EU AI Act Classification for Physical AI: Article 6 Routes, Annex I, Annex III and the Machinery Clock

A Physical AI system can be misclassified when the single question asked is whether its use appears in Annex III. The EU AI Act has two high-risk routes, the 2026 Digital Omnibus changed both the timeline and the treatment of machinery, and the Machinery Regulation and the Cyber Resilience Act run on their own clocks. The unit of analysis is the intended purpose of each AI function and its effect on the physical product and on people, not the model family. This white paper works through the classification clause by clause and offers a method for keeping it stable as the product changes.

Legal and evidence boundary. This is informational analysis, not legal advice. It quotes the Official Journal texts of Regulation (EU) 2024/1689, the amending Regulation (EU) 2026/1744, Regulation (EU) 2023/1230 (applicable from 20 January 2027) with its corrigendum, and the Cyber Resilience Act, Regulation (EU) 2024/2847, as read on 4 November 2026. Reasoning beyond the text is labelled.

Executive summary

• Two routes, not one list. Under Article 6 of the AI Act, a system is high-risk either through Article 6(1), which needs two cumulative conditions tied to Annex I product law, or through Article 6(2) and an Annex III use case.
• The dates moved. Regulation (EU) 2026/1744 applies Chapter III, Sections 1 to 3, except Article 6(5), from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems.
• Machinery changed legal path. The Omnibus moved the Machinery Regulation to Annex I Section B, so for Article 6(1) machinery only Article 6(1), Article 60a and Articles 102 to 112 of the AI Act apply, with AI-specific requirements to come through machinery delegated acts that apply by 2 August 2028.
• The machine clock starts first. The Machinery Regulation applies from 20 January 2027 (Regulation (EU) 2023/1230), and its Annex I Part A covers safety components and embedded systems with fully or partially self-evolving behaviour using machine learning that ensure safety functions.
• A narrower safety-component test. New Article 6(1a) excludes AI used solely for non-safety aspects such as performance optimisation or quality control, but Article 6(1b) treats a system as a safety component where its failure or malfunctioning would endanger health and safety.
• The Article 6(3) derogation is a documented, registered decision. It needs one of four listed conditions, never covers profiling, and requires a prior written assessment and registration.
• Cybersecurity has a bridge. CRA Article 12(1) and new AI Act Article 42(3) link the Cyber Resilience Act to AI Act Article 15, with CRA main obligations from 11 December 2027 and reporting obligations from 11 September 2026.

1. The problem and why it matters now

Four ways a Physical AI classification goes wrong

Classification decides which evidence must exist, which route applies and who signs. Four failure patterns follow from the texts:

1. The Annex III-only check. A team reads the eight Annex III areas, finds no match for a robot or a machine, and stops. Article 6(1) is a separate route that runs through product law.
2. The label check. A label such as "decision support" is treated as a classification. The Act classifies by intended purpose, which Article 3(12) defines as "the use for which an AI system is intended by the provider, including the specific context and conditions of use".
3. The stale memo. A memo based on the earlier machinery classification and high-risk timetable needs reassessment against Regulation (EU) 2026/1744.
4. The single clock. A machine builder plans against the AI Act Annex I date and misses the earlier machinery and CRA dates.

Why the timing changed

Recital 2 of the Omnibus records that "the delayed preparation of standards, which provide technical solutions for providers of high-risk AI systems to ensure compliance with their obligations under that Regulation, and the delayed establishment of the governance and the conformity assessment frameworks at national level has resulted in a compliance burden that is heavier than expected." The amended Article 113 sets 2 December 2027 for Article 6(2) and Annex III systems and 2 August 2028 for Article 6(1) and Annex I systems.

The extra time is not slack: it reflects missing standards and conformity infrastructure, and the same recital says the Commission "should ensure that measures in support of compliance [...] are in place in due time". The Commission's AI Act page reflects the new calendar: "Starting on 2 December 2027, high-risk AI systems will be subject to strict obligations", and "For systems integrated into products such as lifts or toys, the rules will apply from 2 August 2028."

timeline
    title Clocks that touch an AI-enabled machine
    2025-02-02 : AI Act Chapters I and II apply
    2025-08-02 : AI Act governance and GPAI rules apply
    2026-07-27 : Digital Omnibus on AI enters into force
    2026-08-02 : AI Act general date of application
    2026-09-11 : CRA Article 14 reporting obligations apply
    2027-01-20 : Machinery Regulation 2023-1230 applies
    2027-12-02 : AI Act Annex III high-risk rules apply
    2027-12-11 : CRA main obligations apply
    2028-08-02 : AI Act Annex I high-risk rules apply and machinery AI delegated acts apply by this date

