EU AI Act Compliance Checklist

Applies to every organisation. Nothing else in this list can be answered until this is done.

• Build a single inventory of every AI system your organisation builds, buys, embeds or uses.Include systems bundled inside vendor SaaS, which is where most shadow AI is found.
• Record your role for each system: provider, deployer, importer or distributor.Article 3. The role decides which obligations attach, and providers carry almost all of them.
• Check whether you have silently become a provider under Article 25.Putting your name or trademark on a high-risk system, substantially modifying it, or repurposing it so it becomes high-risk transfers the full provider burden to you, unless a contract allocates those obligations otherwise.
• If you supply systems others build on, meet the cooperation duties that now carry a fine.Where a downstream party becomes the provider, the original provider must cooperate and hand over the information and technical access they reasonably need, in writing. Those duties were added to the 15 million euro or 3 percent tier in 2026, so the new exposure sits with the upstream supplier rather than the party that rebranded. It falls away if you clearly specified the system is not to be turned into a high-risk system.
• Confirm territorial scope if you are established outside the EU.The Act reaches non-EU providers and deployers where the output produced by the system is used in the Union.

Binding since 2 February 2025. Two new prohibitions apply from 2 December 2026. Breach carries the highest penalty tier, up to 35 million euro or 7 percent of worldwide annual turnover.

• Screen every system in your inventory against the Article 5 prohibitions.Manipulative or subliminal techniques, exploitation of vulnerability, social scoring, predicting criminal offending from profiling alone, untargeted facial-image scraping, emotion inference, biometric categorisation of protected characteristics, and real-time remote biometric identification for law enforcement.
• Confirm you infer emotions in neither the workplace nor education settings.Prohibited outside genuine medical or safety purposes. This catches interview-analysis, productivity-monitoring and classroom-attention tools.
• Confirm no system categorises people biometrically to deduce protected characteristics.Race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation.
• Prepare for the two new prohibitions applying 2 December 2026.Article 5(1)(ba) and (bb) ban AI that generates or manipulates non-consensual intimate imagery of an identifiable person, and child sexual abuse material.
• If you provide a general-purpose image, video or audio generator, document your safeguards.Article 5(1a) catches placing on the market where such generation is the intended purpose, or where it is a reasonably foreseeable and reproducible outcome without significant technical modification and the system lacks adequate safeguards. Demonstrable technical safety measures plus a misuse-correction process are the route out of the second limb. Deployers are caught only when they use the system for that purpose.
• If you build editing tools, note the manipulation safe harbour.Manipulating material in a way that does not increase the exposure of any depicted intimate parts, or alter the nature of any depicted sexually explicit activities, does not count as manipulation for the new prohibition.
• Check the grandfathering position for systems already on the market, which the prohibitions cut across.High-risk systems placed on the market before the Chapter III dates are largely grandfathered unless their design changes significantly, and the 2026 amendments extended that window by tying it to the deferred dates. The prohibitions are the exception: the legacy provision applies without prejudice to Article 5, so an existing system must still meet the new prohibitions.

Article 4. Binding on every provider and deployer at every risk tier, including minimal risk. In force since 2 February 2025 and softened, not removed, in July 2026.

• Provide AI literacy measures for staff and anyone operating AI on your behalf.Proportionate to their technical knowledge, experience, education and the context of use.
• Document those measures.The current wording requires you to support the development of AI literacy. It expressly does not require you to guarantee any individual level of competence, so your evidence is the programme, not test scores.

Article 50. Live since 2 August 2026 with no grace period, other than the one marking deadline noted below. This is the tranche most organisations wrongly believe was delayed.

• Tell people when they are interacting with an AI system.Required unless it is obvious to a reasonably well-informed, observant and circumspect person. Chatbots and voice agents are the common case.
• Mark synthetic audio, image, video and text in a machine-readable, detectable format.The marking must be effective, interoperable, robust and reliable as far as technically feasible.
• Notify people exposed to emotion recognition or biometric categorisation where its use is permitted.
• Disclose deepfake content, and AI-generated text published to inform the public on matters of public interest.The text duty falls away where the content underwent human review and a person holds editorial responsibility.
• If a synthetic-content system was on the market before 2 August 2026, meet the machine-readable marking duty by 2 December 2026.This transitional deadline sits in Article 111(4) rather than Article 50 itself. It is the only grace period touching the transparency duties, it covers the marking duty alone, and it applies to providers only. Content created before 2 August 2026 needs no retroactive labelling.

