EU Machinery Regulation: What AI in Machines Needs

EU Machinery Regulation: The AI Clause Machine Builders Missed

The EU Machinery Regulation replaces a directive that governed machine safety across Europe for decades, and it becomes mandatory on 20 January 2027. Most of the coverage focuses on the change in legal form, since a regulation applies directly rather than through national implementations. For anyone putting intelligence into machines, the more consequential change sits in the annexes. Components that keep learning after they leave the factory, and machines that lean on a model to keep people safe, now sit in the annex reserved for the highest-risk products, which closes off the self-assessment route most manufacturers rely on today.

Directive Then, Regulation Now

The familiar obligations remain. What changes is how far they reach.

What the EU Machinery Regulation Changes

Three shifts matter to a builder who is adding intelligence to a product. First, software that performs a safety function is now unambiguously a safety component, which brings it inside the conformity assessment rather than leaving it as an engineering detail. Second, the high-risk annex adds categories that did not exist before, aimed squarely at learning systems and AI-driven safety functions. Third, cybersecurity enters the essential requirements, meaning control logic has to hold up against foreseeable malicious interference rather than only against faults.

The practical consequence of appearing on the high-risk list is procedural. Self-assessment is no longer available for those products, so a notified body has to be involved in the conformity route. Notified body capacity is finite, designations under the new regulation have been rolling out since 2024, and the deadline is the same for everyone. Booking that engagement late is the most avoidable risk in this entire transition.

Where the EU Machinery Regulation Meets the AI Act

Machinery is also covered by the broader EU AI Act framework, and the two regimes are designed to interlock rather than duplicate. Put simply, when a model carries a safety function inside a machine that already needs third-party assessment, obligations arrive from both directions at once. The sensible approach is one technical file that satisfies both rather than two parallel projects, and one risk assessment that treats the model as part of the machine rather than as a feature bolted onto it.

Self-Evolving Behaviour Is the Line

The distinction regulators drew is between a system that behaves the same way after deployment and one that does not. A vision model trained once, validated, frozen, and shipped is a component whose behaviour can be characterised and tested. A system that continues to learn from operating data, and therefore may act differently next month, cannot be validated once in the same way.

That has a design implication worth deciding deliberately rather than by default. Many manufacturers will conclude that continuous learning inside a safety function is not worth the conformity burden, and will freeze models between validated releases, keeping adaptation in non-safety functions such as quality analytics or maintenance prediction. That is a legitimate architecture, and it is far easier to evidence.

Evidence is the word worth dwelling on. For a frozen model, a safety case can point at a fixed artefact, a defined dataset, and test results that stay true for that version. For a system that adapts in service, the case has to cover a range of possible future behaviours, which usually means monitoring, bounded adaptation, and a defined trigger for revalidation. That is not impossible. It is simply a much larger programme than most machine builders have scoped.

Six Steps Before January 2027

These steps are sequential, and the first one decides how much work the rest represents. Start with the portfolio review even if the engineering answer feels obvious, because classification disputes surface late and cost the most.

1. Classify your products against the new annexes. Establish whether anything you sell falls into the AI-related high-risk categories.
2. Book notified body capacity early. Availability, not engineering, is the binding constraint for many builders.
3. Decide where learning is allowed. Separate adaptive analytics from safety functions in the architecture.
4. Extend the risk assessment to the model. Include data quality, drift, sensor degradation, and failure behaviour.
5. Add cybersecurity evidence. Show that control logic resists foreseeable tampering and corrupted inputs.
6. Update instructions and documentation. Confirm digital delivery meets the conditions and languages required.

Pros and Cons of Putting AI in a Safety Function

There are real gains available, and a real compliance cost attached to them.

• Pro: better detection than fixed logic. Vision systems can recognise situations that sensors and interlocks cannot express.
• Pro: fewer nuisance stops. Context-aware systems reduce false trips that lead operators to defeat safeguards.
• Pro: differentiation. Certified AI safety functions are hard to copy and support premium positioning.
• Con: third-party assessment. Self-declaration disappears, adding cost, lead time, and scheduling risk.
• Con: validation of a moving target. Anything that keeps learning needs evidence that it stays safe as it changes.
• Con: dual regulatory load. Machinery and AI obligations both apply, so documentation has to satisfy both.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how the EU Machinery Regulation lands in practice rather than presented as verified case studies.

A robotics integrator ships cells with a vision-based safeguarding function trained by the customer during commissioning. Under the new rules the customer-side training may make the behaviour adaptive after placing on the market, which changes both the conformity route and who carries the manufacturer obligations. The fix is contractual and technical, and it needs deciding long before delivery.

A machine builder assumes its AI feature is exempt because it only advises the operator rather than stopping the machine. On review, an interlock elsewhere in the control logic acts on that advice, which makes the model part of a safety function after all. Classification turns on how the system behaves, not on how the feature was described in the brochure.

A component supplier freezes its model between releases and publishes validation evidence with each version. Customers can point at it in their own technical files, and the supplier turns a compliance chore into a sales argument, which is the outcome worth aiming for.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Cybersecurity Is Now Machine Safety

Treating malicious interference as a safety matter is new for many machine builders, and it changes the evidence expected. Control logic has to withstand foreseeable attempts to corrupt it, which for an AI-driven function includes inputs designed to mislead a model. Manipulated markings, adversarial patterns, or simply a camera obscured in a specific way can all move a perception system into the wrong decision.

The discipline resembles what software teams already apply to untrusted input, and it connects to the broader problem of models acting on content they were never meant to trust, which we covered in indirect prompt injection. On a factory floor the consequence is physical rather than reputational, which raises the standard of proof.

Why Talkory Wins

Most of the early work here is interpretive. Does this function count as a safety function, does this product fall inside the annex, does customer-side training change the obligation. Talkory puts the same question and the relevant text in front of GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 at once. Where all six read a requirement the same way, an engineering team can proceed with reasonable confidence. Where they split, the question is genuinely ambiguous and belongs with a notified body or regulatory counsel rather than with an internal assumption. For a compliance programme with a fixed deadline, knowing which questions need expensive advice is most of the battle.

Final Verdict

The EU Machinery Regulation does not ban AI in machines. It moves AI-driven safety functions into a category where somebody independent has to agree that the safety case holds. Classify your portfolio now, book notified body capacity before the queue forms, decide deliberately where learning is permitted, extend risk assessment and cybersecurity evidence to cover the model, and keep one technical file that serves both machinery and AI obligations. January 2027 looks distant until you try to schedule an assessment in it.