EU AI Act · 2026-07-24 · 5 min

EU AI Act: what is actually expected of you on August 2, 2026

The EU AI Act has been in force since August 2024, but enforcement of the core obligations begins on August 2, 2026. For most organizations that is the moment when “we're working on it” no longer suffices. The regulator may then ask for evidence.

Users fall under the law too, not just builders

A persistent misconception is that the AI Act only applies to parties developing AI models. The law also imposes obligations on organizations that use AI systems (“deployers”). Anyone deploying ChatGPT Enterprise, Microsoft Copilot or an AI-driven HR tool has a compliance position of their own and must be able to substantiate it.

What must demonstrably be in place

The core obligations come down to three parts:

  • A current AI register: which AI systems does your organization use, for what purpose, and with which data. Without an inventory, no other obligation can be met.
  • Risk classification per system according to the law's four categories: unacceptable, high, limited and minimal risk. The category determines which requirements apply.
  • Documentation matching the classification: from transparency statements for limited risk to DPIAs, technical documentation and human-oversight protocols for high risk.
UnacceptableBanned outright. Think social scoring and manipulative systems.
HighDPIA, technical documentation, human-oversight protocol and registration.
LimitedTransparency duty: the user must know they are dealing with AI.
MinimalNo extra obligations, but still record it in the register.
The category dictates the obligation. Most business applications land in 'limited' or 'minimal', but you still have to be able to prove it.

The fines are not symbolic: up to €35 million or 7% of global annual revenue, depending on the violation.

Why it goes wrong in practice

The problem is rarely unwillingness. It is visibility. AI usage grows organically: departments buy their own licenses, employees try free-tier tools, and nobody holds the full picture. During audits, the real number of AI systems structurally exceeds what the spreadsheet says. A manually maintained register ages faster than it gets updated.

What you can still do in the coming weeks

  • Start with an inventory across all providers. The admin APIs of OpenAI, Anthropic, Microsoft and Google let you automate most of it.
  • Classify the highest-impact systems first: customer contact, HR decisions, and anything processing personal data.
  • Record per system what already exists (DPA, ZDR status, documentation) and where the gaps are. An honest gap analysis is worth more to a regulator than an empty register.

This article is informational and not legal advice. Consult a lawyer with AI Act expertise for your specific situation.

PRODUCT
AxonGuard

AxonGuard automates the AI register, risk classification and mandatory documentation across all AI providers, from €99/month.

View AxonGuard →