The EU AI Act and Machinery Regulation: What They Mean for Physical Robots

Do robots have to comply with both the EU AI Act and the Machinery Regulation?

Yes — and that’s precisely what makes compliance for physical robots harder than for a chatbot or a recommendation engine. A robot sold in the EU is a physical product with moving parts, so it has always had to comply with machinery safety law. Once that robot runs a machine-learning model to make safety-relevant decisions — say, adjusting force in real time around a human co-worker — it also falls under the EU AI Act, the bloc’s horizontal law on artificial intelligence. Two regulations, two sets of obligations, one product on the factory floor.

The AI Act entered into force on 1 August 2024. The Machinery Regulation (EU) 2023/1230 applies from 20 January 2027, replacing the long-standing Machinery Directive 2006/42/EC with no grace period — any machinery placed on the market from that date must meet the new rules in full. For a robotics manufacturer, the practical question isn’t “which law applies” but “how do the two overlap, and which deadline actually matters right now.” That’s what this article works through, including a legislative change from earlier in 2026 that shifted some of the dates other compliance guides still cite.

What is the EU Machinery Regulation, and how is it different from the old Directive?

The Machinery Regulation replaces a directive — which each EU member state had to transpose into national law, sometimes with local variation — with a regulation that applies uniformly and directly across all member states. That alone tightens the compliance landscape for anyone selling robots into multiple EU markets. But the more consequential change is what the Regulation adds specifically for AI-enabled machinery.

Annex I, Part A of the Regulation introduces two new categories of high-risk machinery tied directly to machine learning:

• Item 5 — safety components with fully or partially self-evolving behaviour that use machine-learning approaches to ensure safety functions.
• Item 6 — machinery that embeds those self-evolving safety components.

Before this Regulation, most machinery manufacturers could self-certify conformity — sign a declaration, apply the CE mark, and place the product on the market. Under items 5 and 6, that option disappears. A robot whose safety functions rely on a self-evolving ML model must go through mandatory third-party conformity assessment by a Notified Body — an independent organisation designated by an EU member state to test and certify products against the relevant standard — before it can carry the CE mark.

Which robots actually count as “high-risk” under the AI Act?

Not every robot with an AI component is automatically high-risk. The AI Act’s high-risk categories are defined by use case, not by the mere presence of a model. A robot becomes high-risk under the AI Act primarily through one of two routes:

1. It is itself a safety component of a product already regulated under EU product-safety law (Annex I of the AI Act references those laws, including the Machinery Regulation), and that product must undergo third-party conformity assessment.
2. It performs a function listed in Annex III — biometric identification, safety components in critical infrastructure, and similar sensitive applications.

This is exactly where the AI Act and the Machinery Regulation intersect for robotics. A robot arm whose collision-avoidance system uses a self-evolving ML model is a high-risk AI system under the AI Act specifically because it is a safety component requiring Notified Body assessment under Machinery Regulation Annex I. A conventional industrial robot running rule-based (non-learning) motion control, by contrast, may never trigger the AI Act’s high-risk provisions at all — the Machinery Regulation still applies, but the AI Act’s heaviest obligations don’t kick in.

AI Act vs. Machinery Regulation: a side-by-side view

	EU AI Act	EU Machinery Regulation 2023/1230
Entered into force	1 August 2024	Published 2023, applies from 20 January 2027
Replaces	New horizontal framework (no prior EU-wide AI law)	Machinery Directive 2006/42/EC
Scope	AI systems across all sectors, risk-tiered	Physical machinery and safety components sold in the EU
Trigger for strictest rules	“High-risk” classification (Annex III use cases, or safety components in Annex I regulated products)	Annex I, Part A items 5 & 6: ML-based self-evolving safety components
Conformity route	Self-assessment for most high-risk systems; third-party review for select cases	Mandatory Notified Body assessment for AI-enabled safety components (no self-certification)
Legal form	Regulation (direct effect, no national transposition)	Regulation (direct effect, no national transposition)
Key 2026 development	Digital Omnibus on AI deferred high-risk deadlines	Deadline itself unchanged; interacts with AI Act via Annex I product link

What changed with the 2026 Digital Omnibus on AI?