2. Technical deep dive: how classification works

2.1 Route A: Article 6(1) has two cumulative conditions

Article 6(1) states that an AI system "shall be considered to be high-risk where both of the following conditions are fulfilled":

• (a) "the AI system is intended to be used as a safety component of a product, or the AI system is itself a product, covered by the Union harmonisation legislation listed in Annex I";
• (b) that product "is required to undergo a third-party conformity assessment, with a view to the placing on the market or the putting into service of that product pursuant to the Union harmonisation legislation listed in Annex I."

Condition (b) is answered by the sector law, not by the AI Act. The Omnibus adds two clarifications:

• Third-party assessment is a trigger, not a consequence. Recital 20 of the Omnibus states that Article 6(1) "should not be understood to require products in which a high-risk AI system is embedded to automatically undergo a third-party conformity assessment involving a notified body." Amended Article 43(3) confirms this for Section A products.
• Non-safety third-party triggers do not count. New Article 6(1c) provides that "a product that is required to undergo a third-party conformity assessment solely due to risks other than risks to health and safety, in particular risks relating to the distribution of radio spectrum or electromagnetic interference that do not affect health and safety, shall not be considered as fulfilling the condition in paragraph 1, point (b)."

This matters for connected robots, because the Radio Equipment Directive 2014/53/EU is listed in Annex I Section A: a third-party assessment required solely for radio-spectrum or electromagnetic-interference risks that do not affect health and safety does not satisfy condition (b).

2.2 What counts as a safety component after the Omnibus

The Omnibus rewrote Article 3(14). A safety component is now "a component of a product or of an AI system which fulfils a safety function for that product or AI system, or the failure or malfunctioning of which endangers the health and safety of persons or property; for the purposes of this definition, a component fulfils a safety function where its intended purpose is to prevent or mitigate risks to health and safety of persons or property".

Two new paragraphs bound the test:

• Article 6(1a): "AI systems that are solely used for non-safety related aspects of user assistance, performance optimisation, service efficiency, automation or convenience or quality control shall not qualify as safety components."
• Article 6(1b): "Notwithstanding paragraph 1a, AI systems the failure or malfunctioning of which would endanger health and safety shall qualify as safety components."

The exclusion in 6(1a) is purpose-based; the override in 6(1b) is consequence-based. A vision model labelled "quality control" falls under 6(1a) only if it is solely used for that; if it also gates a press, a guard or a robot motion, the failure analysis decides. A hazard analysis that traces each AI output to its physical effect is therefore the central classification artefact.

2.3 Annex I: Section A versus Section B

Annex I is split into "Section A. List of Union harmonisation legislation based on the New Legislative Framework" and "Section B. List of other Union harmonisation legislation". The section decides how much of the AI Act applies to a system classified under Article 6(1). The table lists a selection, not the full list; Annex I itself gives every act.

The Omnibus made the move in two edits: "in Section A, point 1 is deleted" (the Machinery Directive 2006/42/EC entry) and, in Section B, point 21 was added for Regulation (EU) 2023/1230. Recital 42 states the intent: "to move to a sectoral approach by moving Regulation (EU) 2023/1230 from Section A to Section B".

For Section A, new Article 2(13) provides that "the application of specific requirements or obligations laid down in Articles 9 to 15 and 17 to 25 may be limited" where the sector legislation "lays down requirements or obligations providing an equivalent or higher level of protection". Under the Omnibus, the Commission must adopt those delegated acts "By 2 August 2027".

2.4 Route B: Article 6(2) and the Annex III areas

Article 6(2) is short: "AI systems referred to in Annex III shall be considered to be high-risk." The quoted Omnibus amendments change Annex I and Annex VIII of the AI Act. The questions below test parts of each Annex III point, not the full entries; read Annex III itself for the complete text.

Critical infrastructure is about safety components, not the sector. Recital 55 of the AI Act explains that such safety components "are systems used to directly protect the physical integrity of critical infrastructure or the health and safety of persons and property but which are not necessary in order for the system to function." It gives as examples "systems for monitoring water pressure or fire alarm controlling systems in cloud computing centres." A utility's load-forecasting model is therefore a different question.

