31 steps · 7 phases
Every step below names the provision it comes from and the date it has to be true by. The phases run in the order the work actually sequences, you cannot classify a system before you know what you hold, or what role you hold it in.
With phase 1, always. Every obligation in the Regulation depends on facts about a specific system (its intended purpose, the role you hold for it, and whether it falls in an Annex III area) so nothing downstream can be assessed until the inventory exists.
Organisations that begin with the legal text rather than the inventory generally have to begin again. The inventory has no deadline of its own and blocks every phase after it, which makes it the only genuinely urgent item on this list.
Three obligations on this checklist are in force now, not pending. AI literacy has applied since 2 February 2025, the first eight prohibitions since the same date, and the Article 50 transparency framework since 2 August 2026.
Every obligation in the Regulation depends on facts about a specific system, so nothing can be assessed until the inventory exists. This phase has no deadline of its own and blocks every phase after it.
Include the intended purpose, the business process it sits in, whether output reaches people outside the organisation, and the vendor. Features added to tools already in use are the most commonly missed category, because no procurement decision marked their arrival.
Duties attach to roles, not to company types, and one organisation is routinely several at once. A deployer becomes a provider if it puts its own name on a system, substantially modifies it, or changes its intended purpose, which is the single most common scoping mistake.
You are in scope if you place a system on the EU market wherever you are established, or if the output is used in the Union. Being outside the EU is not itself an exclusion. Military, defence, national security and purely personal non-professional use are excluded.
Article 5 is absolute: a prohibited practice cannot be made lawful by documentation, consent or safeguards. Screen for it before spending effort anywhere else, because a hit means the system cannot be placed on the market or used at all.
Eight have been unlawful since 2 February 2025. Two more, non-consensual intimate imagery and child sexual abuse material, apply from 2 December 2026. Any material listing eight prohibitions describes the superseded text.
Several prohibitions have narrow statutory qualifications: notably real-time remote biometric identification for law enforcement. Relying on a carve-out is a documented legal position, not an assumption.
Classification decides which regime applies and therefore how much work follows. The two high-risk routes carry different deadlines, so the route matters as much as the conclusion.
A system is high-risk if it is a safety component of a product covered by the Union harmonisation legislation in Annex I, or is itself such a product, and requires third-party conformity assessment under that legislation.
Eight areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and border control, and the administration of justice. This route carries the earlier deadline.
A system in an Annex III area is not high-risk if it does not pose a significant risk of harm, on the grounds listed in Article 6(3). Relying on this requires a documented assessment before placing on the market, and registration in the EU database.
Transparency duties attach regardless of risk tier wherever a system interacts with people, generates synthetic audio, image, video or text, produces deep fakes, or infers emotion or biometric categories. A minimal-risk system can still owe them.
General-purpose AI models are regulated in their own right under Chapter V, separately from systems built on them. Assess whether the model meets the systemic-risk threshold, and notify the Commission if it does.
This is the heavy phase and the reason the deferral to December 2027 is a schedule change rather than a reprieve. Each requirement produces evidence that a conformity assessment will look for.
Not a one-off assessment: a process that runs across the lifecycle, identifying and evaluating known and reasonably foreseeable risks, and adopting mitigation measures. It has to be documented and maintained.
Covers design choices, collection and origin, preparation, assumptions, availability and suitability, and examination for bias. Article 10(5) permits processing special categories strictly for bias detection and correction, subject to safeguards, and Article 4a broadened that basis.
It must exist before the system is placed on the market and be kept current. Annex IV sets the minimum contents; this is the document a market surveillance authority asks for first.
Logs must allow traceability appropriate to the intended purpose. Retention is at least six months unless other Union or national law requires longer.
Including the intended purpose, accuracy and robustness levels, known limitations, foreseeable misuse, the human oversight measures required, and expected lifetime and maintenance.
The person overseeing must be able to understand capacities and limits, monitor for anomalies, resist automation bias, correctly interpret output, decide not to use it, and intervene or stop. A review step that cannot actually override the system does not satisfy this.