Here’s where a lot of older compliance guides now read as outdated, and it’s worth being precise about the timeline rather than picking one date arbitrarily. The AI Act as adopted in 2024 originally set full high-risk applicability for 2 August 2026 (Annex III stand-alone systems) and 2 August 2027 (Annex I regulated products, which is where most robots and machinery sit). Many articles published before mid-2026 still cite those dates, and they weren’t wrong at the time.

That changed on 19 November 2025, when the European Commission published the Digital Omnibus on AI — a package of targeted amendments proposing to defer several high-risk compliance deadlines. On 7 May 2026, the Council of the EU and the European Parliament reached a provisional political agreement on that package, the first amendment to the AI Act since its original adoption. Under the agreed text, high-risk obligations are deferred along two tracks: stand-alone Annex III systems (employment, biometrics, critical infrastructure, education, migration) now apply from 2 December 2027, while AI embedded in Annex I regulated products — machinery, toys, lifts, medical devices, watercraft, and by extension most robots — now applies from 2 August 2028.

The package cleared its final legislative hurdles in June 2026: the European Parliament formally adopted it on 16 June 2026 (423 in favour, 57 against, 174 abstentions), and the Council of the EU gave its final green light on 29 June 2026. So as of this writing, 2 August 2028 is the operative date for high-risk AI Act obligations tied to machinery — not 2027. Robotics manufacturers relying on guidance written before spring 2026 should treat the earlier dates as historical rather than current.

Crucially, this deferral applies to the AI Act’s high-risk obligations. It does not touch the Machinery Regulation’s own timeline: 20 January 2027 remains the date machinery placed on the market must meet EU 2023/1230, including the Notified Body requirement for items 5 and 6. That gap — Machinery Regulation live from January 2027, AI Act high-risk product obligations phased in through August 2028 — is the detail most worth flagging to a compliance team, since it means the machinery-safety assessment can effectively become the earlier practical gate even before every AI Act obligation bites.

How do the two regimes actually interact on a single robot?

Think of it less as two parallel checklists and more as one process with two lenses. The Machinery Regulation asks: is this robot safe to operate near people, and has an appropriate body verified that? The AI Act asks: is the AI system inside it transparent, robust, and subject to appropriate human oversight, given what it does?

For a robot with a self-evolving ML-based safety function, both lenses converge on the same event — the Notified Body assessment. That body will need to evaluate not just mechanical and electrical safety in the traditional sense, but also the behaviour and robustness of the learning component itself: how it was trained, how its outputs are bounded, and how failure modes are handled. This is a meaningfully different audit than the paperwork-based self-declaration that sufficed under the old Machinery Directive for most products, and it’s part of why standards bodies and Notified Bodies are still working through what “conformity assessment for a self-evolving safety component” concretely looks like in practice. Aligning early with established mechanical and functional-safety benchmarks — the kind of layered approach described in our overview of industrial robot safety standards — gives manufacturers a head start, since the AI-specific assessment builds on top of, rather than replaces, that baseline.

What should robotics manufacturers do to prepare now?

Given the two-track timeline, the practical sequencing looks like this:

• Map every robot’s safety architecture against Annex I items 5 and 6 now. If any safety function uses a self-evolving ML approach, that product needs a Notified Body relationship well before January 2027 — assessment slots and documentation review take time, and waiting until the deadline year is the single most common mistake flagged by compliance advisers.
• Separate “self-evolving” from “static” ML use. A model trained offline and then frozen before deployment is a materially different regulatory case than one that continues to adapt in the field. Documenting this distinction clearly can determine whether Annex I items 5/6 even apply.
• Track the Digital Omnibus implementation guidance, not just the headline date. 2 August 2028 is now the reference point for AI Act high-risk obligations on Annex I products, but implementing acts and harmonised standards will keep arriving between now and then.
• Treat CE marking as a single, coordinated exercise, not two separate submissions — a robot that needs both Machinery Regulation and AI Act sign-off benefits from a technical file built to satisfy both from the start, rather than retrofitting one onto the other.

None of this is unique to any one robot form factor — the same overlap applies whether the platform is a fixed industrial arm, a mobile warehouse robot, or a humanoid platform, and it sits inside the broader shift in how capital and regulators are treating the physical AI market as it scales. Manufacturers who build compliance into their engineering roadmap now, rather than at the deadline, will be the ones shipping into the EU without disruption when both regimes are fully in force.

For the wider regulatory and investment context this compliance timeline sits inside, see the Market, Investment & Regulation hub; definitions for terms used throughout this piece are collected in the glossary.