How to prepare for the EU AI Act

The dates moved in 2026, so start by getting them right. Regulation (EU) 2026/1744, the Digital Omnibus on AI, amends the AI Act. It was adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. Under it, stand-alone Annex III high-risk systems must comply by 2 December 2027, and AI embedded in regulated products under Annex I by 2 August 2028.

What did not move matters just as much. Prohibited practices and AI literacy have applied since 2 February 2025. Obligations for general-purpose AI models have applied since 2 August 2025. Transparency duties in Article 50 apply from 2 August 2026, with a grace period to 2 December 2026 for Article 50(2) marking on systems already on the market. A delay of the high-risk date is not a pause on everything else.

Check these dates against the primary source before you publish a plan, because the text was amended recently and commentary varies. Our own EU AI Act timeline tracks them and says when it was last verified.

You cannot classify what you have not found. Include systems you build, systems you buy, AI features inside tools you already use, and agents that act on your behalf. Include the foundation models underneath them. Shadow AI counts, so how to detect shadow AI is a useful companion step.

For each system record the name, the owner, the vendor, the purpose in one sentence, who is affected by its output (customers, employees, applicants), what personal data it uses, and where it runs. Add the model or models behind it and whether you trained or fine-tuned them. Keep the purpose precise. The same model can be minimal risk in one use and high-risk in another, and the Act classifies by intended purpose and use.

Treat the register as a living document with an owner and a review date. A register that is accurate on the day you write it and stale in a quarter will not help when someone asks.

Article 5 lists AI practices that are banned outright, and it has applied since 2 February 2025, so this is the most urgent check. Read the current text of Article 5 with your counsel and compare each purpose in your register against it. The Digital Omnibus adds two more prohibitions, on AI that generates non-consensual intimate imagery and child sexual abuse material, applying from 2 December 2026.

If a system might fall under a prohibition, stop and escalate to legal before doing anything else. Do not wait for the high-risk date. Penalties for prohibited practices are the highest tier in the Act, and the penalties page explains the structure.

Record the result for every system, including a line on why you concluded it was clear. That reasoning is evidence.

The Act puts different duties on different roles. A provider develops an AI system or has one developed and places it on the market or puts it into service under its own name. A deployer uses an AI system under its own authority in a professional setting. Most organisations are deployers of some systems and providers of others.

The role can change. Under Article 25, a deployer, importer or distributor becomes the provider of a high-risk system by putting its own name on it, making a substantial modification, or changing the intended purpose so that the system becomes high-risk. If you fine-tune a model and sell the result inside a product, check whether you have become a provider. If you only run a vendor tool within the vendor's instructions, you are usually a deployer.

General-purpose AI models have their own provider obligations in Chapter V, which sit separately from the high-risk rules. If you build on someone else's model you are usually not that model's provider, but your product can still be high-risk. See GPAI obligations.

Work through the classes in order. Prohibited: done in the previous step. High-risk: a system is high-risk if it falls under one of the use cases in Annex III, or it is a safety component of, or itself, a product covered by Annex I. Annex III has eight areas: biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and the administration of justice. Within each area it lists specific use cases, so being in an area is not enough on its own.

Article 6 allows a provider to conclude that an Annex III system is not high-risk in specific circumstances, for example where it does not pose a significant risk of harm. If you rely on that, write the reasoning down and expect to be asked for it. Commentary on the final Omnibus deal says the registration step for such systems was kept, so confirm this with counsel against the final text.

Transparency: Article 50 sets disclosure duties, for example telling people they are talking to an AI system, and marking AI-generated content. These apply from 2 August 2026. General-purpose models carry their own rules. Everything else is minimal risk and carries no specific obligation under the Act, though other laws such as the GDPR still apply.

Use the high-risk checklist to walk each borderline system. Where two readings are possible, pick the cautious one and get legal review.

Article 4 on AI literacy has applied since 2 February 2025. The Digital Omnibus softened it: commentary on the amendment says providers and deployers are required to take measures to support AI literacy among their staff, rather than to guarantee a specific level. That is still an obligation, and a regulator will ask what you did.

Match the training to the role. People who approve AI outputs need to know the system's limits. People who buy tools need to ask the right questions. Engineers need to know the rules for the classes they build. Keep a record: who attended, the date and the content. The Article 4 guide covers what to include.

For high-risk systems, providers face the requirements in Articles 9 to 15 (risk management, data governance, technical documentation, record-keeping, transparency to deployers, human oversight, accuracy, resilience to errors and attacks, and cybersecurity), a quality management system under Article 17, conformity assessment, registration and post-market monitoring. The Omnibus simplified some documentation and quality-management requirements for smaller companies; check whether you qualify.

Deployers of high-risk systems have duties under Article 26: use the system according to its instructions, assign competent human oversight, monitor operation, keep the logs the system generates for an appropriate period (our EU checklist summarises the minimum as six months; confirm in the text), and inform workers and affected people. Some deployers must also carry out a fundamental rights impact assessment under Article 27. Public bodies, private entities providing public services and deployers of certain Annex III credit and insurance use cases are named in the checklist.

Where you rely on a vendor, ask for what you need: the provider's instructions for use, its documentation, and a way to get logs. Put those requests into contracts now, because you will need them well before December 2027.

Work backwards from 2 December 2027. Conformity assessment and documentation take longer than the calendar suggests, which is why starting early matters even after the delay.

Evidence is what turns a plan into something you can show. For each system keep the register entry, the classification reasoning, the role decision, the risk assessment, the data sources and their governance, test and evaluation results, the human oversight design, the logs and their retention, and the training record. How to validate your AI explains how to produce test results that stand up, and how to audit AI systems explains how an auditor would check them.

Where personal data is involved, pair this with a data protection impact assessment. How to do a DPIA for AI covers it, and the two documents can share evidence.

Review on a schedule. Set a quarterly check that the register matches reality, the dates still match the Official Journal, and no system has changed purpose.