EU AI Act - Annex I High-Risk Systems (2 Aug 2028)

High-risk obligations for AI that is a safety component of products covered by EU harmonisation legislation (machinery, medical devices, vehicles, aviation). Deferred from August 2027 by the Digital Omnibus on AI.

JurisdictionEuropean Union
CategoryAI Regulations
StatusUpcoming
Effective date
Latest development

Analysis

The EU AI Act’s Annex I product-safety obligations apply from 2 August 2028 under the 2026 AI Omnibus amendment, according to the European Commission. They concern AI systems that are safety components of regulated products—or are themselves regulated products—where the underlying product must undergo third-party conformity assessment. The controlling legal text is Regulation (EU) 2024/1689, as amended by the AI Omnibus.

Key Requirements

1. Determine whether the system is an Annex I high-risk system

An AI system is high-risk under Article 6(1) when both conditions apply:

  • It is intended to function as a safety component of a product, or is itself a product, covered by EU harmonisation legislation listed in Annex I.
  • The product must undergo a third-party conformity assessment before being placed on the EU market or put into service.

These requirements are set out in Article 6(1) of the EU AI Act and Annex I.

Annex I covers product legislation involving, among other areas:

  • Machinery
  • Toys
  • Recreational craft
  • Lifts
  • Equipment and protective systems for explosive atmospheres
  • Radio equipment
  • Pressure equipment
  • Personal protective equipment
  • Motor vehicles and related vehicle systems
  • Civil aviation and aviation safety equipment
  • Marine equipment
  • Rail interoperability
  • Medical devices and in-vitro diagnostic medical devices

The precise scope depends on the legislation listed in Annex I and whether the relevant product requires third-party assessment. See Annex I of Regulation (EU) 2024/1689 and the Commission’s AI regulatory framework overview.

2. Apply the high-risk requirements in Chapter III

Providers of covered high-risk AI systems must generally establish a documented risk-management system, use appropriate data and data-governance controls, prepare technical documentation, maintain records, provide information to deployers, enable appropriate human oversight, and meet accuracy, robustness, and cybersecurity requirements.

The principal obligations are in Chapter III, Section 2 of the AI Act, including:

  • Risk management: establish and maintain a continuous, documented risk-management process.
  • Data and data governance: training, validation, and testing datasets must satisfy the Act’s quality and governance requirements where applicable.
  • Technical documentation: maintain documentation demonstrating conformity and describing the system’s design, development, evaluation, and operation.
  • Record-keeping: design the system to generate and retain logs appropriate to its intended purpose.
  • Transparency and information: provide deployers with information needed to understand capabilities, limitations, operating conditions, and necessary oversight.
  • Human oversight: design the system so that humans can oversee, interpret, interrupt, or override its operation as appropriate.
  • Accuracy, robustness, and cybersecurity: achieve appropriate performance and resilience throughout the system lifecycle.

The requirements are specified in Articles 8–15 and summarized by the European Commission in its guidelines for providers and deployers of high-risk AI systems.

3. Integrate AI conformity assessment with product conformity assessment

For Annex I systems, AI conformity assessment must be coordinated with the conformity-assessment process required by the underlying product legislation. The provider must determine the applicable assessment route, technical documentation, quality-management controls, notified-body involvement, and EU declaration requirements.

Relevant provisions include:

Organizations should avoid treating the AI Act assessment as a separate software-only exercise: the legally relevant conformity route may be governed substantially by machinery, medical-device, vehicle, aviation, rail, or another sectoral framework.

4. Establish a quality-management system

Providers must operate a quality-management system addressing regulatory compliance, design controls, development procedures, testing, data management, risk management, post-market monitoring, incident handling, and corrective action.

This obligation is established by Article 17 of the AI Act. Useful implementation references include:

5. Register systems and cooperate with authorities

Providers must meet applicable registration, monitoring, incident-reporting, corrective-action, and cooperation obligations. The AI Act also assigns responsibilities to importers, distributors, deployers, authorized representatives, notified bodies, and market-surveillance authorities.

