Menu
© 2026 World Compliance Association

27/07/2026

Europe amends the AI Act and postpones some of the obligations for high-risk artificial intelligence systems

The new Digital Omnibus Regulation on Artificial Intelligence extends certain deadlines, simplifies some obligations and strengthens oversight, but does not establish a general moratorium on the AI Act

Europe amends the AI Act and postpones some of the obligations for high-risk artificial intelligence systems

By Iván Martínez, CEO of Intedya

The European Union has just introduced the first major reform of the Artificial Intelligence Regulation since its adoption. The Official Journal of the European Union has published Regulation (EU) 2026/1744 of 8 July 2026, known as the Digital Omnibus on Artificial Intelligence, which amends the AI Act, the European Civil Aviation Regulation and the Machinery Regulation. The new legislation aims to simplify the implementation of the European artificial intelligence framework, reduce duplication and facilitate compliance without compromising the protection of health, safety and fundamental rights. The reform comes after months of debate over the lack of harmonised standards, technical specifications, guidance and bodies equipped to enforce certain obligations under the Artificial Intelligence Regulation. The most notable development is the postponement of the requirements applicable to certain high-risk AI systems. However, it would be a mistake to interpret this amendment as a general suspension of the AI Act.

Europe has granted more time to comply with certain high-risk obligations, but the Artificial Intelligence Regulation remains fully in force and many of its provisions are already enforceable.

The Artificial Intelligence Regulation came into force on 1 August 2024 and set out a phased implementation timetable. During the first few months of implementation, European institutions and economic operators have identified significant difficulties: delays in the development of harmonised standards, a lack of sufficiently detailed technical criteria, national supervisory structures that are still incomplete, and the risk of duplication with other sector-specific legislation.  The Regulation itself acknowledges that these delays could result in a greater regulatory burden than initially anticipated and hinder the simple, uniform and effective application of the AI Act. However, the reform emphasises that simplification measures must not reduce the level of protection of health, safety or fundamental rights.  The Digital Omnibus therefore aims to address implementation issues without altering the core philosophy of European regulation: artificial intelligence that is people-centred, reliable and managed in accordance with a risk-based approach.

The new timetable for high-risk AI systems

The most significant change concerns the dates of application of the obligations set out for artificial intelligence systems classified as high-risk.

The new timetable distinguishes between two main categories.

2 December 2027: high-risk systems listed in Annex III

The obligations set out in Chapter III, Sections 1, 2 and 3, will apply from 2 December 2027 to systems classified as high-risk in accordance with Article 6(2) and Annex III of the AI Act.

This group includes certain systems used in particularly sensitive areas, amongst them:

  • Biometrics.
  • Education and vocational training.
  • Recruitment, hiring and management of workers.
  • Access to essential private and public services.
  • Credit assessment and certain financial services.
  • Migration, asylum and border control.
  • Administration of justice.
  • Democratic processes and certain uses in the public sector.

Not all systems used in these areas are automatically high-risk. Classification will depend on their purpose, context, operation and the exceptions set out in the Regulation itself.

2 August 2028: high-risk systems linked to regulated products

The obligations will apply from 2 August 2028 to systems classified as high-risk in accordance with Article 6(1) and Annex I.

These are, primarily, AI systems that constitute a product regulated by specific European legislation or that function as a safety component of such a product.

This category may include, amongst others, medical devices, vehicles, lifts, personal protective equipment, toys, radio equipment and other products subject to European harmonisation legislation and, in certain cases, to conformity assessment.

The new Regulation expressly sets out these two dates: 2 December 2027 for high-risk systems listed in Annex III and 2 August 2028 for systems linked to products listed in Annex I.

There is no general moratorium on the AI Act

The postponement affects certain requirements for high-risk systems, but does not bring the Artificial Intelligence Regulation to a standstill.

This is probably the main point of confusion that organisations need to avoid.

The AI Act continues to apply in accordance with its timetable, and the provisions relating to, amongst others, the following remain relevant:

  • Prohibited artificial intelligence practices.
  • Obligations of providers of general-purpose AI models.
  • Governance and supervisory authorities.
  • AI literacy.
  • Transparency of certain systems.
  • Identification and labelling of synthetic content.
  • Contractual responsibilities between technology providers.
  • Obligations that have not been expressly deferred.
  • Penalty regime and market surveillance measures.

Companies must not interpret the new timetable as authorisation to halt their compliance projects.

On the contrary, the additional time should be used to complete inventories, classify systems, assign responsibilities, review contracts, establish controls and develop a proper governance structure.

AI literacy does not disappear, but its wording changes

One of the most significant changes concerns Article 4 of the AI Act, relating to artificial intelligence literacy. The previous wording required suppliers and those responsible for deployment to ensure a sufficient level of literacy amongst their staff. The new wording requires suppliers and those responsible for deployment to take measures to support the promotion of AI literacy amongst those responsible for the operation and use of the systems. These measures must be tailored to:

  • The technical knowledge of the individuals concerned.
  • Their experience, education and training.
  • The intended context of use.
  • The AI systems used.
  • The individuals or groups on whom the systems may be used.