Employment can be triggered downstream. A pick-allocation algorithm is logistics until its output allocates work "based on individual behaviour or personal traits"; configuration alone can cross that line.

2.5 The Article 6(3) derogation, clause by clause

Article 6(3) states that an Annex III system "shall not be considered to be high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making." It applies where any of four conditions is fulfilled: the system is intended "to perform a narrow procedural task"; "to improve the result of a previously completed human activity"; "to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review"; or "to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III."

Three constraints sit around it:

• Profiling override. "Notwithstanding the first subparagraph, an AI system referred to in Annex III shall always be considered to be high-risk where the AI system performs profiling of natural persons." Profiling takes its meaning from Article 4(4) of the GDPR, which covers processing "to evaluate certain personal aspects relating to a natural person, in particular to analyse or predict aspects concerning that natural person's performance at work".
• Documentation and registration. Under Article 6(4), a provider relying on the derogation "shall document its assessment before that system is placed on the market or put into service" and "shall be subject to the registration obligation set out in Article 49(2)". The Omnibus deleted points 7 and 9 of Annex VIII Section B from that registration, but recital 22 states that the provider "remains obligated to document its assessment" and that "national competent authorities should be able to request that assessment."
• Guidance status. The Commission's guidelines page describes "draft guidelines" and states: "Though not legally binding, they reflect the Commission's interpretation and will guide enforcement."

Reasoning, not settled law: a system that monitors or evaluates worker performance processes personal data to evaluate "performance at work", the first example the GDPR names. A team relying on Article 6(3) for a worker-facing function should expect the profiling question to be tested first.

flowchart TD
    A["AI function with a written intended purpose"] --> B{"Safety component or product under Annex I legislation?"}
    B -- "Yes" --> C{"Product requires third-party conformity assessment for health and safety risks?"}
    B -- "No" --> F{"Matches an Annex III entry?"}
    C -- "Yes" --> D{"Annex I Section A or Section B?"}
    C -- "No" --> F
    D -- "Section A" --> E1["High-risk via Article 6(1), from 2 Aug 2028"]
    D -- "Section B incl. machinery" --> E2["High-risk via Article 6(1), only Art 6(1), 60a, 102-112 apply, requirements via sector law"]
    E1 --> X1{"Same function also matches an Annex III entry?"}
    E2 --> X2{"Same function also matches an Annex III entry?"}
    X1 -- "Yes" --> Y1["Dual route, follow the Section A procedure under Article 43(3), check the Annex III date"]
    X2 -- "Yes" --> Y2["Dual route in a Section B product, treatment not stated, classify per function and ask counsel"]
    X1 -- "No" --> N1["Article 6(1) route only"]
    X2 -- "No" --> N1
    F -- "No" --> G["Not high-risk under Article 6, check other obligations for its functions"]
    F -- "Yes" --> H{"Profiling of natural persons?"}
    H -- "Yes" --> I["High-risk via Article 6(2), from 2 Dec 2027"]
    H -- "No" --> J{"Article 6(3) condition met and no significant risk?"}
    J -- "No" --> I
    J -- "Yes" --> K["Not high-risk, document before placing on market and register under Article 49(2)"]

2.6 Dual-route systems

One product can carry several AI functions, and one function can sit on both routes. Amended Article 43(3) settles one case: "where a high-risk AI system is both covered by the Union harmonisation legislation listed in Section A of Annex I and it falls within one of the categories listed in Annex III, the provider of that system shall follow the relevant conformity assessment procedure as required pursuant to the relevant Union harmonisation legislation listed in Section A of Annex I."

The texts read contain no such sentence for Section B. The amended Article 2(2) excerpt read here lists the provisions that apply to Section B products; it does not establish how a separate Annex III function in the same machine, such as worker-performance monitoring, is treated. Classify per function, and ask counsel how a function meeting both routes in a Section B product is treated.

3. Interplay with the Machinery Regulation and the Cyber Resilience Act

3.1 The Machinery Regulation decides condition (b) for AI in machines

For an AI-enabled machine, Article 6(1)(b) asks whether the Machinery Regulation requires third-party conformity assessment. For Annex I Part A categories, Article 25(2) offers EU type-examination (module B) followed by conformity to type based on internal production control (module C), conformity based on full quality assurance (module H), or conformity based on unit verification (module G). Internal production control (module A) is not an option for Part A; Article 25(3) lists it for Part B, where the machinery is designed and constructed in accordance with the harmonised standards or common specifications covering all the relevant essential health and safety requirements, and Article 25(4) applies it to categories not listed in Annex I. In module B, "a notified body examines the technical design of machinery or a related product".

