PiirZ← Back to home
AI Act after 2 August 2026: what applies now and what moved to 2027
RegulationNovember 2026

AI Act after 2 August 2026: what applies now and what moved to 2027

What has applied since 2 August 2026

The AI Act (Regulation (EU) 2024/1689) enters into application in stages. Prohibited practices have been banned since 2 February 2025. Obligations for general-purpose AI models have applied since 2 August 2025.

Since 2 August 2026 two further layers are active. The first is transparency under Article 50: people must be told when they interact with an AI system, synthetic content must be marked, emotion recognition and biometric categorisation must be disclosed to the people concerned, deepfakes must be labelled. Providers of generative systems already on the market before that date have until 2 December 2026 for the marking obligation. The other duties apply in full.

The second layer is supervision. National market surveillance authorities and the European AI Office now enforce the regulation. Breaches of the transparency rules can lead to fines of up to 15 million euros or 3% of worldwide annual turnover. For prohibited practices the ceiling is 35 million euros or 7%.

What the Digital Omnibus postponed

The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It moves the obligations for the high-risk systems listed in Annex III, such as those used in recruitment, credit scoring or access to education, from 2 August 2026 to 2 December 2027. For AI embedded in the products covered by Annex I, the date moves to 2 August 2028. The same regulation simplifies documentation duties for SMEs and small mid-caps.

The content of the obligations does not change: risk management across the lifecycle, data governance, technical documentation, logging, human oversight, accuracy, robustness and cybersecurity.

What the postponement changes for organizations

Sixteen additional months do not shorten the path: classifying systems, mapping data flows, writing technical documentation and designing human oversight into processes take the same time in 2027 as in 2026. The postponement only widens the margin for correcting mistakes.

Organizations that use the postponement to start the inventory have time to correct course. Organizations that read it as a reason to wait will face the same work on a compressed timeline, while the transparency duties are already enforceable.

Compliance is an architecture problem

Meeting the AI Act takes three competences at once: interpretation of the regulation, assessment of technical risk, knowledge of how the systems are built.

A lawyer interprets the articles but rarely classifies the risk of an automated scoring model. A data scientist knows the model but not necessarily the data governance the regulation requires. A system integrator implements, but without a governance framework builds on fragile ground.

The three meet in the architecture: where data enters, which component decides, who can intervene, what is logged. Compliance designed at that level can be audited. Compliance added afterwards produces documents that describe a system other than the one in production.

How PiirZ works on it

PiirZ starts from an inventory of the AI systems in use, including those embedded in third-party software. Each system is classified against the regulation, gaps are measured against the applicable requirements, and interventions are ordered by risk and by date of application.

The result is operational governance owned by the client: documented roles, controls built into the systems, records that hold up under audit. PiirZ does not sell software and does not resell licenses.