Starting EU AI Act Compliance Late: A Catch-Up Plan
This page is for teams that have done little or nothing on the EU AI Act and are starting now, after enforcement powers went live. It sets out a sequence ordered by exposure, meaning what can be penalised today comes first and what is penalisable in 2027 and 2028 comes later. It is not a restatement of the Act in article order, because article order is a bad work plan when you are behind.
What the Act says, and which parts are already live
Three groups of obligations are in force and enforceable right now.
Article 5 bans a defined set of practices and has applied since 2 February 2025. These include manipulative or deceptive techniques that materially distort behaviour and cause significant harm, exploitation of vulnerabilities linked to age, disability or socio-economic situation, social scoring, certain predictive policing based solely on profiling, untargeted scraping of facial images to build recognition databases, emotion inference in workplaces and education outside medical and safety uses, and biometric categorisation inferring sensitive attributes. Article 99(3) sets the penalty ceiling for Article 5 breaches at 35 million euro or 7 per cent of total worldwide annual turnover, whichever is higher.
Article 50 covers transparency: telling people they are interacting with an AI system unless it is obvious, marking synthetic audio, image, video and text in a machine readable format, disclosing deepfakes, informing people subjected to emotion recognition or biometric categorisation, and labelling AI generated text published to inform the public on matters of public interest. Article 53 covers general purpose AI model providers: technical documentation for the model, information packages for downstream providers who integrate it, a copyright policy addressing the text and data mining reservation, and a sufficiently detailed public summary of training content. Both have been enforceable since 2 August 2026. Article 99(4) sets a ceiling of 15 million euro or 3 per cent of turnover for most non Article 5 breaches, and Article 101 sets a ceiling of 15 million euro or 3 per cent for GPAI model providers.
Article 4 requires providers and deployers to ensure a sufficient level of AI literacy among staff and others operating systems on their behalf. It has applied since 2 February 2025 and is the cheapest item on this list to close.
Provider or deployer, decided before anything else
Your obligations differ sharply by role, so fix the role first. You are a provider if you develop an AI system or model and place it on the market or put it into service under your own name or trade mark. You are a deployer if you use an AI system under your own authority in a professional context. Article 25 converts a deployer or distributor into a provider in three situations: putting your name or trade mark on an existing high-risk system, making a substantial modification to one, or modifying the intended purpose of a system so that it becomes high-risk. Teams that wrap a third party model, rename the output as their own product and change what it is used for are frequently providers and have not realised it.
Deployer duties under Article 26, including using systems according to the instructions for use, assigning competent human oversight, monitoring operation, and keeping automatically generated logs under your control for at least six months unless other law says otherwise, attach to high-risk systems and therefore follow the high-risk dates, not the 2026 dates. Transparency duties in Article 50 bind deployers now.
The catch-up sequence
Phase one, zero to 30 days: stop anything that is already unlawful
Inventory every AI feature shipped or used internally, including procured tools. Screen each one against Article 5. If a feature infers emotion from employees or students, scores people on unrelated behaviour, or categorises people biometrically into sensitive groups, it is off until legal sign off, not after the next sprint. Record the date of the decision and who made it. Article 99(7) requires authorities to weigh the intentional or negligent character of an infringement, mitigating action taken, and cooperation with the competent authority when setting a fine. A dated remediation log is the evidence that supports every one of those factors.
Phase two, 30 to 90 days: close the transparency and GPAI gaps
These are the cheapest currently enforceable obligations to satisfy. Add an interaction disclosure to every chat, voice and agent surface. Apply machine readable marking to generated media, for example C2PA content credentials or embedded provenance metadata, and visible labelling to deepfakes. If you train or substantially fine tune and distribute a general purpose model, publish the training content summary, write the copyright policy and assemble the Article 53 documentation set, including the downstream information package your API customers will ask for.
Phase three, 90 to 180 days: classify properly and decide product strategy
Run every system through Article 6 and Annex III. The Annex III areas are biometrics, critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services including credit scoring and life and health insurance pricing, law enforcement, migration and border control, and administration of justice. Article 6(3) allows a derogation where the system performs a narrow procedural task, improves a prior human activity, detects decision patterns without replacing human assessment, or performs preparatory work, but profiling of natural persons always counts as high-risk. If you claim the derogation, document the assessment and register it; do not leave it as an internal opinion.
Phase four, 180 days to 2 December 2027: build the high-risk file
For each confirmed Annex III high-risk system you need a risk management system (Article 9) running across the lifecycle, data governance covering training, validation and testing sets (Article 10), technical documentation to Annex IV (Article 11), automatic logging (Article 12), instructions for use for deployers (Article 13), human oversight measures (Article 14), accuracy, robustness and cybersecurity (Article 15), a quality management system (Article 17), conformity assessment (Article 43), the EU declaration of conformity (Article 47), CE marking (Article 48), registration in the EU database (Article 49), post-market monitoring (Article 72) and serious incident reporting to the market surveillance authority (Article 73). This is the work that takes four quarters, which is roughly what remains.
The dates that bind you
| Obligation | Status |
|---|---|
| Article 5 prohibited practices | Applies since 2 February 2025 |
| Article 4 AI literacy | Applies since 2 February 2025 |
| Article 50 transparency and Article 53 GPAI | Enforceable since 2 August 2026 |
| Annex III standalone high-risk systems | 2 December 2027, 414 days away, per the Digital Omnibus provisional agreement of May 2026 which is still pending formal adoption |
| Annex I embedded high-risk systems | 2 August 2028 |
Plan against 2 December 2027 as the binding date but keep the original schedule in your risk register, because the extension is provisional until the omnibus is formally adopted.
Five actions to take this quarter
- Produce a single AI inventory with, for each entry, the role you hold (provider, deployer, or both), the Article 6 and Annex III classification, the model or vendor behind it, and the named owner. One spreadsheet is acceptable. No inventory means no defensible position in an information request.
- Run the Article 5 screen and shut down or carve out failures within 30 days, with a written, dated record of the decision, the reasoning and the remediation.
- Ship transparency changes as a product task, not a legal task. Interaction disclosure, provenance marking on generated media, deepfake labelling, and notification for any emotion recognition or biometric categorisation still in use.
- Fix logging now, ahead of the high-risk deadline. Automatic, tamper evident logs of inputs, model version, outputs and human overrides, retained for the lifetime of the system and at minimum six months. Retrofitting logging into a live system takes longer than writing any policy document.
- Deliver AI literacy training with an attendance record to engineering, product, support and anyone operating a system on your behalf, and repeat it on a fixed cycle.
Where late starters go wrong
Two failure modes dominate. The first is starting with policy documents because they are easy to write, which produces a governance binder attached to systems that still have no logs, no evaluation records and no documented human oversight. The second is treating Annex III classification as a question to settle later, which leaves the longest piece of work, the technical file and conformity assessment, compressed into the final quarter before 2 December 2027. Classification is the gate on everything else, so resolve it by month six at the latest.
If you want a starting point for the inventory and classification step, the free screener at https://www.getactcomply.com/check walks through role, Annex III category and the Article 6(3) derogation and returns the list of articles that apply to each system you enter.