The EU AI Act and the 2 August 2026 deadline

The regulation entered into force on 1 August 2024, but its obligations apply in stages over several years. That is deliberate — and the reason 2 August 2026 is the date most organizations should have in their calendar:

• From 2 February 2025: the prohibited practices (unacceptable risk) apply — certain AI uses are banned.
• From 2 August 2025: obligations for providers of general-purpose AI (GPAI) models, plus parts of the governance structure.
• From 2 August 2026: the core obligations for high-risk AI systems under Annex III, and the penalty framework, become fully applicable.
• From 2 August 2027: obligations for high-risk systems that act as a safety component of already-regulated products (Annex I).

For the common case — AI used independently in one of the high-risk areas — 2026 is the decisive deadline. 2027 mainly concerns products that already fall under sectoral safety regulation.

The following become particularly relevant on that date:

• Requirements for many high-risk AI systems (including those under Annex III of the regulation).
• Transparency, documentation and oversight duties for providers and deployers.
• The penalty framework — fines up to €15M or 3% of global annual turnover, whichever is higher.
• National market surveillance, competences and the authority structure across the member states.

The EU AI Act does not regulate “AI” across the board, but by risk. Before thinking about obligations, classify your applications into one of four classes:

Unacceptable risk

Prohibited practices — such as manipulative behaviour or certain forms of social scoring. These systems have been banned since February 2025.

High risk

Systems in sensitive areas such as critical infrastructure, employment, credit scoring, education or justice. Here the full catalogue of obligations applies — this is the class 2026 is about.

Limited risk (transparency)

For example chatbots or AI-generated content. The core duty is transparency: users must know they are interacting with an AI or that content is AI-generated.

Minimal risk

The majority of today’s applications. Largely free of new obligations — voluntary codes of conduct remain possible.

Crucially: the same technology can fall into different classes depending on its purpose. It is not the model that determines the classification, but the concrete use case.

If an application falls into the high-risk class, the regulation requires a whole bundle of measures. In short, it is about making the system controllable and provable across its entire lifecycle:

• A risk-management system across the whole lifecycle — continuous, not a one-off check.
• Data governance: quality, relevance and representativeness of the data used.
• Technical documentation that makes it possible to demonstrate conformity.
• Automatic logging of events during operation — the basis of traceability.
• Transparency and provision of information so deployers can use the system appropriately.
• Effective human oversight — a person must be able to intervene.
• Appropriate accuracy, robustness and cybersecurity.

These requirements cannot be bolted on afterwards. They assume you have insight into your system — exactly what is often missing with a closed third-party service.

The regulation distinguishes roles. A common misconception is that responsibility lies with the model provider alone:

• Providers carry the main burden: conformity assessment, technical documentation, conformity marking.
• Deployers — that is you, when you run AI in production — have their own duties: use in line with the instructions, ensuring human oversight, monitoring operation and keeping the logs the system generates.

So anyone running a high-risk system cannot rely entirely on the provider — the logging and oversight duties in particular sit with you. And you can only meet them if you own the logs.

Fines are tiered and follow the higher of a fixed amount and a share of turnover:

• Up to €35M or 7% of global annual turnover — for breaches of the prohibited practices.
• Up to €15M or 3% — for breaches of the other obligations (including those for high-risk systems).
• Up to €7.5M or 1% — for incorrect or incomplete information to authorities.

Smaller companies benefit from caps that take their size into account. The point stands: the order of magnitude is chosen so that non-compliance becomes a board topic, not an IT detail.

The regulation demands traceability: who used which model in which version, and under whose oversight? With public cloud AI services you often hit a wall on exactly these questions:

• The audit trail ends at the provider’s system boundary — what happens behind it, you cannot see.
• Model versions can change without you controlling it or documenting it completely.
• Human oversight and intervention are shaped by the provider’s product, not by you.
• Data sovereignty: especially sensitive data should not leave the EU legal area — a provider outside that legal order cannot guarantee this.

None of these questions is unsolvable — but they are far easier to answer when the infrastructure is yours.

On-premise operation is not an end in itself — it delivers the evidence an audit demands, because everything stays in the house:

• Build an inventory: which AI systems are in use — and which fall into the high-risk category?
• Clarify roles: for a given system, are you the provider, the deployer, or both? Which duties follow?
• Review data flows: which data leaves your jurisdiction with cloud AI?
• Ensure traceability: is there a complete audit trail per request — and do you own it?
• Document human oversight: who can intervene, how, and is it evidenced?
• Record model provenance: which models, in which version, with which approval?
• Anchor governance: put roles, permissions and responsibilities in writing — before the deadline, not after.