Relevant provisions include:

Compliance Challenges

Scope and classification

  • Determining whether AI is a “safety component” is often difficult. The relevant question is not merely whether the software improves safety, but whether its failure or malfunction can affect health or safety and whether it is integrated into a product within Annex I’s scope. The classification test is in Article 6, while product coverage is in Annex I.
  • The same AI function may have different regulatory treatment depending on deployment. An AI model used for a non-safety administrative function may not be classified in the same way as the model embedded in a vehicle-control, medical-device, machinery, or aviation-safety function. This follows from the Act’s product-specific classification structure in Article 6.

Overlap with sectoral product regimes

  • Organizations must reconcile AI Act obligations with existing product-safety, medical-device, vehicle, aviation, machinery, and cybersecurity requirements. The Act expressly links Annex I systems to Union harmonisation legislation rather than replacing those regimes; see Article 6 and Annex I and the Commission’s AI Act framework page.
  • Evidence duplication is a common risk: engineering teams may maintain separate AI, functional-safety, cybersecurity, and product-quality documentation that describes the same controls inconsistently. A unified evidence architecture is therefore preferable, using the AI Act’s technical-documentation requirements alongside the relevant product regime.

Safety validation and data limitations

  • AI systems may change behavior when exposed to new operating conditions, data distributions, sensor configurations, or adversarial inputs. This complicates validation of accuracy, robustness, foreseeable misuse, and fail-safe behavior under Article 15.
  • Medical, vehicle, industrial, and aviation applications frequently have limited failure data because serious failures are rare or unacceptable to induce experimentally. Organizations must therefore combine simulation, scenario testing, historical data, expert review, stress testing, and post-market monitoring. The AI Act requires lifecycle risk management under Article 9 and monitoring under Article 72.

Supply-chain and responsibility allocation

  • Responsibility may be distributed among the AI provider, product manufacturer, component supplier, importer, distributor, and deployer. Contractual arrangements must allocate access to training-data information, model versions, test evidence, incident data, change notices, and corrective actions. The relevant obligations are in Articles 24–27.
  • Third-party and open-source dependencies can create traceability problems, particularly where a product manufacturer embeds a model supplied by another organization. Providers should establish formal supplier controls and technical documentation requirements under Article 17 and Article 25.

Standards are still developing

  • Harmonized standards are important because compliance with cited standards can provide a presumption of conformity for relevant requirements, but organizations cannot defer all preparation until standards are finalized. The Commission’s standards and support activities are described on the AI Act implementation page, while the AI Act’s conformity framework is in Article 40.

Implementation Best Practices

1. Build an inventory and classification register

Create a controlled register containing:

  • AI system name, version, owner, supplier, and deployment location.
  • Product into which the system is integrated.
  • Intended purpose and safety function.
  • Applicable Annex I legislation.
  • Whether third-party conformity assessment is required.
  • Applicable AI Act classification and transition date.
  • Product lifecycle status and planned changes.

Use Article 6, Annex I, and the Commission’s high-risk-system guidance as the classification baseline.

2. Map one integrated compliance matrix

Map every AI Act requirement to:

  • An accountable owner.
  • A product-safety requirement.
  • A document or evidence artifact.
  • A test or control.
  • A release gate.
  • A monitoring metric.
  • A corrective-action process.

A useful structure is to map Articles 8–15 to existing quality, functional-safety, medical-device, aviation, vehicle, and cybersecurity processes rather than creating an isolated AI compliance program.

3. Establish lifecycle risk management

Implement documented procedures for:

  • Hazard and risk identification.
  • Foreseeable misuse and reasonably foreseeable operating conditions.
  • Data and model-risk assessment.
  • Human-oversight design.
  • Residual-risk acceptance.
  • Change-impact assessment.
  • Decommissioning and rollback.

The legal basis is Article 9, supported by ISO/IEC 23894 and the NIST AI RMF.

4. Create an evidence and traceability architecture

Maintain version-controlled records connecting:

  • Requirements to design decisions.
  • Training and validation data to model versions.
  • Model versions to test results.
  • Test results to release decisions.
  • Incidents to corrective actions.
  • Product changes to renewed conformity assessments.

This supports the AI Act’s technical-documentation requirement, automatic logging requirement, and quality-management requirement.

5. Implement model, software, and product change control

Define which changes require:

  • Re-testing.
  • Re-training.
  • Reassessment of risks.
  • Notification to a notified body.
  • An updated EU declaration of conformity.
  • A new or amended product certification.
  • Customer or regulator notification.