Declared metrics go in the instructions for use. Robustness covers resilience to errors and inconsistencies; cybersecurity covers attempts to alter use, behaviour or performance, including data poisoning and adversarial examples.
A documented system covering regulatory compliance strategy, design and verification procedures, data management, the risk management system, post-market monitoring, incident reporting and accountability.
Most organisations deploy far more systems than they provide, and deployer duties are frequently missed because the provider's documentation is mistaken for the whole obligation.
Oversight must be assigned to named people with the competence, training and authority to exercise it, and the authority matters as much as the training.
Suspend use and inform the provider where you have reason to consider that use in accordance with the instructions may present a risk. Keep logs for at least six months.
Workers and their representatives must be informed before a high-risk system is put into use in the workplace. People subject to a decision informed by an Annex III system must be told the system was used.
Required for deployers that are public bodies, or private entities providing public services, and for certain creditworthiness and life or health insurance pricing uses.
The gate between a compliant build and lawful placing on the market. Which route applies depends on the classification route and on whether harmonised standards were applied.
Annex VI internal control, or Annex VII involving a notified body. For most Annex III systems internal control is available where harmonised standards were applied in full; biometrics is the notable exception.
The declaration is kept for ten years after placing on the market and made available to authorities on request.
Applies to providers of Annex III high-risk systems, to those relying on the Article 6(3) derogation, and to certain public-authority deployers.
Required for providers established outside the Union, by written mandate, before placing a high-risk system on the market.
Compliance is a standing condition, not a launch milestone. These obligations run for as long as the system is on the market.
Proportionate to the nature and risks of the system: collect, document and analyse performance data across the lifetime, and feed it back into the risk management system.
Report to the market surveillance authority of the Member State where the incident occurred. The deadlines are short and vary with the nature of the incident, so the path has to exist before it is needed.
Providers of generative systems must mark output as artificially generated or manipulated. Systems already on the market when the transparency rules began have until 2 December 2026 to comply.
Applies to every provider and deployer regardless of risk tier, and has since 2 February 2025. The Digital Omnibus softened this from ensuring a level of literacy to taking measures to support it, which changes what you should be recording, not whether you owe it.
Regulation (EU) 2026/1744 changed deadlines, added two prohibitions, softened Article 4 and created new AI Office enforcement powers. Deadlines and prohibition counts in older material are likely to be wrong.
A checklist is only useful once each line has an owner and a date against it. We will send this as a working document you can assign and track, with the provision links intact.
Seven phases in the order the work sequences: build an inventory of your AI systems and fix your role for each, screen against the ten Article 5 prohibitions, classify each system against both high-risk routes and the Article 50 transparency duties, build the Chapter III requirements for anything high-risk, meet the separate deployer duties, clear conformity assessment and registration, then operate post-market monitoring and incident reporting. AI literacy under Article 4 applies throughout, regardless of tier.
With an inventory. Every obligation depends on facts about a specific system (its intended purpose, the role you hold for it, and whether it falls in an Annex III area) so scope cannot be assessed until you know what you hold. Organisations that start with the legal text rather than the inventory generally have to start again.
Yes, for three reasons. The Article 5 prohibitions apply in every tier. The Article 50 transparency duties apply in every tier, and catch any system that interacts with people or generates synthetic content. And the Article 4 AI literacy duty applies to every provider and deployer regardless of risk. You also need documented reasoning for the conclusion that a system is not high-risk.
It depends which obligation. Prohibited practices and AI literacy have applied since 2 February 2025, general-purpose AI model obligations since 2 August 2025, and the transparency framework since 2 August 2026. Two further prohibitions and the legacy synthetic-content marking deadline fall on 2 December 2026. High-risk obligations apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products.
The Regulation places obligations on the organisation as an operator rather than on a named officer, so accountability follows your existing governance. In practice the inventory sits with technology risk, classification with legal, the Chapter III requirements with engineering and product, and the deployer duties with the business owner using the system.
No. A checklist organises the work and shows what has been considered; it is not itself evidence. What a market surveillance authority asks for is the artefacts: the technical documentation to Annex IV, the risk management records, the logs, the declaration of conformity and the post-market monitoring data.