AI Omnibus Strips Machinery from High-Risk AI Rules — What It Means for European Workers

What Actually Changed

To understand what happened, you need to know how EU law handles AI systems embedded in physical products. Annex I of the AI Act originally listed products subject to pre-existing harmonisation legislation — including the Machinery Regulation (EU) 2023/1230 — that would automatically fall under the AI Act's high-risk category if they contained an AI safety component. Think of an AI-powered emergency braking system in a forklift, collision avoidance in a robot arm, or pedestrian detection in an autonomous guided vehicle. Under the original AI Act, these were classified as high-risk and subject to requirements on data governance, bias mitigation, cybersecurity, traceability, and human oversight.

The Digital Omnibus on AI — proposed in November 2025, politically agreed in May 2026, and in force since 27 July 2026 — changed that. Machinery products were removed from Annex I, Section A. The official reasoning was to "avoid duplication" between the Machinery Regulation and the AI Act. In practice, it means AI safety components in machinery are no longer subject to the AI Act's directly applicable high-risk obligations until replacement rules are written into the Machinery Regulation itself.

The European Commission now has until August 2028 to incorporate those AI-specific safeguards into the Machinery Regulation through a delegated act — a fast-track legislative procedure that avoids reopening the entire regulation.

What's at Stake: The Safety Component

The term "safety component" is central to this debate. Under the AI Act, an AI system qualifies as high-risk if it is intended to serve as a safety component of a product covered by the listed legislation. The European Trade Union Confederation (ETUC) warns that any further narrowing of what counts as a safety component could push critical systems — AI-supported braking, collision avoidance, pedestrian recognition — outside the high-risk category entirely. If that happens, products containing those systems could reach the market without undergoing the AI Act's conformity assessment procedures.

Claes-Mikael Ståhl, deputy general secretary of the ETUC, and Giulio Romani, ETUC confederal secretary, put it bluntly in their analysis: "When new technology creates new risks at work, responsibility must remain with those who design, manufacture, and deploy it, not be transferred to the workers operating alongside it."

Standards Are Not Law

During the transition period, manufacturers can rely on AI Act standards developed by standardisation bodies like CEN/CENELEC. This offers a bridge between the two regulatory regimes. But here's the catch: standards cannot create legal protections that the legislation does not guarantee. A harmonised standard can tell a manufacturer how to build a compliant system; it cannot invent a requirement that the law itself never imposed.

Think of it this way: if the Machinery Regulation doesn't explicitly require bias testing for an AI vision system that detects pedestrians in a warehouse, a standard cannot magically add that obligation. The gap is legal, not just technical.

Why This Matters for Developers and Deployers

If you're a developer building computer vision systems for industrial machinery, or a company deploying AI-powered safety equipment on a factory floor, this uncertainty affects you in concrete ways:

Compliance planning is frozen. You don't know which specific AI requirements will make it into the Machinery Regulation. Should you invest in bias audits, logging infrastructure, and human-oversight mechanisms now, or wait? If you build to the original AI Act standard and the final Machinery Regulation rules are lighter, you've spent money you didn't need to. If you wait and the rules are equally strict, you'll be scrambling at the deadline.

Liability is ambiguous. If an AI safety component fails during this transition period — say, an vision-based emergency stop system doesn't detect a worker — which regulatory framework governs the investigation? The AI Act's traceability and logging requirements no longer apply directly. The Machinery Regulation, as it stands today, lacks equivalent obligations. This is not a theoretical concern; it's the kind of question that ends up in court.

Cross-border sales get complicated. Different EU member states may interpret the transition differently, leading to a fragmented market where the same product faces different expectations in Germany versus Poland. For a sector that thrives on standardisation, that's a problem.

The Timeline: A Two-Year Window

Here's the regulatory timeline in plain terms:

• August 2024: AI Act enters into force.
• November 2025: European Commission proposes the AI Omnibus, aiming to simplify rules.
• May 2026: Political agreement reached on the AI Omnibus. Machinery removed from Annex I.
• 27 July 2026: AI Omnibus enters into force. Machinery AI safety components lose direct high-risk status under the AI Act.
• 2 August 2026: AI Act itself becomes applicable (general provisions).
• August 2028: Deadline for the Commission to transfer AI safeguards into the Machinery Regulation via delegated act.
• 2 August 2028: Original deadline for high-risk AI systems in Annex I products — but now uncertain because the Annex I machinery entry was deleted.

The gap is the period between July 2026 and August 2028. Two years without clear, legally binding AI-specific obligations for safety-critical machinery. Two years during which new products with AI safety components can enter the market while the rules that should govern them are still being drafted.

The European Dimension

This is not an abstract regulatory debate. Europe has a large machinery industry — from German automotive manufacturing to Italian packaging equipment to Czech industrial automation. These sectors are increasingly integrating AI. The question is whether the regulatory framework keeps pace.

The AI Act became fully enforceable on 2 August 2026. Transparency obligations — such as labelling AI-generated content — also kicked in on that date. But the machinery carve-out means a significant category of high-risk AI systems now operates under transitional rules while workers and their representatives watch the clock.

The ETUC's position is clear: simplification should not become a cover for dilution. The AI Act already had mechanisms to avoid regulatory duplication; the deletion from Annex I created a problem that the delegated act must now solve.

What Should Happen Next

The Commission needs to deliver three things before the 2028 deadline:

1. A detailed mapping of every AI Act high-risk requirement and exactly how it will be reflected in the Machinery Regulation — not a vague commitment, but a line-by-line correspondence.
2. A clear timeline with milestones, so manufacturers and trade unions can track progress rather than waiting until the last minute.
3. No narrowing of the safety component definition. As AI becomes more deeply embedded in obstacle detection, collision avoidance, and other safety functions, the definition must capture the full scope of systems that protect workers.

For the European Parliament and the Council, the job is oversight: scrutinise the delegated act, ensure it provides genuinely equivalent protection, and reject any version that waters down the AI Act's original safeguards.

Our Take

As someone who runs production AI systems — pipelines that generate articles, transcribe audio, and benchmark language models — I pay close attention to how regulation shapes the software we build. A regulatory gap may sound abstract until you're the developer wondering whether your computer vision module needs a full conformity assessment or can ship with lighter documentation. That uncertainty has a real cost: it slows deployment, increases legal risk, and — in the worst case — puts workers in the path of systems whose safety hasn't been verified under clear, enforceable rules.

The AI Omnibus was sold as simplification. Simplification is fine — nobody wants duplicative paperwork. But when the machinery that protects factory workers, warehouse operators, and construction crews loses its direct high-risk status under the AI Act, simplification has crossed into something else. The Commission has two years to prove it hasn't.