Piece-picking systems no longer handle only pre-taught items but mixed and unfamiliar ones, which means they learn from what they encounter. And they increasingly work in aisles alongside people rather than behind a guard.
Until July, a system of that kind faced two sets of European rules. Regulation (EU) 2023/1230 on machinery (the Machinery Regulation) governs the product itself and replaces Directive 2006/42/EC (the 2006 Directive). Regulation (EU) 2024/1689 on artificial intelligence (the AI Act) would have added its own requirements wherever the machine already needed a notified body. Neither set of high-risk rules has started to apply. That is not a reason to defer the question, because the systems designed now are the ones that will be placed on the market under them. On 27 July 2026, Regulation (EU) 2026/1744 (the Digital Omnibus) changed the relationship between the two.
What Changed on 27 July 2026
The Digital Omnibus moved the Machinery Regulation from Section A to Section B of Annex I to the AI Act. The classification is the same either way: AI that serves as a safety component of a listed product is high-risk where that product requires a third-party conformity assessment (AI Act, Article 6(1)). The consequence is not. For Section A products the AI Act’s own requirements for high-risk systems apply in addition to the sectoral ones. For Section B products they do not (AI Act, Article 2(2)). Risk management, data governance, technical documentation, human oversight and the provider obligations will therefore not reach AI in machinery through the AI Act.
The Commission must instead add health and safety requirements to Annex III of the Machinery Regulation that reflect those of the AI Act. The delegated acts are to apply by 2 August 2028 (Machinery Regulation, Article 8, as amended). The AI Act’s conformity assessment procedure is not among the requirements to be reproduced, because the Machinery Regulation prescribes its own, and neither are the provider obligations, which the Machinery Regulation covers through its duties on manufacturers. Recital 42 of the Omnibus describes the change as a sectoral approach, adopted to avoid a legal gap
The Machinery Regulation Was Already Written for Learning Machines
That approach is workable because the Machinery Regulation was not drafted for static machines. Its essential health and safety requirements already address systems that change their own behaviour (Machinery Regulation, Annex III). A control system that evolves its own behaviour must be built so that the machine cannot act beyond its defined task and movement space. Safety-related decisions have to be recorded and kept for a year. It must be possible to correct the machine at any time to keep it inherently safe. And the manufacturer’s risk assessment has to cover the hazards that are foreseeable as an intended evolution of that learning behaviour. None of this was borrowed from the AI Act. It has been in the Machinery Regulation since 2023.
The Question That Decides the Route
Under the Machinery Regulation, the assessment route depends on where the product sits in Annex I. Part A requires a notified body, without exception. Part B allows the manufacturer to assess the product itself, but only if it has applied the relevant harmonised standards in full. Anything not listed is self-assessed. All three routes are set out in Article 25 of that Regulation.
Part A covers safety components whose behaviour is self-evolving through machine learning, whether they are sold separately or built into the machine. Part B covers protective devices designed to detect the presence of persons, and logic units that ensure safety functions. In a cell that works alongside people, a vision system that stops the arm when someone enters the aisle is a candidate for those entries. The grasping model that decides where to close the gripper is not, at least in principle.
That determines whether a notified body is required, on the Machinery Regulation’s own terms. The AI Act builds on the outcome rather than shaping it: its high-risk classification presupposes that the product requires a third-party assessment in the first place (AI Act, Article 6(1)(b)). Where the Machinery Regulation calls for no notified body, no AI in that machine is high-risk on that basis, whatever it does. Where one is required, the AI Act applies a test of its own, and it is broader than purpose alone. AI used solely for efficiency, automation or convenience is excluded (AI Act, Article 6(1a)), but AI whose failure would endanger health and safety qualifies regardless (AI Act, Article 6(1b)). The Commission’s draft guidance takes the same view in an adjacent case: efficiency may well be the purpose, and the consequence of failure still decides. For a grasping model that can drop a load from height, that qualification is not academic.
The Dates, and What They Mean for Buyers
The move brought no date forward. It removed a layer that would have arrived in August 2028, and the delegated acts that replace that layer arrive by the same date. The Machinery Regulation still applies from 14 January 2027, with the requirements on learning behaviour already in it, as it did before the Omnibus. The relief is real in relation to the AI Act and irrelevant to the Machinery Regulation. Equipment placed on the market before that date stays under the 2006 Directive. For anything ordered now and delivered next year, where that line falls is a question for the purchase agreement.
The obligations sit with the manufacturer. A buyer can still ask which entry in Annex I to the Machinery Regulation this product was assigned to, and which assessment procedure follows from it. And, if it was self-assessed as a Part B product, which harmonised standards were applied and whether they were applied in full. The absence of a notified body does not mean the manufacturer treated the AI as safety-irrelevant. It may mean that the harmonised standards covered the relevant requirements.
A person who makes a substantial modification after installation becomes a manufacturer and must run the conformity assessment (Machinery Regulation, Article 18). But a change counts as substantial only where the manufacturer did not foresee it, where it affects safety, and where it calls for additional guards or protective measures. Since the manufacturer has to assess the intended evolution of learning behaviour anyway, retraining that stays within that assessment is foreseen. Having the assessment in writing before signing is the practical step.
For operators, the AI Act’s duties on deployers of high-risk systems will not reach this equipment at all (AI Act, Article 26). They will reach the rest of the AI in a warehouse: shift allocation, picker ranking, performance monitoring. The AI Act steps back from the robot. It does not step back from the building around it.
The practical question in 2027 will not be whether the AI Act or the Machinery Regulation applies. That one now has an answer, and it is the easier of the two. What remains is where in the system the safety function sits, and whether that placement was a decision or an accident. That depends on which model receives the safety-rated sensor’s output and may act on it, which one only sees the tote, and whether the two are kept apart at all. It is settled early in development, not in a compliance review before shipping. In July the legislature disapplied the AI Act’s high-risk requirements for these products. What replaces them is still being drafted. The decisions that determine which of them apply are being taken now.
