What already applies and what has been postponed
Regulation (EU) 2024/1689 entered into force on 1 August 2024 and its obligations phase in. From 2 February 2025 the bans on prohibited AI practices apply, together with Article 4, the obligation to ensure an adequate level of competence among people who operate AI systems. From 2 August 2025 the rules on general-purpose models, governance structure and penalties apply.
High-risk systems under Annex III, including safety components in the management of critical infrastructure, were to be subject to the rules from 2 August 2026. The political agreement on the Digital Omnibus on AI of 7 May 2026 moves this date to 2 December 2027, and for systems embedded in products under Annex I to 2 August 2028. Publication of the omnibus in the Official Journal of the EU has not been confirmed in the sources checked, so base your implementation plan on both dates and check the status of the text before an audit.
Does this apply to your plant
The AI Act distinguishes between the provider of a system and the deployer. A plant that buys and runs a ready-made tool is usually a deployer. A plant that builds its own model or substantially modifies someone else's may take on provider obligations. This distinction decides the scope of documentation.
The first step is straightforward: collect a list of all tools with an AI element, including those built into software that teams have used for a long time. In an industrial plant typical entries are a ticket assistant, instruction translation, camera image analysis, machine failure prediction and dispatcher support. For each entry record who uses it, for what, and which decisions rely on it.
Fine limits under Article 99 of the AI Act
Risk classification step by step
You classify each system separately. The order of the questions limits errors.
- Does the system carry out a prohibited practice under Article 5? If so, switch it off immediately.
- Is it a safety component in the management of critical infrastructure (energy, water, transport, utilities supply)? If so, it is probably high risk.
- Does it decide on employment, employee evaluation or access to training? That is also a high-risk area under Annex III.
- Does it only support a person in an auxiliary task, for example summarising a report? Record the reasoning for low risk and the scope of human verification.
- Does the system generate content visible to people? Check the transparency obligations under Article 50.
AI literacy: the cheapest obligation to prove
Article 4 does not require a certificate, but it does require people who operate AI systems to have adequate knowledge, taking into account their role and the context of use. The auditor will ask about three things: who uses AI, what they were taught and when.
The evidence is easy to collect: a list of users with roles, a short training programme (the model's limits, errors, a ban on pasting sensitive data, the duty to verify the output), a completion confirmation with a date and a refresher date. If you run the training in a system that records the test result and the expiry date, the evidence is created as a by-product of the work, not a week before an inspection.
Fines and what follows from them
Article 99 provides for up to EUR 35,000,000 or 7% of worldwide annual turnover for prohibited practices. Breach of the obligations of providers, importers, distributors and deployers, including for high-risk systems, carries up to EUR 15,000,000 or 3% of turnover. Supplying incorrect information to the authorities carries up to EUR 7,500,000 or 1%. For small and medium-sized enterprises, the lower of the amounts applies.
The fines relate to different infringements, so do not add them into a single risk figure. For the director the practical conclusion matters more: the largest fine concerns practices that a simple policy can rule out, and most of the work is documentation and human oversight.
Audit evidence: the minimum package
A package that supports meeting the requirement and provides audit evidence fits into five documents.
- A register of AI systems with risk classification and a review date.
- The role and name of the person responsible for oversight of each system (in an internal document, not a public one).
- Evidence of Article 4 training with expiry dates.
- Instructions for use and rules for human verification of the output.
- Usage logs where the system influences safety-related decisions.
Video
Frequently asked questions
Does the AI Act apply to a plant that only uses a ready-made AI tool?
Yes, as a deployer. Among other things, the bans on practices, Article 4 on competence and human oversight apply to it when the system is high risk.
Is an AI training certificate required?
Article 4 requires an adequate level of competence, not a specific certificate. The evidence is the programme, the list of participants and the completion date.
Which date should be adopted for high-risk systems?
Plan for 2 December 2027 (Annex III) and 2 August 2028 (Annex I), but check the publication status of the Digital Omnibus before budget decisions.
Sources
Put your AI evidence in order before the audit
Let us talk about the register of systems, training and the decision trail in your plant.
Let's talk