The AI Act requires lifecycle controls and post-market processes under Articles 16–17 and Article 72.

6. Use recognized management and engineering frameworks

Useful resources include:

These frameworks support implementation, but certification to a voluntary standard does not by itself establish compliance with every mandatory AI Act or sectoral product requirement.

Recent Updates

AI Omnibus transition amendment

The European Commission states that the AI Omnibus changed the application date for high-risk AI systems embedded in regulated physical products to 2 August 2028. The Commission specifically identifies products such as machinery, toys, lifts, and similar regulated products. See the official AI Omnibus announcement and the Commission’s updated AI Act overview.

Distinction between Annex I and Annex III

The extended date applies to Annex I product-related systems. It should not be confused with high-risk systems classified under Annex III, for which the Commission identifies 2 December 2027 as the application date under the updated framework. See the Commission’s high-risk-system guidance and AI Act regulatory framework.

Original AI Act timeline

The original AI Act entered into force on 1 August 2024, with the general high-risk product provisions

Recent developments

  • — The European Commission published draft **guidelines for providers and deployers of high-risk AI systems**, launching a targeted stakeholder consultation (open until 23 July 2026) that will shape how obligations for Annex I high-risk systems are interpreted and implemented.[8][1] (source)
  • — Analysis of the EU’s “AI Omnibus” amendments confirms that **Annex I high-risk AI systems** (AI embedded in regulated products) now have their Chapter III obligations deferred to **2 August 2028**, extending the previous deadline and giving product-based AI sectors an extra year to comply.[9][1][6] (source)
  • — A detailed compliance update explains that, under Regulation (EU) 2026/1744, **high-risk AI systems in Annex I** will see core Chapter III obligations apply from **2 August 2028**, outlining practical steps for providers and deployers to prepare for the extended timeline and integrate AI Act requirements with existing product-safety regimes.[6][1] (source)
  • — Client alert reports that the EU formally approved **delays and amendments** to certain AI Act obligations, including postponing the effective date for **Annex I high-risk product/safety-component AI systems** from 2 August 2027 to **2 August 2028**, impacting sectors such as medical devices, radio equipment, toys and other regulated products.[13][1] (source)
  • — The European Commission’s official AI Act page confirms that, following the political agreement on the “AI Omnibus” simplification package, **rules for high-risk AI systems embedded into regulated products in Annex I** benefit from an extended transition period until **2 August 2028**, distinguishing them from Annex III high-risk use cases which apply earlier.[1][3] (source)
  • — News coverage of the EU “omnibus” deal reports that lawmakers agreed to **delay implementation of high-risk AI rules**, with stand‑alone high-risk systems (Annex III) now applying from 2 December 2027 and **Annex I high-risk systems embedded in products** applying from **2 August 2028**, prompting companies to reassess compliance roadmaps.[14][1] (source)
  • — Cloud Security Alliance research note analyses the **enterprise readiness gap** for high-risk AI compliance, noting the binding high-risk obligations in August 2026 for many systems while emphasizing that **Annex I product-based high-risk AI** faces a longer runway to 2028 but still demands early investment in governance, monitoring and incident reporting.[5][6] (source)
  • — Industry guidance for companies on high-risk AI explains phased preparation steps for the AI Act, highlighting that **Annex I high-risk AI systems embedded in regulated products** have a later application date and require early inventory, risk classification, technical documentation and conformity assessment aligned with the deferred obligations.[7][11] (source)
  • — A technical readiness guide details how enterprises should prepare for AI Act **high-risk obligations**, distinguishing between earlier obligations for Annex III systems and the later compliance window for **Annex I product-route high-risk systems**, and stressing the need to build risk management and documentation capabilities ahead of the 2028 enforcement date.[12][6] (source)
  • — An August 2026‑updated comprehensive summary of the EU AI Act explains the classification of **high-risk AI systems under Annex I** and confirms that their obligations will **apply from 2 August 2028**, prompting software and product engineering teams to align lifecycle processes, quality assurance and human‑oversight mechanisms with the delayed but stringent requirements.[11][2] (source)

Related regulations

Put it into practice

Browse all regulations · Compliance deadlines · Latest updates