33 applies specifically when AI ensures a safety function with behaviour that evolves after being put into service. A machine that uses AI only to optimise production, for non-critical predictive maintenance, or for business analytics, without that component being involved in the safety of persons, is not automatically subject to this reinforced procedure.

What exactly does “evolving behaviour” mean in the text of the Regulation?

It refers to systems whose behaviour changes after being placed on the market without a redesign intervention by the manufacturer, typically through continuous learning, parameter updates using field data, or periodic retraining. An AI model that is trained once and then frozen (with no subsequent learning) is closer to the treatment of a traditional control system, although the notified body must assess this on a case-by-case basis.

Which notified bodies can assess machinery with safety-related AI?

They must be notified bodies designated under Article 27 of Regulation 2023/1230 with the specific technical scope for AI-based safety systems, since the assessment requires competencies in machine learning in addition to traditional machine-safety expertise. Not all currently notified bodies have this extended scope, so it is worth verifying before starting the procedure.

How does Article 33 relate to the AI Act?

The AI Act classifies as high-risk, in its Annex III, AI systems that are safety components of products already regulated under EU harmonisation legislation, including the Machinery Regulation. In practice, the manufacturer must simultaneously meet the AI Act's risk-management and data-governance requirements and the Article 33 conformity assessment procedure, and the notified body will review both sets of evidence in a coordinated manner.

Is it mandatory to monitor the AI model after it is placed on the market?

Regulation 2023/1230 requires, through the manufacturer's general post-market surveillance obligations (Article 10), monitoring of the product's behaviour once in use. For AI components with continuous learning, this translates in practice into monitoring model drift and documenting any significant behavioural change that could affect the original risk assessment.

Since when is this requirement enforceable?

Regulation 2023/1230 applies in full from 20 January 2027, from which date any new machine placed on the market with a safety function based on AI with evolving behaviour must comply with the procedure under Article 33. For machines already lawfully placed on the market under Directive 2006/42/EC before that date, the transitional provisions of Article 52 apply.