Physical AI in the EU: How Machinery Rules, the AI Act, and Cybersecurity Fit Together

Define the machine and each AI function

Here, physical AI means AI used to perceive an environment and influence a machine’s physical actions, such as selecting objects, planning a grasp or navigating a warehouse. The legal assessment starts with the actual product and function. The AI Act has its own definition of an AI system; a supplier’s ‘physical AI’ label does not determine the applicable rules.

Record the robot, tooling, controller, sensors, software versions, intended users and operating environment. Separate production tasks from functions that protect people. Identify who supplies the complete machine, integrates the application, supplies the AI and maintains connected software. This guide focuses on industrial machinery; medical robots, road vehicles, personal-data processing and workplace use can require additional or different rules.

Three frameworks answer different questions

Use a separate scope conclusion for each framework, then connect the supporting evidence. A machine’s declaration, an AI model description and a penetration-test report answer different parts of the review. Check other applicable product legislation, including radio or electromagnetic-compatibility rules where relevant.

Machinery safety includes software and learning behavior

The Machinery Directive 2006/42/EC remains the applicable product framework before the transition. Regulation (EU) 2023/1230 replaces it from 20 January 2027. Establish when the individual machine will first be placed on the EU market or put into service; its design date or a model’s earlier launch does not settle the transition.

Under the Regulation, assess foreseeable evolution of autonomous behavior, protection of safety-related software and data, and control-system failures. Define task and movement limits and show how learning or changed settings are prevented from creating hazards. Include remote access, updates and the response to sensor or communication failures in the review.

Choose the conformity route for the actual category. Annex I Part A includes certain machine-learning safety components and embedded systems ensuring safety functions; these require a route involving a notified body. An AI feature used for an ordinary production task does not by itself put the whole robot in that category.

Check the AI Act’s current machinery route

For the product route in Article 6(1), both conditions matter: the AI must be a safety component, or itself a covered product, and the relevant product must require third-party conformity assessment for health and safety. AI used solely for non-safety production optimization or quality control is distinguished from safety components. Examine the consequences of failure: a function whose failure endangers health and safety can still qualify.

Regulation (EU) 2026/1744 entered into force on 27 July 2026 and moved machinery to Annex I Section B of the AI Act. For Article 6(1) high-risk AI associated with that machinery, Article 2(2) limits direct AI Act application. Relevant AI safety requirements are to be integrated through delegated acts under the Machinery Regulation, applying by 2 August 2028. Plan around that sectoral route and check the applicable delegated acts before release.

Assess separate AI uses individually. For example, worker-management or biometric functions can raise their own Annex III or prohibited-practice questions. The machinery route is specific to the qualifying product-related AI; it is not a blanket conclusion for every AI system used in a factory.

Add product cybersecurity and a reporting process

Screen the robot, separately supplied software and necessary remote processing against the CRA. Check device-to-device connections as well as internet connectivity, and examine the exclusions. For covered products, prepare cybersecurity risk assessment, vulnerability handling, security updates and a defined support period. Connect these records to the machinery assessment where a cyberattack could affect motion or protective functions.

Manufacturers’ CRA reporting duties have applied since 11 September 2026, including for covered products placed on the market before the main requirements apply. An actively exploited vulnerability and a severe incident affecting product security are separate reporting triggers. Set up the Single Reporting Platform process: early warning within 24 hours of awareness and notification within 72 hours, without undue delay. Final-report deadlines differ between vulnerabilities and incidents.

Put the different dates in your release plan

As reviewed on 29 October 2026, the dates below concern different obligations. The AI Act’s general application date was 2 August 2026, with specific exceptions. Its later high-risk dates do not postpone existing machinery safety duties or CRA reporting.

Example: an AI picking cell for March 2027

Suppose a supplier will first place a fenced robot cell on the EU market in March 2027. A vision model selects parts, separate protective equipment controls access and stopping, and remote diagnostics connect the cell to the supplier. This hypothetical example illustrates the questions to resolve; it is not a classification of a particular product.

• Machine: review the complete cell under the Machinery Regulation, including tooling, integration, safety controls and the correct assessment route.
• AI function: document whether part selection is solely non-safety-related and whether failures could endanger people. Confirm the protective system’s independence and coverage before relying on that conclusion.
• Connected product: check CRA scope and the reporting arrangements already required. Plan separately for later units supplied after 11 December 2027, and for modifications to earlier units.

Ask the supplier for a connected evidence package

Use the following as a purchasing review checklist. Some underlying records belong in the manufacturer’s technical documentation rather than a mandatory delivery pack; agree what supporting evidence the supplier will make available.

1. Scope and roles. The exact configuration, intended purpose, AI functions, supply date and businesses responsible for the machine, integration, AI and software support.
2. A reasoned regulatory map. Applicable laws, exclusions, the AI safety-component analysis, conformity routes and the dates used for each conclusion.
3. Safety evidence. Relevant risk-assessment findings, operating limits, validation results, instructions and behavior when sensors, communications or AI outputs fail.
4. Cybersecurity and support. Remote-access controls, component and software identification, vulnerability contact, update process, support end date and incident-reporting responsibilities.
5. Release and change records. Applicable declarations and certificates matched to the model and configuration, plus responsibility for assessing model updates, new tools or changed tasks.

Reopen the assessment when the application changes

Record changes to the model, safety functions, payload, speed, tooling and operating space. Under the Machinery Regulation, a physical or digital change can make the modifier a manufacturer when the complete substantial-modification test is met. An update alone does not establish that result. Recheck the legal definition, affected risks and required protective changes, then document the conclusion before deployment.

Research your physical AI requirements with Regunow

Use Regunow to investigate machinery, AI and cybersecurity requirements, examine source references and analyze supplier documents. Bring the findings together for your team’s review.