When your AI-enabled machinery must satisfy both the EU Machinery Regulation and the AI Act. Two examiners, three deadlines, one evidence stream.
Who this is for: Machine safety engineers responsible for CE marking, AI governance leads managing EU AI Act compliance, Notified Body assessors evaluating machinery with ML safety functions, product liability counsel, and robotics OEM compliance teams preparing for dual certification.
Prerequisite: Familiarity with EU Machinery Regulation 2023/1230 and the EU AI Act (Regulation 2024/1689). Readers new to either regulation should start with the dedicated crosswalk guides linked in Section 10.
Assessor Notice: SWT3 witness anchors provide timestamped, tamper-evident records of operational state. They supplement -- but do not replace -- assessor judgment, Notified Body examination, or regulatory authority. Each regulation retains its own assessment authority. Evidence anchors prove what happened. Assessors determine what it means.
Not every AI system in a factory triggers dual compliance. The overlap is specific: it occurs when a single product falls under both the Machinery Regulation and the AI Act at the same time. Three questions determine whether your system is in that overlap.
| Is your system "machinery" under Regulation 2023/1230 Art. 2? | Machinery means an assembly with at least one moving part, fitted with or intended to be fitted with a drive system other than directly applied human or animal effort. This includes safety components placed on the market separately. |
| Does the machinery contain AI or ML performing a safety function? | A safety function is any function whose failure or malfunction endangers persons. If your system uses machine learning for collision avoidance, emergency stop logic, hazard detection, or safe speed monitoring, the answer is yes. |
| Is the AI component classified as high-risk under the AI Act? | Safety components in machinery are listed in AI Act Annex I. Standalone high-risk categories (biometrics, critical infrastructure, employment decisions) are listed in Annex III. If either applies, the AI Act imposes additional obligations beyond the Machinery Regulation. |
If all three answers are yes: your system is dual-regulated. The Machinery Regulation governs the safety conformity assessment. The AI Act governs risk management, transparency, human oversight, and documentation for the AI component itself. Both apply. Neither replaces the other.
If the first two are yes but the AI is not high-risk: the Machinery Regulation applies in full. The AI Act still requires transparency (Art. 50) if the system interacts with people, and the prohibited practices (Art. 5) apply regardless of risk classification.
Mandatory third-party assessment. If your ML safety component falls under Annex I Part A of the Machinery Regulation, self-assessment is not an option. A Notified Body must perform EU Type-Examination. This is not a recommendation. It is a legal requirement effective January 20, 2027.
On July 27, 2026, Regulation (EU) 2026/1744 (the "Digital Omnibus on AI") entered into force. It amended the AI Act in ways that directly affect machinery manufacturers. Three changes matter most.
The Omnibus moved the Machinery Regulation from AI Act Annex I Section A to Section B. In practical terms, this means the Machinery Regulation's conformity assessment procedure takes precedence for safety-related questions. The Notified Body assessing your machinery under 2023/1230 is the primary examiner for AI safety function compliance, not the AI Office.
However, the AI Act's substantive requirements -- risk management (Art. 9), data governance (Art. 10), human oversight (Art. 14), accuracy and robustness (Art. 15) -- still apply to the AI component. The Omnibus changed who leads the assessment, not what must be assessed.
Annex III standalone high-risk obligations moved from August 2, 2026 to December 2, 2027. Annex I embedded safety-component obligations moved to August 2, 2028. These deferrals give manufacturers more time for AI Act compliance, but they do not delay the Machinery Regulation.
The Machinery Regulation's January 20, 2027 application date was not affected by the Omnibus. It is unchanged. Transparency obligations under AI Act Art. 50 are already enforceable. Prohibited practices under Art. 5 are already enforceable. AI literacy requirements under Art. 4 are already enforceable.
For the full 8-amendment breakdown, timeline comparison, and procedure mapping, see EU AI Act Digital Omnibus Update (July 2026).
A single AI-enabled machine placed on the EU market in early 2027 faces four regulatory deadlines within 21 months of each other. These are not sequential. They overlap, and evidence you build for the earliest deadline directly serves the later ones.
Directive 2024/2853 must be transposed by all EU member states. AI software becomes a "product" subject to strict liability. The development risk defense is unavailable for defects arising from post-deployment software updates. Every OTA model update to a factory robot eliminates this defense.
Regulation 2023/1230 replaces the Machinery Directive. Products placed on the EU market after this date must comply. ML safety components require mandatory Notified Body EU Type-Examination under Annex I Part A. No deferral. No grace period.