Annex I Part A includes two AI-specific entries:

• point 5: "Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions";
• point 6: "Machinery that has embedded systems with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions that have not been placed independently on the market, in respect only of those systems."

Recital 55 of the Machinery Regulation says these provisions "should only apply to systems with a fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions" and "should not apply to software incapable of learning or evolving, and programmed only to execute certain automated functions".

Original insight (Hyperion reasoning): "self-evolving behaviour using machine learning approaches" is the hinge for both regimes at once. If an ML safety function falls under Part A point 5 or 6, the machine needs a notified-body procedure under the Machinery Regulation, which satisfies AI Act Article 6(1)(b) and makes the function high-risk under the AI Act. If it does not, neither consequence follows on that ground. The texts quoted here do not define "self-evolving", so record the interpretation used for freezing weights after validation, or for allowing on-site adaptation, as a classification decision with its rationale, and obtain a system-specific assessment.

3.2 What machinery AI must meet, and when

The Machinery Regulation applies from 20 January 2027 (see the corrigendum and the consolidated text). Section 1.2.1 of its Annex III already addresses control systems "with fully or partially self-evolving behaviour or logic that are designed to operate with varying levels of autonomy": they "shall not cause the machinery or related product to perform actions beyond its defined task and movement space"; recording of "data on the safety related decision-making process for software based safety systems ensuring safety function including safety components" must be enabled, with that data "retained for one year after its collection"; and "it shall be possible at all times to correct the machinery or related product in order to maintain its inherent safety."

The Omnibus adds a second layer. It provides for delegated acts under the Machinery Regulation so that "the relevant requirements set out in Chapter III, Section 2, and Articles 17, 19, 72 and 73 of Regulation (EU) 2024/1689 are reflected." "Those delegated acts shall apply by 2 August 2028." Meanwhile, new Article 20(10) of the Machinery Regulation (applicable from 20 January 2027) provides that "common specifications adopted pursuant to Articles 40 and, respectively, 41 of Regulation (EU) 2024/1689 shall be presumed to be in conformity with the essential health and safety requirements set out in Annex III to this Regulation".

Derived observation: for more than eighteen months, from 20 January 2027 to 2 August 2028, the Machinery Regulation's own requirements apply to an AI-enabled machine while the AI-specific delegated acts may not yet apply. Machinery design evidence is needed first.

3.3 The Cyber Resilience Act and AI Act Article 15

The CRA "applies to products with digital elements made available on the market, the intended purpose or reasonably foreseeable use of which includes a direct or indirect logical or physical data connection to a device or network." Article 2(2) to (4) excludes products covered by the medical-device, in vitro diagnostic and vehicle general-safety regulations, products certified under Regulation (EU) 2018/1139, and marine equipment under Directive 2014/90/EU.

Its Article 71 states: "This Regulation shall apply from 11 December 2027. However, Article 14 shall apply from 11 September 2026". Article 14 sets "Reporting obligations of manufacturers", starting with actively exploited vulnerabilities; Article 69(3) concerns products "that have been placed on the market before 11 December 2027", so assess it when determining whether those duties apply to products already on the market.

CRA Article 12(1) provides that, "without prejudice to the requirements relating to accuracy and robustness set out in Article 15" of the AI Act, in-scope products "classified as high-risk AI systems pursuant to Article 6" of that Regulation "shall be deemed to comply with the cybersecurity requirements set out in Article 15" where three conditions are met: the products fulfil the essential cybersecurity requirements in Part I of CRA Annex I; the manufacturer's processes comply with those in Part II; and the required level of protection "is demonstrated in the EU declaration of conformity issued under this Regulation". New AI Act Article 42(3), inserted by the Omnibus, applies the same presumption where "the conditions laid down in Article 12(1) of that Regulation are fulfilled". One cybersecurity evidence set can serve both regimes; accuracy and robustness evidence stays separate.

sequenceDiagram
    participant S as "AI component supplier"
    participant M as "Machine manufacturer"
    participant N as "Notified body"
    participant D as "Deploying site"
    S->>M: "Agree information and technical access in writing"
    M->>M: "Classify each AI function and record the self-evolving decision"
    M->>N: "Part A procedure where self-evolving ML ensures a safety function"
    N-->>M: "Type-examination outcome"
    M->>D: "Instructions for use, oversight measures, log handling"
    D->>M: "Operational monitoring and incident information"
    M->>M: "CRA vulnerability reporting from 11 Sep 2026"

4. Research evidence: what the sources are, and what they are not

No peer-reviewed study of misclassification rates is cited, no prevalence is claimed and no vendor claim is relied on. The evidence is legal text of differing weight, and treating those texts as interchangeable is itself a risk.