Chapter V. Binding since 2 August 2025 on providers of GPAI models. Skip this section if you only use models built by others.

• Maintain technical documentation for the AI Office and provide integration documentation to downstream providers.
• Operate a copyright policy that identifies and respects text-and-data-mining rights reservations.Rights reserved under Article 4(3) of the Copyright in the Digital Single Market Directive, using state-of-the-art technology.
• Publish a sufficiently detailed summary of training content using the AI Office template.
• If training compute exceeds 10^25 FLOP, notify the Commission within two weeks and meet the systemic-risk duties.Model evaluation including documented adversarial testing, Union-level systemic-risk mitigation, serious-incident reporting to the AI Office, and adequate cybersecurity. The open-source exemption does not apply to systemic-risk models.

The obligations are deferred to 2 December 2027 (Annex III) and 2 August 2028 (Annex I), but the Commission's classification guidance was carved out of that deferral and applies from 2 August 2026. Classify now.

• Classify each system against Annex I and Annex III.Annex I covers AI as a safety component of regulated products. Annex III covers eight stand-alone areas including biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and administration of justice.
• If you rely on the narrow-task derogation, document the assessment before the system reaches the market.The derogation does not apply where the system profiles natural persons, it does not remove the registration duty, and national authorities can demand the assessment.
• Plan on the deferred dates being final.The deferral carries no standards-readiness trigger and no automatic extension, so the dates bind whether or not harmonised standards exist.

Chapter III. The build programme behind the 2027 and 2028 dates. Conformity work realistically takes longer than the time remaining.

• Providers: stand up the seven build-time requirements.Risk management across the lifecycle, data governance, technical documentation per Annex IV, automatic logging, instructions for use, human oversight by design, and accuracy, robustness and cybersecurity.
• Providers: operate a quality management system and retain documentation for ten years.
• Providers: choose your conformity route, then declare conformity, affix CE marking and register.Most Annex III systems use internal control. Biometrics generally requires a notified body where harmonised standards do not exist or were not fully applied. A substantial modification triggers a fresh assessment.
• Deployers: assign human oversight to competent, trained and authorised people, and keep logs for at least six months.Also use the system per its instructions, monitor it, and suspend use and notify the provider or authority when risk emerges.
• Deployers: inform workers' representatives and affected workers before putting a high-risk system into service in the workplace.
• Deployers: run a fundamental rights impact assessment before first use where it applies to you.Required of public bodies and private providers of public services, and of deployers of creditworthiness and of life and health insurance pricing systems. Critical infrastructure systems are expressly excepted. The results are notified to the market surveillance authority.

The new enforcement architecture went live on 2 August 2026, not with the deferred high-risk tranche.

• Plan for EU database registration of Annex III high-risk systems before they are placed on the market.Critical infrastructure systems are excepted and register nationally instead, and registrations for biometrics, law enforcement and migration go to a non-public section. Note a genuine open question: the registration article was not deferred, but the classification article it depends on was, so how registration operates before December 2027 is unresolved. The 2026 amendments trimmed the submission required on the narrow-task derogation track, without abolishing it.
• Identify your national competent authority and its single point of contact.
• Work out whether the AI Office, rather than your national authority, supervises you.It has exclusive competence over AI systems built on GPAI models where model and system come from the same provider or undertaking, subject to carve-outs, and over systems that constitute or are integrated into a very large online platform or search engine. Serious incidents for those systems are reported to the AI Office.
• Brief your board on the real exposure, including the 2026 additions.Up to 35 million euro or 7 percent of worldwide annual turnover for prohibited practices, 15 million or 3 percent for operator duties, and 7.5 million or 1 percent for supplying misleading information. Within the AI Office's perimeter there are also periodic penalty payments of up to 5 percent of average daily worldwide turnover per day.