Standalone high-risk AI obligations become enforceable (deferred from August 2026 by the Omnibus). If the AI component in your machinery also falls under Annex III categories -- such as biometric identification or safety components of critical infrastructure -- these requirements apply.
High-risk obligations for AI embedded in regulated products (including machinery) become enforceable. By this point, 18 months of continuous witness evidence from Machinery Regulation compliance directly serves AI Act conformity documentation.
Consider a robotic welding cell with ML-based collision avoidance. The Product Liability Directive makes the AI software a product on December 9, 2026. The Machinery Regulation requires NB assessment of the ML safety function on January 20, 2027. The AI Act's standalone high-risk requirements arrive in December 2027. And the embedded safety-component rules arrive in August 2028. These deadlines overlap. Evidence you build for the Machinery Regulation NB assessment directly serves AI Act compliance over a year later.
The Machinery Regulation does not wait for the AI Act. January 20, 2027 is a hard boundary with no Omnibus deferral. Products placed on the EU market after that date without NB EU Type-Examination for AI/ML safety functions are non-compliant.
Dozens of Notified Bodies are registered under Regulation 2023/1230, but very few hold notification scope covering machine-learning safety components (Annex I Part A, categories 5 and 6). Fewer still are scoped for embedded ML safety systems. The pool of qualified examiners is small relative to the number of manufacturers that need EU Type-Examination for AI-equipped machinery.
Lead times for NB engagement are running 6 to 9 months. If you have not started the NB engagement process, your first available examination slot may not arrive until Q3 or Q4 of 2027 -- well past the January 20 enforcement date.
The dual-examiner challenge adds complexity. The Notified Body assessing your machinery safety is not necessarily the same organization that will evaluate your AI Act conformity when those obligations become enforceable. Two different examiners may request overlapping evidence in different formats, creating reconciliation risk if evidence was gathered separately.
This is where a unified witness protocol changes the calculus. One protocol feeds both regulatory technical files simultaneously. The NB assessing your machinery safety function under 2023/1230 Annex III 1.2.1 receives timestamped, tamper-evident evidence of safe state transitions, drift detection, and human override capabilities. The same evidence anchors -- identical fingerprints, identical timestamps -- serve the AI Act assessor evaluating Art. 9 risk management, Art. 14 human oversight, and Art. 15 accuracy and robustness.
Two examiners, one evidence stream. No duplication. No reconciliation. No conflicting audit trails. The assessor reviewing your file in January 2027 sees the same cryptographic fingerprints that the AI Act assessor will review in late 2028. The evidence was gathered once, witnessed once, and anchored once. It serves both regimes because both regimes ask fundamentally the same question: was this system operating safely, and can you prove it?
For NB engagement templates, conformity assessment checklists, and pilot package materials, see Notified Body Pilot Package.
Nine SWT3 procedures serve obligations under both the Machinery Regulation and the AI Act. These are "dual-benefit" procedures: a single witness anchor satisfies requirements in both regulatory files simultaneously. The coverage matrix below shows which obligation each procedure addresses in each regulation.
| Procedure | Machinery Reg Provision | AI Act Article | Evidence Captured |
|---|---|---|---|
| AI-SAFE.1 | Annex III 1.2.1 | Art. 15 | Safe state transition, trigger, suspended actions |
| AI-DRIFT.1 | Art. 18 | Art. 9(2)(a) | Baseline hash, drift metric, threshold |
| AI-EMRG.1 | Annex III 1.2.1 | Art. 14 | Override trigger, authorization, fallback state |
| AI-HITL.1 | Annex III 1.2.1 | Art. 14(4) | Oversight mode, intervention capability |
| AI-INCIDENT.1 | Art. 18 | Art. 62 | Incident class, timeline, corrective actions |
| AI-TRANS.1 | -- | Art. 50 | Disclosure method, content, timing |
| AI-CYBER.1 | Annex III 1.1.9 | Art. 15(4) | Assessment scope, vulnerabilities, mitigation |
| HBOM-INV.1 | Technical file | Art. 11 | Component manifest, firmware, manifest hash |
| AI-CHAIN.1 | Art. 5 | Art. 25 | Supplier attestation, component provenance |
Note: AI-TRANS.1 is included because AI Act transparency obligations (Art. 50) are already enforceable and apply to the same physical system, even though the Machinery Regulation does not have a separate transparency requirement.
Machinery Regulation: Annex III 1.2.1 requires control systems to enter a safe state upon failure detection. The system must reach a safe condition without creating additional hazards during the transition.
AI Act: Art. 15 requires high-risk AI to be designed so that output errors do not lead to safety risks. The system must fail safely and predictably.
