The AI Act set to be amended again even before many of its provisions have come into effect. We explain what the planned reform of the AI Act entails, the potential implications for businesses, and why companies should be cautious about relying on newly suggested timelines.
Omnibus IV – the initiative for competitiveness in the digital sector
The digital omnibus for AI is part of a series of amendments proposed by the European Commission to reduce bureaucracy. The aim is to ease the burden on small and medium-sized enterprises in particular.
In addition to the reform of the General Data Protection Regulation, the European Commission will also be amending the AI Act.
Note: The Commission’s draft amendment of 19 November 2025 has already been significantly amended during the European Union’s legislative process and is currently before the European Parliament.
Why should the AI Act be amended?
Firstly, it is unusual that the amendment to the AI Act is to take place before almost all existing provisions have become applicable. The initiative to amend the AI Act is therefore being taken without individual issues having been clarified in court and without any practical experience in implementing the provisions or coordinating with the supervisory authorities.
The Commission, together with various Member States, is convinced that economic competitiveness in the field of AI can only be achieved if the regulatory handbrake (in the form of the AI Act) is applied later – and that this competitiveness is still within reach.
Most provisions of the AI Act are not intended to be amended; rather, their applicability is simply being postponed. The assumption is that, during the additional one and a half years without regulation, companies will emerge or grow that can compete with the investment-heavy (U.S.) tech giants and will then be able to meet the compliance requirements without economic disadvantages. In addition, smaller companies are not supposed to have their growth hindered by compliance obligations at all.
Alongside the easing of certain training requirements, some obligations for high-risk AI systems are also being relaxed. Beyond that, the amendment streamlines many procedures for supervisory authorities and market surveillance and creates new opportunities for AI regulatory sandboxes.
The proposed amendments to the AI Act
The AI Act will see a range of provisions amended and supplemented. On the one hand, this includes an expansion of certain provisions to cover companies belonging to the SME group (micro, small and medium-sized enterprises) and SMCs (small mid-cap enterprises). The aim is to ensure that extensive compliance obligations apply only to large companies, and that the economic development of smaller and now also medium-sized or mid-cap companies (fewer than 750 employees and an annual turnover not exceeding EUR 150 million) is not hindered.
For example, the simplification of technical documentation and the design of quality management systems is planned to be explicitly extended to SMCs. The Commission is expected to issue guidelines for the simplified quality management system.
SMEs and mid-cap companies are also to be taken into account when setting administrative fines. For these companies, fines should no longer be imposed based on whichever amount is higher – between the percentage of annual turnover and the fixed amounts in Art. 99(3)–(5) AI Act (ranging from EUR 7.5 to 35 million) – but instead based on whichever of the two is lower.
On the other hand, the deadlines for the applicability of various provisions are being pushed back by up to one and a half years.
- For high-risk systems as defined in Annex III (AI used in recruitment processes, credit checks, and law enforcement), the application of the requirements will be postponed from 2 August 2026 to 2 December 2027, i.e. by approximately one and a half years.
- For high-risk systems as defined in Annex I (AI integrated into specific products such as medical devices or motor vehicles), the application of the requirements is postponed from 2 August 2027 to 2 August 2028, i.e. by approximately one year.
For high-risk AI systems under Art. 6(1) of the AI Act and Annex I, the reformed Art. 60a and Art. 111 AI Act would apply.
- In particular, Art. 111 AI Act has far-reaching implications. It provides that such systems which are placed on the market and put into service before the date on which the requirements become applicable must comply with the requirements of the AI Act only if they undergo a substantial modification after that date. For public authorities using such systems, the long implementation period until 2 August 2030 continues to apply.
- With the new Art. 60a, high-risk AI systems may also under certain circumstances be tested outside AI regulatory sandboxes. This is subject to the conditions of a “Real-World Testing Agreement” concluded between the Member States and the Commission.
In addition, providers are no longer required to register when they classify a system listed in Annex III as not being a high-risk AI system. However, under Art. 6(4) of the proposed amendment, this classification must be documented before the service is offered or put into operation.
Providers of AI systems capable of generating artificial audio, image, video, or text content and that were already made available before 2 August 2026 must comply with the labelling obligation for AI-generated content under Art. 50(2) AI Act by 2 November 2026.
Furthermore, the obligation to ensure AI competence will be relaxed (Art. 4 AI Act). Providers and operators of AI are no longer required to ‘ensure’ such competence through measures such as training or further education, but rather to ‘support the improvement’ of it. Furthermore, this obligation does not extend to guaranteeing a specific level of AI competence for individual persons. The Commission is to provide guidance on the practical implementation of these obligations.
The new Art. 4a incorporates, first, the current Art. 10(5) AI Act. This provision constitutes an exception allowing the processing of special categories of personal data or other sensitive data (health data, political opinions, etc.) for the purpose of detecting biases or distortions in AI systems and correcting them in high-risk systems if this is absolutely necessary. This introduces a clear purpose limitation for such processing. Secondly, paragraph 2 expands this exception for deployers of high-risk AI systems where the processing is necessary and proportionate for the prevention of biases, insofar as such biases are likely to affect the health or safety of individuals, have a negative impact on fundamental rights, or lead to discrimination prohibited by EU law. The proposal extends the scope for processing special categories of personal data for the purpose of detecting and preventing bias in high-risk AI systems. At the same time, this processing is subject to strict conditions such as necessity and purpose limitation.
AI regulatory sandboxes, which enable new AI applications to be tested under real-world conditions already during development, may, under the proposed Art. 57(3a), also be established by the AI Office. In addition, the real-world conditions governing such tests are to be set out in a dedicated document.
In addition, the list of prohibited practices in the field of AI (Art. 5) is being expanded. AI systems that are capable of generating realistic sexual or intimate videos and photos of natural persons without their consent (so-called deepfakes) will be prohibited, unless effective safeguards have been put in place to prevent this.
Outlook
The planned reform eases several requirements of the AI Act and, in particular, broadens the group of companies that will be permitted to implement these requirements in a less comprehensive manner. However, the amendments primarily aim to give companies additional time to comply with the obligations.
Affected companies should adjust their existing implementation timelines only once the reform is certain. The remaining time should then be used to establish robust AI management systems to ensure reliable compliance.