The regulation clarifies that the obligation does not require ensuring that every individual achieves a specific level of knowledge. However, this does not remove the need to educate, raise awareness amongst and train staff. The new Article 4 maintains a specific organisational obligation: to adopt and be able to demonstrate reasonable and proportionate measures for AI literacy.  Therefore, cancelling training programmes would be a decision that would be difficult to justify. The sensible approach would be to review their scope and link them to risk, to each person’s role and to the actual use of the systems.

New prohibited practices: intimate images and child sexual abuse material

The reform is not limited to simplifying obligations. It also expressly expands the list of prohibited practices. The new Regulation introduces a ban on placing on the market, putting into service or using certain AI systems that generate or manipulate realistic images, videos, audio recordings or other content of an intimate nature relating to an identifiable person without their consent.  Certain systems designed to generate or manipulate material relating to child sexual abuse are also prohibited. The prohibition applies both to systems specifically designed to produce such content and to those in which such an outcome is reasonably foreseeable and reproducible, and where adequate safeguards are not in place. For those responsible for deployment, the ban applies when the system is used for the purpose of generating or manipulating such material. These new bans will come into force on 2 December 2026. The amendment is a direct response to the rise of ‘nudification’ tools, non-consensual synthetic intimate content and other forms of abuse facilitated by generative AI.

Changes to transparency obligations regarding synthetic content

The reform also introduces a transitional period for certain providers of generative AI systems. Providers of systems – including those based on general-purpose models – that generate synthetic audio, images, video or text, and which were already on the market before 2 August 2026, must take the necessary measures to comply with the labelling obligations set out in Article 50.2 by 2 December 2026 at the latest. This obligation is particularly relevant for platforms, developers of generative tools and companies that market or integrate systems capable of producing artificial content. Organisations must review not only the content visible to the user, but also the technical mechanisms for labelling, traceability, tagging and detection.

Greater proportionality for SMEs and mid-cap companies

The Regulation extends several simplification measures initially intended for micro-enterprises. SMEs, including start-ups, will be able to comply with certain elements of the quality management system required of high-risk system providers in a simplified manner. Simplified technical documentation forms are also provided for, along with a proportionate application of certain requirements taking into account the organisation’s size and resources.

The reform also extends certain benefits to so-called ‘small mid-caps’ – companies that exceed the SME thresholds but still face similar difficulties in terms of costs and administrative burden. These measures do not constitute an exemption from compliance. Simplification must maintain the required level of protection and the necessary rigour to demonstrate the conformity of high-risk systems.

A more precise definition of ‘security component’

The Digital Omnibus amends the definition of a safety component, a key concept for determining when a system integrated into a product should be classified as high-risk. The new regulation clarifies that a system fulfils a safety function when its intended purpose is to prevent or mitigate risks to the health and safety of persons or property. Systems used solely for the following purposes shall not be considered safety components:

  • User assistance.
  • Performance optimisation.
  • Service efficiency.
  • Automation.
  • Convenience.
  • Quality control not related to safety.

However, where a system failure or malfunction could endanger health or safety, it may be considered a safety component. This clarification will reduce overly broad classifications, but will require the intended purpose, failure modes and reasonably foreseeable consequences to be carefully documented.

Existing AI systems: the importance of significant modifications

The new Regulation clarifies the treatment of high-risk systems placed on the market or put into service before the new application dates. As a general rule, the AI Act will apply to these systems when, after the relevant date, they undergo significant changes to their design. The regulation further clarifies that, where a unit of a specific type and model has been lawfully placed on the market, other units of the same type may benefit from the transitional regime provided their design remains unchanged. This makes it particularly important to monitor and document:

  • The date of first placing on the market.
  • System versions.
  • Changes to algorithms and models.
  • Changes to data or functionalities.
  • Extensions of intended use.
  • Updates that may be considered substantial.

An apparently routine update could trigger the full application of obligations if it significantly alters the system’s design, purpose or behaviour.

Further contractual obligations within the technology supply chain

The reform also strengthens cooperation between the various operators in the supply chain. Where a third party supplies models, tools, services, components or processes used in a high-risk system, there must be a written agreement setting out the information, capabilities, technical access and support necessary to enable the end supplier to fulfil its obligations.

These agreements must cover aspects such as:

  • Access to technical documentation.
  • Information on limitations and failure modes.
  • Testing and validation capabilities.
  • Assistance with conformity assessment.
  • Incident management.
  • Cooperation with the authorities.
  • Allocation of responsibilities.
  • Communication of relevant changes.

Generic technology contracts will be insufficient in many cases. Organisations will need to incorporate specific clauses relating to compliance with the AI Act.

Processing of sensitive data to detect and correct bias

The reform introduces a new legal basis for the exceptional processing of special categories of personal data where strictly necessary to detect and correct bias. This possibility is subject to stringent conditions, including:

  • That it is not possible to achieve the objective using synthetic or anonymised data.
  • The implementation of advanced security and privacy measures.
  • Pseudonymisation.
  • Strict access controls.
  • A ban on disclosure to third parties.
  • Deletion of the data once the bias has been corrected.
  • Documentation of the necessity of the processing.