Evidence anchored: Timestamp of state transition, trigger type (sensor, watchdog, operator), suspended actions, recovery path availability, time to safe state.
Verify that the safe state definition documented in the witness anchor matches the risk assessment. The NB will check whether the AI's safe state is consistent with the machinery's overall safety concept. Cross-reference with ISO 13849-1 performance levels.
Machinery Regulation: Art. 18 defines "substantial modification" as any change affecting safety. A model drift that changes the behavior of a safety function may constitute a substantial modification, triggering re-assessment.
AI Act: Art. 9(2)(a) requires the risk management system to identify and analyze known and foreseeable risks, including those from self-evolving behavior.
Evidence anchored: Baseline model hash, current model hash, drift metric value, configured threshold, detection timestamp.
A drift anchor that exceeds the configured threshold is a flag for potential substantial modification. The NB should evaluate whether the behavioral change affects the safety function's performance level. Under the AI Act, continuous drift monitoring satisfies the Art. 9 requirement for ongoing risk identification.
Machinery Regulation: Annex III 1.2.1 requires every machine to have an emergency stop device that overrides all operating modes. The AI must not prevent or delay emergency stop execution.
AI Act: Art. 14 requires that humans can effectively oversee the AI system and, where necessary, override or reverse its decisions.
Evidence anchored: Override trigger source, authorization level (automatic/manual), fallback state, recovery availability, time from trigger to safe state.
The critical question is latency. If the AI component introduces any delay between the emergency stop signal and the physical stop, it must be documented and justified. The witness anchor records the timeline. Compare the anchored time-to-safe-state against the values declared in the risk assessment.
Machinery Regulation: Annex III 1.2.1 requires operators to be able to override automated functions and regain manual control at any point during operation.
AI Act: Art. 14(4) specifies three oversight modes: human-in-the-loop (approval before action), human-on-the-loop (monitoring with intervention capability), and human-in-command (strategic direction with veto authority).
Evidence anchored: Active oversight mode, intervention capability status, override latency, operator qualification level.
The Machinery Regulation requires override capability. The AI Act requires documented oversight mode. A single anchor captures both. Verify that the oversight mode in the anchor matches the mode declared in the conformity documentation.
Machinery Regulation: Art. 18 obliges manufacturers to conduct post-market surveillance. Serious incidents involving machinery must be reported to the market surveillance authority of the member state where the incident occurred.
AI Act: Art. 62 requires providers and deployers of high-risk AI to report serious incidents to the relevant market surveillance authority. The report must include details of the AI's role in the incident.
Evidence anchored: Incident classification, event timeline, affected systems, AI component role, corrective actions taken, reporting status.
Dual reporting is likely. The same incident may need to be reported under both Art. 18 (machinery) and Art. 62 (AI Act) to the same or different market surveillance authorities. The witness anchor provides a single authoritative timeline that both reports can reference.
AI Act: Art. 50 requires that persons interacting with an AI system are informed that they are doing so. This obligation is enforceable now (not deferred by the Omnibus) and applies to the same physical system.
Machinery Regulation: No separate transparency requirement, but the Machinery Regulation's instructions for use (Annex III 1.7.4) must describe the system's intended use, including AI capabilities.
Evidence anchored: Disclosure method, disclosure content, timing relative to interaction, acknowledgment status.
Art. 50 transparency is already enforceable. If the machinery interacts with operators or bystanders and uses AI, transparency disclosure evidence should be in place now, not deferred until the Annex I deadline.
Machinery Regulation: Annex III 1.1.9 introduces Essential Health and Safety Requirements (EHSRs) for cybersecurity -- new in 2023/1230, absent from the old Machinery Directive. Machinery must be protected against corruption and unauthorized modification.
AI Act: Art. 15(4) requires cybersecurity measures proportionate to the risk level, including protection against adversarial manipulation of training data or model parameters.
Evidence anchored: Assessment scope, vulnerability state, mitigation status, last assessment date.
Cybersecurity is new territory for machinery NBs. Many are accustomed to assessing physical safety, not digital threats. The witness anchor provides structured evidence that helps bridge the gap between traditional safety assessment and the new cybersecurity EHSRs.
Machinery Regulation: The technical file must include a full description of the machinery, including hardware components. For AI safety functions, the computing hardware running the ML model is part of the safety-relevant bill of materials.
AI Act: Art. 11 requires technical documentation that describes the computational resources used by the AI system, including hardware specifications.
Evidence anchored: Component manifest, firmware versions, manifest hash, baseline comparison delta.
