The Machinery Directive is going. What governs the AI inside

A manufacturer building AI into a machine has had a question with no settled answer. Is the matter decided by the Machinery Directive and its successor, by the AI rules, or by both at once? Until July 2026 the answer was both. Anyone building a self-learning safety function into a machine had to reckon with the machinery requirements and the EU AI Act's requirements stacked on top of each other, assessed inside a single procedure.

That answer no longer holds. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and entered into force three days later. It amends three legal acts at once, and for anyone placing machinery on the EU market the change that matters is a move between two sections of one annex.

The Machinery Directive runs until January 2027

Directive 2006/42/EC is still the instrument in force. It is repealed by the Machinery Regulation (EU) 2023/1230, and Arbetsmiljöverket, the Swedish work environment authority, states that the regulation takes full effect on 20 January 2027, replacing the requirements in that authority's own machinery rules at the same time.

That calendar is worth keeping apart from the AI calendar, because the two do not meet until much later. A machine placed on the market during 2027 is CE marked under the Machinery Regulation, and nothing in the AI rulebook changes that procedure.

AI in machinery no longer sits under the EU AI Act's requirements

Annex I to the EU AI Act is split into two sections. Section A gathers the product legislation whose products pull in the entire high-risk regime: the requirements in Chapter III Section 2, the conformity assessment in Article 43, the notified bodies, market surveillance and the penalties. Section B gathers the legislation that largely handles the question in its own machinery, such as civil aviation and rail interoperability.

Machinery sat in Section A, through a reference to the Machinery Directive. The amending regulation deletes that point from Section A and instead adds the Machinery Regulation (EU) 2023/1230 as a new point in Section B.

What survives of the EU AI Act is a classification rule, a testing rule, and the final provisions.

The consequence sits in the amended Article 2(2). For high-risk AI systems related to products covered by Section B, only Article 6(1), Article 60a and Articles 102 to 112 apply. The requirements in Chapter III Section 2 fall away. Article 43 on conformity assessment falls away. The market surveillance chapter falls away, and so do the administrative fines in Article 99. Articles 57, 58 and 59 on regulatory sandboxes apply only to the extent the high-risk requirements have been written into the sectoral legislation, which they have not been yet.

That does not mean nobody is watching and nothing can be penalised. The Machinery Regulation carries its own market surveillance and its own penalty provisions, and that is the route that applies to a machine with AI in it.

One job remains: naming which machines count

Article 6(1) survives, and it starts to apply on 2 August 2028. It says an AI system is high-risk where it is a safety component of a product covered by the legislation in Annex I, or is itself such a product, and where that product has to undergo a third-party conformity assessment. Both conditions have to be met.

What the classification now settles is not which requirements apply, but which machines the coming requirements will reach. The amended Article 8 of the Machinery Regulation obliges the Commission to add health and safety requirements for precisely those AI systems classified as high-risk under Article 6(1) because they are a safety component in a machine, or are a machine themselves.

The classification rule stayed behind as an address book after the rest of the regime moved out.

The requirements themselves are not written yet

The delegated acts are to reflect Chapter III Section 2 together with Articles 17, 19, 72 and 73 of the EU AI Act: the quality management system, the documentation, post-market monitoring and the reporting of serious incidents. They must apply by 2 August 2028 at the latest. The Commission's power to adopt them runs for five years from 27 July 2026.

Until then there is a transitional rule. A new paragraph in Article 20 of the Machinery Regulation lets a manufacturer applying harmonised standards or common specifications adopted under the EU AI Act be presumed to meet the coming AI requirements in Annex III to the Machinery Regulation. It is a useful rule resting on uncertain ground. The amending regulation's own second recital describes the drafting of those standards as delayed, and names that delay as one of the reasons for the whole amendment.

Machinery is not alone in being simplified. The same amending regulation empowers the Commission to limit Articles 9 to 15 and 17 to 25, by delegated acts due no later than 2 August 2027, for the products that remain in Section A, where the sectoral legislation there sets requirements giving an equivalent or higher level of protection. A manufacturer of a different Annex I product has reason to follow that question, even though the outcome there is not settled.

In Sweden, the file is read by Arbetsmiljöverket

Which national body carries out market surveillance is a matter of national designation, and in Sweden machinery sits with Arbetsmiljöverket. The list of responsibilities published by Marknadskontrollrådet, the Swedish market surveillance council, puts machinery, personal protective equipment for occupational use, work platforms, ladders and trestles, mine hoists and guard rails on that authority's row.

On 12 June 2026 the Swedish government named five authorities as national competent authorities under the EU AI Act, in decision Fi2026/01365: Post- och telestyrelsen, Integritetsskyddsmyndigheten, Finansinspektionen, Läkemedelsverket and Swedac. Arbetsmiljöverket is not among them, and does not need to be, because the market surveillance chapter of the EU AI Act no longer reaches AI inside machinery.

The answer would have been the same by the old route. Article 74(3) already pointed to the authority designated under the sectoral legislation as market surveillance authority for Section A products. The difference is that the route no longer has to be taken, and that market surveillance of AI inside a machine cannot be split between two authorities.

Anyone who built against the EU AI Act did not build wrong

Here sits the strongest objection to everything above. The amending regulation's recitals state that the Commission is to secure a level of protection consistent with the EU AI Act, and the transitional rule in Article 20 of the Machinery Regulation points back at the same standards. A manufacturer who has already built risk management, data governance and technical documentation against Chapter III Section 2 has done the right work and gets to keep it. Read that way, the move is bookkeeping.

Equivalence that so far exists only in a recital is an intention, not a text to document against.

The difference lies in what can be shown today. A delegated act that must apply by 2 August 2028 has not been adopted, and it will take a different form from the EU AI Act's articles: essential health and safety requirements in an annex rather than obligations in a chapter. The work carries over; the format does not. That argues for doing the work now and rereading the text when it lands, not for waiting on the text. The same pattern runs through the high-risk deadlines generally, which moved without the neighbouring duties moving with them.

The answer is in your own statement of intended purpose

The amending regulation tightened the definition of a safety component. A component serves a safety function where its intended purpose is to prevent or reduce risks to the health and safety of people or to property. What intended purpose means is not in that sentence but in the regulation's list of definitions: the use for which the provider intends the system. The starting point is therefore your own description.

A new paragraph in Article 6 adds the converse. AI used solely for non-safety-related aspects of user support, performance optimisation, service efficiency, automation, convenience or quality control is not a safety component. That is the same line that sits underneath the rules for AI in manufacturing, and it is where most systems in a workshop land. Reading what the machine logs already contain is one example: it can be done before the integration is built and it controls nothing on the line.

Then comes the paragraph after that, the one sentence in the whole amendment a manufacturer cannot write its way out of. An AI system whose failure or malfunction would endanger health and safety is to be treated as a safety component, whatever the preceding paragraph says.

The criterion is therefore in two parts, and it gives different readers different answers. First: what does your own statement of the system's intended purpose say? Second: what happens if the system stops working mid-run? Two manufacturers with identical technology in the machine can get different answers to the first question and the same answer to the second. Anyone who answers quality control to the first and danger to the operator to the second already has the verdict, and that assessment belongs in the same inventory and classification as everything else in AI governance that is meant to hold.