AI Act: what applies on 2 August 2026, and what does not

On 2 August 2026 the AI Act becomes broadly applicable. The specific obligations for high-risk AI systems are not part of that: they arrive later. It is exactly this distinction that is missing from many project plans, where 2 August 2026 stands as the one deadline.

The two later dates come from Regulation (EU) 2026/1744; for the Annex I systems, 2 August 2027 was originally foreseen. For MedTech the 2028 route is the relevant one, but only under the condition named above: without third-party conformity assessment for product conformity, that date does not apply, and not every AI-based medical device is subject to it.

Article 4 deserves a note, because the obligation is older than all of these postponements and was still recast: it has applied since 2 February 2025, and its current wording since 27 July 2026. What it requires are measures that support the development of AI literacy among the people involved with AI systems, taking into account their prior knowledge, experience, training and the context of use. It expressly does not require guaranteeing a particular level for any individual; the earlier version required using best efforts to ensure a sufficient level.

What actually becomes applicable on that day

What comes first in practice on that day are the transparency obligations under Article 50. Here it matters whether a company is the provider or the deployer of the AI system in question, and what specific function that system performs:

• Providers of AI systems intended to interact directly with natural persons must ensure that the persons concerned can recognise that they are interacting with an AI system, unless this is already obvious.
• Providers of AI systems that generate synthetic audio, image, video or text content must, as a rule, ensure that those outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated.
• Deployers of emotion recognition or biometric categorisation systems must inform the persons concerned that the system is being used.
• Deployers must disclose deepfakes as such.
• Deployers must label certain AI-generated or manipulated text where it is published to inform the public on matters of public interest. Labelling is not required where the text has undergone genuine expert review or editorial control and a person or organisation holds editorial responsibility.

For AI systems that were already placed on the market before 2 August 2026, a limited transitional period until 2 December 2026 applies solely to the technical marking and detectability obligation for synthetic content under Article 50(2).

For a hospital with a patient chatbot, 2 August 2026 becomes immediately relevant, because patients must be able to recognise that they are interacting with an AI system. Labelling or disclosure obligations may also arise for AI-generated content used publicly. AI-assisted preliminary findings, by contrast, do not fall under Article 50 by virtue of their diagnostic function alone; what matters is whether one of the interaction or content functions governed there is additionally present.

For the specific high-risk requirements applying to an AI-based diagnostic system assessed under the MDR or IVDR with the involvement of a notified body, 2 August 2028 is in principle the decisive date. Independently of that, individual transparency obligations under Article 50 may already be relevant from 2 August 2026.

Why the postponement is no reason to pause the project

The obvious conclusion to draw from the postponement is: we will pick the topic up later. Three reasons argue against it.

First, the postponement does not change the regulatory classification of the system. Whether an AI system is to be classified as high-risk follows from Article 6 of the AI Act. For medical devices and IVDs, what matters in particular is whether the AI system is itself a regulated product or a safety component of one, and whether third-party conformity assessment is required for product conformity. Anyone who does not carry out that classification today gains no regulatory clarity from the postponement.

Second, the AI Act requirements are not meant to be dealt with only after the MDR or IVDR conformity assessment has been completed. They have to be integrated into product development, the quality management system, the technical documentation and the existing conformity assessment. Anyone who adds the AI-specific requirements only afterwards risks substantial rework on documents, processes and evidence that already exist.

Third, the data basis is often the real bottleneck. Missing evidence on the origin, selection, quality, representativeness or preparation of training, validation and test data may later no longer be fully and reliably reconstructable. What then exists is not merely a documentation gap, but potentially a gap in the data basis that can no longer be closed.

What to do now

• Record the regulatory classification of the AI system in writing, including the reasoning. It forms the basis for determining which requirements and deadlines apply.
• Embed the AI literacy obligation, which has applied since 2 February 2025, in the organisation: identify the roles affected, define the competence required, and document suitable training or qualification measures.
• Review your own AI portfolio for which systems and use cases fall under Article 50. Alongside your own products, this must also cover systems purchased or used internally.
• Plan implementation backwards from the legal date of application and from the next relevant conformity assessment. The statutory deadline remains binding; the notified body's processing and assessment times have to be planned in addition and agreed early.

Entourage supports manufacturers and healthcare providers with the classification of their AI systems, with dovetailing AI Act requirements into an ongoing MDR or IVDR procedure, and with the question of which obligations already apply today.