EU AI Act Enforcement Has Started: Where You Stand Now
This page answers one question: now that enforcement powers are live, what can a national authority or the AI Office actually do to your company, and what should you do first if you have not started. It covers the practical effect of 2 August 2026, the split between provider and deployer duties, and the remaining dates that still bind.
What changed on 2 August 2026
2 August 2026 was the general date of application set out in Article 113. Before it, most of the Regulation existed on paper but without a supervisory apparatus pointed at you. From that date, the enforcement chapters apply, national market surveillance authorities exercise their powers, and the penalty regime in Article 99 is operative in the Member States that have transposed it.
Three bodies of obligation are now live and being supervised:
- Article 5 prohibited practices. These have applied since 2 February 2025, but until August 2026 there was, in most Member States, no designated authority empowered to fine you for breaching them. There is now.
- Article 50 transparency obligations. Disclosure that a user is interacting with an AI system, machine readable marking of synthetic audio, image, video and text, disclosure of emotion recognition and biometric categorisation to exposed persons, and labelling of deep fakes.
- Article 53 obligations for providers of general purpose AI models. Technical documentation, information for downstream providers, a copyright policy, and a sufficiently detailed public summary of training content. Article 55 adds model evaluation, systemic risk assessment and mitigation, incident tracking and cybersecurity for models classified under Article 51.
Alongside these, Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among staff operating their systems. It has applied since 2 February 2025 and is cheap to satisfy, which makes it an awkward one to be caught short on.
What supervisory authorities can do from today
Member States were required under Article 70 to designate notifying and market surveillance authorities, and under Article 99 to lay down and notify their penalty rules, by 2 August 2025. Where that has happened, the authority operates with the market surveillance powers of Regulation (EU) 2019/1020 as applied by Article 74. In practice that means it can demand your technical documentation, require information about how a system was built and tested, and where necessary to assess conformity of a high-risk system, obtain access to training data and in defined circumstances source code.
Article 79 governs what happens if an authority finds an AI system presents a risk. It can require the operator to take corrective action, bring the system into compliance, withdraw it from the market or recall it within a prescribed period. Non-compliance with that instruction escalates to restriction or prohibition on the national market.
For general purpose AI models, enforcement sits with the Commission through the AI Office. Article 91 lets it request documentation and information from providers, Article 92 lets it conduct evaluations of a model, and Article 93 lets it require providers to take measures including restricting the model on the market, withdrawing it or recalling it.
On money, Article 99(3) sets the ceiling for breach of Article 5 at 35 million euro or 7 per cent of total worldwide annual turnover, whichever is higher. Article 99(4) sets 15 million euro or 3 per cent for breach of most other operator obligations, including the Article 50 transparency duties and deployer duties under Article 26. Supplying incorrect, incomplete or misleading information to authorities carries up to 7.5 million euro or 1 per cent under Article 99(5). For SMEs and start-ups, Article 99(6) applies the lower of the percentage and the fixed sum rather than the higher. Fines on GPAI model providers are set separately in Article 101 at up to 15 million euro or 3 per cent of worldwide annual turnover.
Provider or deployer, and why the answer changes your exposure
Article 3 defines a provider as the entity that develops an AI system or model, or has one developed, and places it on the market or puts it into service under its own name or trade mark. A deployer is the entity using an AI system under its own authority, other than in a purely personal non-professional activity.
If you build and ship an AI feature in your own product, you are a provider of that system. If you also train or substantially fine-tune a general purpose model you then place on the market, Article 53 applies to you as a GPAI provider in addition to anything that applies to the system. If you buy a vendor tool and run it on your staff or customers, you are a deployer, and your live duties are narrower but real: Article 50 transparency where your use is customer facing, Article 4 literacy, and from the high-risk dates, Article 26.
Article 25 is the one that catches integrators. Put your own name or trade mark on a high-risk system, make a substantial modification to it, or change its intended purpose so that it becomes high-risk, and you become the provider with the full provider obligation set. Reselling a model under your own brand is not a neutral act.
What the live obligations concretely require
- Article 5: no deployment of the listed practices at all. The ones that reach ordinary software companies are emotion inference in workplace and education settings, untargeted scraping of facial images to build recognition databases, social scoring, and exploitation of vulnerabilities linked to age, disability or social and economic situation. There is no documentation route that makes these lawful.
- Article 50: a disclosure at first interaction for chat and voice interfaces, unless it is obvious to a reasonably well informed person. Machine readable marking of generated or manipulated output in a format that survives distribution, which in practice means C2PA style provenance metadata or watermarking, not a footer line. Deep fake content labelled as artificially generated.
- Article 53: technical documentation of the model kept current, a documentation package for downstream providers covering capabilities and limitations, a policy to comply with Union copyright law including respecting text and data mining reservations, and a public summary of training content following the AI Office template.
- Article 4: recorded training appropriate to role for the people who build, procure and operate your AI systems.
The dates that still bind
| Date | What applies |
|---|---|
| 2 February 2025 | Article 5 prohibitions and Article 4 AI literacy. In force. |
| 2 August 2026 | General application, enforcement powers, Article 50, Article 53. In force. |
| 2 December 2027 | Annex III standalone high-risk obligations, per the Digital Omnibus provisional agreement of May 2026, still pending formal adoption. |
| 2 August 2028 | Annex I embedded high-risk obligations, for AI that is a safety component of a product already covered by Union harmonisation law. |
The practical planning number is 415 days to 2 December 2027. A quality management system under Article 17, data governance under Article 10, logging under Article 12 and a conformity assessment are not twelve week projects.
If you have done nothing yet
- Inventory every AI system and model you ship or use, this month. One row per system: what it does, whether you built it or bought it, whether you are provider or deployer, whether any Annex III use case is in scope, and the named internal owner. Authorities asking questions under Article 74 start here, and so should you.
- Screen the inventory against Article 5 first. Anything that infers emotion from employees or students, scores people on unrelated behaviour, or scrapes faces at scale comes out of production now, not after a review cycle. This is the only category where the fine ceiling is 7 per cent.
- Fix Article 50 disclosure in the product, not the terms of service. Add the interaction notice to every chat and voice surface, and add provenance metadata to generated media output. Record the date each change shipped, because that date is your evidence of when you came into compliance.
- If you fine-tune or distribute a model, assemble the Article 53 pack. Model card, evaluation results, downstream documentation, copyright policy including your TDM reservation handling, and the public training content summary. Decide explicitly whether you are relying on the GPAI Code of Practice.
- Start the Article 11 technical documentation for anything heading towards Annex III. Write it as a living document from today rather than reconstructing design history in late 2027. Pair it with Article 12 logging: retain logs for the lifetime of the system and in any event at least six months, and make sure they capture inputs, outputs and human override events.
The honest position of a company that has done nothing is not catastrophic, but it is no longer a planning problem. Article 5 and Article 50 exposure is immediate and can be closed in weeks. High-risk readiness is a 2027 programme that needs to begin now. If you want a quick view of which of these buckets your systems fall into, the free screener at https://www.getactcomply.com/check walks through the classification questions and returns a per-system answer you can bring to your next engineering planning session.