The texts support the findings above. The quoted Commission page describes the classification guidelines as draft. The texts read here do not establish whether the machinery delegated acts, the Article 2(13) delegated acts or the delayed harmonised standards have since been adopted.

5. Business and decision analysis

5.1 What the classification outcome drives

Classification sets a programme's cost structure. The drivers are lead time with external bodies, evidence volume, the number of signing parties and how fast evidence goes stale.

5.2 Ownership and the value chain

Roles are functional. Article 3(3) includes placing a system on the market or putting it into service under one's own name or trademark in its provider definition; assess the complete definition when assigning roles. A deployer uses "an AI system under its authority". Article 25(1) considers a distributor, importer, deployer or other third party to be a provider of a high-risk AI system where it puts "their name or trademark on a high-risk AI system already placed on the market or put into service" (without prejudice to contractual arrangements allocating the obligations otherwise), makes "a substantial modification" to such a system so that it "remains a high-risk AI system pursuant to Article 6", or modifies the intended purpose of a system not classified as high-risk so that it "becomes a high-risk AI system in accordance with Article 6".

• Intended purpose has one owner. Record the intended purpose under Article 3(12): "the use for which an AI system is intended by the provider, including the specific context and conditions of use".
• Supplier terms can be an obligation. Amended Article 25(4) requires the provider of a high-risk AI system and a third party that supplies "an AI system, AI model, tools, services, components, or processes that are used or integrated in a high-risk AI system" to "specify the necessary information, capabilities, technical access and other assistance" "by written agreement". It does not apply to third parties making tools, services, processes or components, other than general-purpose AI models, accessible to the public under a free and open-source licence. For a Section B product such as machinery, amended Article 2(2) applies only Article 6(1), Article 60a and Articles 102 to 112, so identify whether Article 25(4) applies to the particular AI function. The amended list expressly includes "AI model".
• Installed fleets have a different trigger. Amended Article 111(2) contains a condition, "only if, as from that date, those systems are subject to significant changes in their designs"; assess its complete scope and exceptions before relying on it. A design-change log is the record that condition turns on.

6. A method: the Four-Hinge Classification Trace

This is a Hyperion reasoning device derived from the texts, not field evidence or a legal test.

Every classification turns on four hinges. If one is undocumented or controlled by someone else, the classification can flip without anyone deciding it.

1. Purpose hinge. The intended purpose, including "the specific context and conditions of use", as the provider states it in the "instructions for use, promotional or sales materials and statements" that Article 3(12) refers to.
2. Authority hinge. What the output can do: a safety function or failure consequence (Articles 3(14), 6(1a), 6(1b)), or material influence on a decision (Article 6(3)).
3. Product-law hinge. The Annex I legislation, its section, and whether it requires third-party assessment for health and safety reasons, including the self-evolving question for machinery.
4. Change hinge. Who can alter any of the above after release: configuration, retraining, new data sources, new users or a new site.

A fragility ratio helps prioritise work:

$F = \frac{u}{d}$

Here d is the number of facts that determine the classification of one AI function, and u is the number of those facts not fixed in a document the provider controls. F = 0 means every determinative fact is pinned; F near 1 means the classification depends on decisions made elsewhere. It is a prioritisation aid, not a measure of legal risk.

Worked example (hypothetical)

The company, product and figures below are invented to illustrate the method. A manufacturer sells autonomous mobile robots for warehouses, with three AI functions.