Hardware changes can affect ML inference behavior (different GPU, different floating-point precision, different latency). If the hardware manifest hash changes between anchors, investigate whether the ML safety function's validated performance levels still hold.
Machinery Regulation: Art. 5 places obligations on manufacturers that extend to components sourced from suppliers. If a third-party AI model is embedded as a safety component, the manufacturer remains responsible for conformity.
AI Act: Art. 25 defines obligations for distributors and importers of high-risk AI components. The provider of the AI component must supply sufficient technical documentation for downstream conformity assessment.
Evidence anchored: Supplier attestation, component provenance, delegation chain, documentation completeness.
If the ML model was developed by a third party and integrated by the machinery manufacturer, verify that the supply chain anchor documents which entity bears responsibility for which aspect of the conformity assessment. The Machinery Regulation makes the manufacturer ultimately responsible; the AI Act distributes obligations across the supply chain.
Consider a Tier 1 automotive supplier operating a robotic welding cell in a German factory. The cell uses industrial robots conforming to ISO 10218 and includes two AI components:
The system receives quarterly over-the-air model updates from the AI provider. It is being placed on the EU market for the first time in February 2027.
Machinery Regulation (Jan 20, 2027): The ML collision avoidance system is a safety component with self-evolving behavior. It falls under Annex I Part A. A Notified Body must perform EU Type-Examination. The technical file must include evidence of safe state behavior (AI-SAFE.1), emergency stop override (AI-EMRG.1), cybersecurity protection (AI-CYBER.1), hardware inventory (HBOM-INV.1), and supply chain documentation for the third-party AI model (AI-CHAIN.1).
Product Liability Directive (Dec 9, 2026): The AI software is a "product" subject to strict liability. Each OTA update eliminates the development risk defense for defects introduced by that update. The manufacturer needs a continuous evidence trail showing what version was running when, and what testing preceded each deployment (AI-DRIFT.1).
AI Act Annex I (Aug 2, 2028): The collision avoidance system is an AI safety component embedded in machinery. When Annex I obligations become enforceable, the manufacturer must demonstrate risk management (Art. 9), human oversight (Art. 14), and accuracy/robustness (Art. 15). The adaptive weld-path optimizer, if not classified as high-risk, still requires transparency disclosure (Art. 50) because factory workers interact with it.
| When | Event | Anchor Minted | Serves |
|---|---|---|---|
| Nov 2026 | Hardware manifest captured at commissioning | HBOM-INV.1 | Machinery Reg technical file + AI Act Art. 11 |
| Nov 2026 | Supply chain documentation for third-party ML model | AI-CHAIN.1 | Machinery Reg Art. 5 + AI Act Art. 25 |
| Dec 2026 | Collision avoidance triggers protective stop | AI-SAFE.1 | Machinery Reg Annex III 1.2.1 + AI Act Art. 15 |
| Dec 2026 | Operator hits emergency stop during maintenance | AI-EMRG.1 | Machinery Reg Annex III 1.2.1 + AI Act Art. 14 |
| Jan 2027 | Q1 OTA update deployed; baseline drift measured | AI-DRIFT.1 | Machinery Reg Art. 18 (substantial modification?) + AI Act Art. 9 |
| Jan 2027 | NB receives technical file with 60+ days of witness anchors | -- | Machinery Reg EU Type-Examination |
| Ongoing | Transparency disclosure to factory workers | AI-TRANS.1 | AI Act Art. 50 (enforceable now) |
| Dec 2028 | AI Act assessor reviews same evidence chain | -- | AI Act Annex I conformity |
The Notified Body reviewing the technical file in January 2027 sees a continuous, tamper-evident record of safety behavior going back to commissioning. Over a year later, the AI Act assessor sees the same record -- the same fingerprints, the same timestamps. No duplicate evidence gathering. No conflicting narratives. One evidence stream, two regulatory files.
This phased approach starts with the most urgent deadline and builds evidence that compounds across all three regulations.
