The EU AI Act places its obligations mainly on providers and deployers of high-risk AI systems, and the dates from which they apply were moved in July 2026. Below: what the Act requires, which dates apply, and five questions internal audit can test, with the evidence for each.
What the EU AI Act asks of organisations
The Act sorts AI systems by risk. Certain practices are prohibited (Article 5). High-risk systems carry an extensive set of requirements. Certain other systems, including those that interact with people or generate content, carry transparency obligations (Article 50). For the remaining systems the Act sets few specific requirements.
The obligations sit mainly with the provider, who develops a system, or has it developed, and places it on the market or puts it into service under its own name or trademark, and the deployer, who uses it under its own authority. Personal, non-professional use falls outside that role. For high-risk systems the requirements cover seven subjects:
- a risk management system across the whole lifecycle (Article 9);
- data and data governance (Article 10);
- technical documentation (Article 11);
- record-keeping of events in logs (Article 12);
- transparency and information for the deployer (Article 13);
- human oversight (Article 14);
- accuracy, robustness and cybersecurity (Article 15).
These requirements rest primarily on the provider. Article 26 gives the deployer its own obligations, including using the system in line with the instructions for use, assigning human oversight, and keeping the logs to the extent these are under its control. Which obligations apply depends on the system's risk class and the organisation's role.
The dates that apply
Most of the AI Act has applied since 2 August 2026. Parts applied earlier: the prohibited practices and AI literacy since 2 February 2025, the rules for general-purpose AI models since 2 August 2025 (for models already on the market before that date, the final date is 2 August 2027). The requirements for high-risk systems did not take effect on 2 August 2026: the Digital Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026, entered into force on 27 July 2026 and moved them.
| Date | What applies |
|---|---|
| 2 August 2026 | Most of the Act applies, including the transparency obligations of Article 50. For the marking of generated content (Article 50(2)), existing generative systems have a transitional period until 2 December 2026. |
| 2 December 2026 | Two new prohibited practices (Article 5(1), points ba and bb), relating to non-consensual intimate imagery and to child sexual abuse material. Also the final date for machine-readable marking by generative AI systems already on the market before 2 August 2026 (Article 50(2)). |
| 2 December 2027 | Requirements for high-risk systems listed in Annex III (with a caveat: see below). |
| 2 August 2028 | Requirements for high-risk AI in regulated products under Annex I: the system is a safety component of such a product or is itself such a product, and the product requires a third-party conformity assessment (Article 6(1)). |
A use case being listed in Annex III is not the whole answer. Under Article 6(3), an Annex III system is not high-risk after all if it poses no significant risk to health, safety or fundamental rights and meets at least one of the conditions listed in that paragraph (for example a narrow procedural task or a preparatory task); this does not apply to systems that profile natural persons, and a provider relying on the exception must document its assessment (Article 6(4)). For high-risk systems placed on the market before the date of application, the requirements in principle apply only if the design is significantly changed afterwards; for systems intended for public authorities the Act has its own transitional period (Article 111). Test each system, not only the list.
This is the position on the date of publication. The Act may change again. For an audit plan or an opinion, the Official Journal and the amending regulations are what counts; the consolidated text on EUR-Lex is a reading aid with no legal force.
The postponement means the high-risk requirements cannot yet be tested as legal obligations in force. What they call for (inventory, role determination, oversight, logging) can already be assessed as controls.
Five questions internal audit can test
Each question comes with the evidence to request and a red flag. A red flag is an indication for further work; whether an obligation has been breached depends on classification and role.
- Is there an inventory of AI systems, with a risk class for each system? Evidence: the register, showing owner, purpose, supplier, risk class and the reasoning behind that class for each system. Reconcile the register with procurement records and the software list. Red flag: a register that holds only models built in-house, or a class without reasoning and a date.
- Is it recorded for each system whether the organisation is provider or deployer? Evidence: the role determination per system, the contract and the supplier's instructions for use. Red flag: a purchased system that was modified internally or is offered under the organisation's own name, with no reassessment of the role. For high-risk systems, Article 25 provides that the role can change when the organisation puts its own name or trademark on the system, makes a substantial modification, or changes the intended purpose so that the system becomes high-risk.
- Is human oversight defined, and does it work in practice? Evidence: the appointment of the person performing oversight, with authority to intervene and the training received, and a sample of decisions where the system's output was reviewed. Red flag: a sample in which the system's advice was never changed (this may point to nominal oversight, but does not prove it), or oversight without the authority to stop the system.
- Are logs, incidents and model changes recorded and reviewed? Evidence: the log settings and retention period, the incident register, the model change log with approval and retest, and records of periodic review. Red flag: logs that nobody reads, or a new supplier model version put into use without review.
- Is AI from suppliers under the same controls? Evidence: the procurement requirements, the contract terms on documentation, logging, change notifications and incidents, and the outcome of the supplier assessment. Red flag: AI features switched on through an update to existing software that appear in no register.
What Audirium offers for this
Audirium builds the software that audit and risk teams use to test these questions. The inventory is kept in Modelio, the EUC and model register: for each AI application the owner, the role of the organisation, the Annex III category and the assessment under Article 6(3). The AI Act is in the Audirium library with the 113 articles of the legal text and 27 controls at article level, alongside ISO/IEC 42001, the NIST AI Risk Management Framework and the IIA AI Auditing Framework. An engagement can then be scoped against the obligations that apply to the organisation. Evidence is linked to a control once and reused where the AI Act overlaps with NIS2 or privacy requirements.
Audirium supports the testing and the evidence trail. It does not certify AI systems and does not guarantee compliance.
Further reading
- Seven Digital Deadlines Between August 2 and January 1: the AI dates next to the other digital rules of 2026.
- Model Risk & EUC: From Blind Spot to Control: inventory and validation of models and self-built tools.
- From processing register to compliance meter: privacy management in one connected file.