F3 is stable only because the architecture keeps thermal protection outside the model; if the model set charge limits, Article 6(1b) would need examining. F2 has the highest ratio: the provider ships a logistics optimiser, and the customer's configuration can turn it into an Annex III employment system, at which point Article 25(1)(c) may make the customer a provider. The design response is to give picker-level weighting its own documented intended purpose, or remove it. F1 hinges on the self-evolving decision, kept under change control.

stateDiagram-v2
    state "Not high-risk, purpose documented" as NHR
    state "Annex III high-risk" as A3
    state "Article 6(3) derogation, registered" as D63
    state "Article 6(1) high-risk via product law" as P61
    state "Reassess both Article 6 routes and the complete Article 6(3) test" as R
    [*] --> NHR
    NHR --> A3: "Customer enables per-worker task weighting"
    NHR --> R: "New Annex III use"
    NHR --> R: "Output becomes part of a safety function"
    P61 --> R: "Safety function moved to a non-AI protection layer"
    R --> D63: "Article 6(1) not met, Annex III use, complete Article 6(3) test met, assessment documented"
    R --> A3: "Annex III use, Article 6(3) test not met"
    R --> P61: "Article 6(1) conditions met"
    R --> NHR: "Neither route applies"
    D63 --> R: "Output starts to influence the decision"
    D63 --> A3: "Profiles natural persons"
    A3 --> A3: "Retrain or new site triggers re-assessment"

7. Implementation roadmap and checklist

Before 20 January 2027

• Inventory every AI function, its physical effect and the people it affects, and write its intended purpose.
• For each machine, record the Annex I Part A and self-evolving decision, the Article 25 procedure and the section 1.2.1 control-system measures.
• Operate CRA reporting, which applies from 11 September 2026.

Before 2 December 2027

• Test every Annex III entry against real workflows and customer configurations; for any Article 6(3) reliance, complete the assessment, profiling analysis and Article 49(2) registration first.
• For Annex III high-risk systems, map Articles 9 to 16 to owned artefacts and give deployers what Article 26 needs.
• Prepare CRA evidence that can also support the Article 15 presumption, ahead of 11 December 2027.

Before 2 August 2028

• Section A products: integrate AI requirements into the sector assessment and track Article 2(13) delegated acts.
• Machinery: track the delegated acts amending Annex III of the Machinery Regulation and the related standards.

Always

• Log which hinge each change touches and re-run classification when u rises; where Article 25(4) applies, put its terms into supplier contracts; revalidate against Official Journal texts, not summaries.

Related resources: the Physical AI Product System, the Product Decision Review, the EU AI Act classifier and the EU AI Act tracker. A live case needs qualified counsel and conformity-assessment expertise.

8. Limitations and open questions

• "Self-evolving" is not defined in the texts quoted here. Whether a model trained offline and frozen before release falls under Annex I Part A points 5 and 6 is not answered by those texts.
• Delegated acts are pending. The machinery AI requirements and the Article 2(13) limitations for Section A are empowerments, not adopted content, in the texts read.
• Guidelines are draft; final examples may change how borderline cases are read.
• Section B with Annex III. The treatment of a function meeting both routes inside a Section B product is not stated in the texts read; section 2.6 records the open question.
• No prevalence data. The failure modes described here are structural, derived from the texts, not measured.

9. FAQ

Is every robot or AI-enabled machine high-risk under the EU AI Act?

No. Article 6(1) requires both a safety-component or product link to Annex I legislation and a third-party conformity assessment requirement under it, and Article 6(1a) excludes AI used solely for non-safety aspects unless Article 6(1b) applies.

When do the high-risk obligations apply to an AI function in a machine?

The Machinery Regulation itself applies from 20 January 2027. For AI Act Chapter III, Sections 1 to 3, except Article 6(5), the dates are 2 December 2027 for Article 6(2) and Annex III systems and 2 August 2028 for Article 6(1) and Annex I systems. Machinery AI requirements follow the sectoral route described above, with delegated acts to apply by 2 August 2028.

Can a quality-inspection model avoid high-risk classification?

Article 6(1a) names quality control among excluded non-safety uses, but only where the system is solely used for that. If its output also controls a hazardous motion or guard, Article 6(1b) brings the failure consequence back in.

Can a provider rely on Article 6(3) for a worker-facing system?

Only if one of the four conditions is met, the system poses no significant risk, and it does not profile natural persons. Because the GDPR names performance at work in its profiling definition, worker-performance functions face a high bar.

Does meeting the Cyber Resilience Act cover AI Act cybersecurity?

CRA Article 12(1) and AI Act Article 42(3) create a presumption for the Article 15 cybersecurity requirements when the stated conditions are met. The accuracy and robustness requirements of Article 15 are not covered by that presumption.