| Phase | Timeline | Focus | Key Procedures |
|---|---|---|---|
| Phase 1 | Now to Dec 9, 2026 | PLD readiness. Establish the evidence chain before the Directive makes AI a "product." Capture hardware manifests and supply chain provenance. Document what version is running and where it came from. | HBOM-INV.1, AI-CHAIN.1, AI-TRANS.1 |
| Phase 2 | Dec 2026 to Jan 20, 2027 | Machinery Regulation readiness. Complete NB engagement. Technical file must include safety state evidence, emergency override records, cybersecurity assessment, and human oversight documentation. Begin drift baselines for every deployed model version. | AI-SAFE.1, AI-EMRG.1, AI-CYBER.1, AI-HITL.1, AI-DRIFT.1 |
| Phase 3 | Jan to Dec 2027 | Dual accumulation. Continue minting anchors under normal operations. Evidence accumulates for the AI Act Annex III deadline (Dec 2027). Each anchor serves both the machinery technical file and the emerging AI Act conformity documentation. Incident reporting procedures are exercised and documented. | AI-INCIDENT.1, all Phase 1-2 procedures (ongoing) |
| Phase 4 | Jan to Aug 2028 | Annex I completion. AI Act embedded safety-component obligations become enforceable. By this point, 18+ months of continuous witness evidence exists. The conformity documentation is substantially complete. The remaining work is formatting and submission. | Documentation compilation (no new procedures) |
Start date matters. Organizations that begin witnessing in Phase 1 present the NB with 60+ days of continuous evidence at the Machinery Regulation deadline. Organizations that wait until January 2027 present an empty file.
These examples show dual-regulation witnessing. The jurisdiction and legal_basis fields tag each anchor for both regulatory files, making it clear to assessors which frameworks the evidence serves.
from swt3_ai import Witness
witness = Witness(
endpoint="https://your-instance.example.com/api/v1/witness",
api_key="axm_your_key",
agent_id="welding-cell-ML-01",
jurisdiction="DE",
legal_basis="EU-2023-1230,EU-2024-1689"
)
# Safety state transition (Machinery Reg Annex III 1.2.1 + AI Act Art. 15)
witness.witness_event(
procedure="AI-SAFE.1",
factor_a="proximity_threshold",
factor_b="actuator_write,weld_arc",
factor_c="recovery_available"
)
# Drift detection after OTA update (Art. 18 + Art. 9)
witness.witness_event(
procedure="AI-DRIFT.1",
factor_a="a9f3c2e1",
factor_b="7b21e843",
factor_c="drift=0.023,threshold=0.05"
)
# Hardware inventory for technical file (Machinery file + Art. 11)
witness.witness_hardware(
component_count=47,
manifest_hash="d4e8a1...",
baseline_delta=0
)
witness.flush()import { Witness } from '@tenova/swt3-ai';
const witness = new Witness({
endpoint: 'https://your-instance.example.com/api/v1/witness',
apiKey: 'axm_your_key',
agentId: 'welding-cell-ML-01',
jurisdiction: 'DE',
legalBasis: 'EU-2023-1230,EU-2024-1689'
});
// Emergency override (Machinery Reg Annex III 1.2.1 + AI Act Art. 14)
witness.witnessEvent({
procedure: 'AI-EMRG.1',
factorA: 'safety_sensor',
factorB: 'automatic',
factorC: 'full_stop,recovery_available'
});
// Human oversight mode (Machinery Reg 1.2.1 + AI Act Art. 14(4))
witness.witnessEvent({
procedure: 'AI-HITL.1',
factorA: 'human_on_the_loop',
factorB: 'intervention_capable',
factorC: 'override_latency_ms=12'
});
await witness.flush();For full SDK documentation, adapter guides, and additional examples, see SDK Documentation.
Two separate teams, two separate evidence sets, two separate timelines. The result: duplicate work, conflicting narratives, and unnecessary reconciliation when the NB asks about the same safety function the AI Act assessor will review later. A unified evidence approach eliminates this entirely.
The Machinery Regulation does not wait. If you place ML-enabled machinery on the EU market after January 20, 2027 without NB assessment of the AI safety function, you are non-compliant regardless of the AI Act deferral. The Omnibus extended the AI Act timeline. It did not extend the Machinery Regulation timeline.
The Omnibus routes the conformity assessment procedure through the Machinery Regulation for safety functions. It does not exempt the AI component from the AI Act's substantive requirements. Risk management (Art. 9), human oversight (Art. 14), transparency (Art. 50), and documentation (Art. 11) still apply. As Bird & Bird's analysis notes, both regulations "were not designed to be in conflict, and largely they are not; but they were designed independently, and where they meet, they leave significant questions unanswered."
Only 3 Notified Bodies are currently certified for ML safety components under Annex I Part A. Lead times are 6 to 9 months. If you begin the engagement process in December 2026, your first available examination slot may not arrive until Q3 2027. Start now.
Both regulations require ongoing evidence, not just a one-time assessment. Art. 18 of the Machinery Regulation requires post-market surveillance including investigation of complaints and non-conformity. Art. 9 of the AI Act requires a risk management system that operates throughout the lifecycle. Witness anchors provide the continuous evidence record both examiners expect for post-market monitoring.