The measure does not create a general obligation to use sensitive data to assess bias, but provides an exceptional legal basis where such processing is essential. Organisations must coordinate these activities with their data protection systems, impact assessments, security controls and governance procedures.

Greater powers for the European Artificial Intelligence Office

Another major change is the strengthening of the AI Office. The Office will have exclusive jurisdiction over certain systems based on general-purpose AI models, particularly where the model and the system have been developed by the same provider or by companies belonging to the same group. It will also assume jurisdiction over systems integrated into online platforms or search engines considered to be of very large scale.

Its powers include:

  • Requesting information.
  • Accessing systems and documentation.
  • Carrying out remote or on-site inspections.
  • Examining data, books and records.
  • Demanding explanations.
  • Order corrective measures.
  • Make corporate commitments legally binding.
  • Impose fines and penalty payments.
  • Withdraw or restrict systems from the market.

The Office may impose daily penalty payments which, in certain cases, may amount to 5 per cent of average daily revenue or global annual turnover for the previous financial year. The reform thus confirms that regulatory simplification does not equate to a reduction in supervisory capacity.

Controlled testing and experimentation environments under real-world conditions

The Regulation expands the possibilities for testing AI systems in controlled environments and under real-world conditions. Member States must ensure that at least one operational controlled testing environment is in place by 2 August 2027 at the latest. The AI Office may also establish a Europe-wide controlled testing environment, with priority access for SMEs, start-ups and small mid-cap companies.

The scope for conducting real-world testing of certain high-risk systems is also extended, provided that adequate plans, oversight mechanisms and safeguards are put in place.

These measures are intended to facilitate innovation, but do not remove the need to manage risks, protect personal data and safeguard the rights of participants.

The opportunity and risk for organisations

The new timetable offers additional leeway to implement the more complex requirements for high-risk systems. However, this leeway can be interpreted in two ways. An organisation may use it to delay decisions, maintain unregistered systems and rely on further extensions being granted. Or it can use it to build a robust governance framework before the obligations become fully enforceable. The second option is the only reasonable one from the perspective of compliance, risk management and corporate responsibility. Classifying a system, completing its technical documentation, establishing a risk management system, ensuring data quality, implementing human oversight, reviewing contracts and creating post-market surveillance mechanisms requires months of work; in complex organisations, it may take years.

Urgent ten-point guide to adapting to the new Digital Omnibus Act

1. Update the internal AI Act timetable

Compliance plans must incorporate the new deadlines of 2 December 2026, 2 December 2027 and 2 August 2028, whilst not overlooking obligations that are already in force.

2. Compile an inventory of all artificial intelligence systems

The inventory must include corporate tools, solutions integrated into processes, in-house developments, generative models, systems contracted from third parties and applications used informally by staff.

3. Determine the organisation’s role

It is essential to establish whether the company acts as a supplier, a party responsible for deployment, an importer, a distributor, a product manufacturer or an authorised representative.

A single company may fulfil several roles simultaneously.

4. Identify potentially high-risk systems

The classification must analyse the intended purpose, the sector, the individuals affected, the degree of autonomy, the consequences of decisions and the relationship with regulated products.

5. Distinguish between Annex III systems and those linked to products

This distinction determines the timetable, obligations and conformity assessment procedure.

6. Maintain governance and digital literacy programmes

The new wording of Article 4 does not remove the requirement for training. It calls for measures tailored to individuals’ knowledge, roles and context of use.

7. Review transparency obligations

Systems that interact with people, recognise emotions, generate synthetic content or produce artificial images, audio, video or text must be analysed.

8. Document existing systems and their modifications

It is essential to retain evidence of dates, versions, functional changes, updates and design modifications.

9. Update technology contracts

Agreements with suppliers must expressly regulate access to documentation, data, testing, technical limitations, incidents, audits and liabilities.

10. Implement an AI Act compliance system

Compliance cannot rely on isolated actions. It requires a permanent framework of policies, functions, controls, evidence, risk assessment, training, supervision and continuous improvement.

Conclusion: more time, but no less responsibility

The Digital Omnibus Regulation represents a far-reaching reform. Europe recognises that the original timetable for high-risk systems was not viable without sufficiently developed standards, guidance and assessment frameworks. For this reason, it extends certain deadlines, introduces simplifications and seeks greater coordination with sector-specific legislation. However, the reform does not repeal or suspend the AI Act. Prohibited practices remain enforceable. The obligations for general-purpose models remain in force. Transparency regarding synthetic content is advancing. AI literacy remains necessary. Oversight is being strengthened and the AI Office is being granted greater investigative and sanctioning powers. The conclusion is clear: Europe is granting more time for certain obligations relating to high-risk systems, but is not granting a general moratorium on compliance with the Artificial Intelligence Regulation. Organisations that do not yet have a governance and compliance framework for the AI Act should begin implementing one immediately.

reCAPTCHA protection