EU AI Act Enforcement Begins: Penalties and Powers

EU AI Act enforcement formally began on August 2, 2026 — but not the enforcement most headlines describe. What activated is the European Commission’s power, exercised through the EU AI Office, to investigate and fine general-purpose AI model providers, plus exercisable penalties for prohibited AI practices and new Article 50 transparency duties for chatbots, synthetic media, and deepfakes.

What did not activate is the tier that dominated two years of compliance planning: the high-risk system obligations covering hiring tools, credit scoring, and education. The Digital Omnibus — Regulation (EU) 2026/1744, in force since July 27, 2026, roughly a week before enforcement day — deferred those duties to December 2, 2027. Misreading which track your systems sit on leads to over-reacting in one direction or sleepwalking in the other.

This guide maps the enforcement machinery itself: the penalty tiers and the two distinct fining articles, who actually enforces what, the complaint paths, and what day-one enforcement realistically looks like given that only a fraction of member states have staffed their side. It is deliberately not another compliance checklist — for that, see the risk-tier compliance checklist and the broader EU compliance guide.

Strip away the “AI Act enforcement day” framing and three specific things went live. First, per Wilson Sonsini’s client alert, the Commission — acting through the EU AI Office — became formally entitled to exercise investigative and enforcement powers over general-purpose AI model providers and the prohibited-practices rules. The GPAI obligations themselves have applied since August 2, 2025; what is new is the teeth. Second, penalties for Article 5 prohibited practices — social scoring, manipulative systems — became exercisable. Third, Article 50 transparency duties took effect for chatbots, synthetic content, deepfakes, and emotion-recognition systems.

The named-company framing matters here. CNBC’s enforcement-day coverage put Anthropic, OpenAI, and Google at the front of the new scrutiny — frontier labs whose models fall squarely into the GPAI category the AI Office now polices. The backdrop is already tense: EU regulators fined Google $1 billion in July 2026 in a separate Digital Markets Act case — not an AI Act action — which drew a tariff threat from President Trump. AI Act enforcement now unfolds inside that wider EU–US friction.

The Digital Omnibus on AI is now law: Regulation (EU) 2026/1744. Council and Parliament reached political agreement on May 7, 2026; the text was published in the Official Journal on July 24 and entered into force on July 27 — three days after publication, and roughly a week before AI Act enforcement day arrived. Its timing is the whole point: the postponement took legal effect before the deadline it postponed.

The substance, per Gibson Dunn’s analysis: stand-alone high-risk AI systems under Annex III — employment and HR screening, education, credit scoring, law-enforcement-adjacent uses — moved from August 2, 2026 to December 2, 2027. High-risk systems embedded in regulated products under Annex I, such as medical devices and machinery, moved from August 2, 2027 to August 2, 2028. Wilson Sonsini’s alert makes the complementary point explicit: the Omnibus postponed the Act’s principal high-risk requirements but does not affect any of the August 2, 2026 developments covered in this guide.

The Act runs two distinct penalty tracks, and press coverage routinely merges them. Article 99 sets the general framework in three tiers: up to €35 million or 7% of global annual turnover — whichever is higher — for prohibited AI practices; up to €15 million or 3% for breaches of most other operator obligations, including Article 50 transparency; and up to €7.5 million or 1% for supplying incorrect, incomplete, or misleading information to authorities or notified bodies.

The track that matters for frontier labs is Article 101: a GPAI-provider-specific cap of 3% of global annual turnover or €15 million, whichever is higher. Triggering conduct includes infringing the GPAI obligations, failing to respond accurately to an Article 91 information request, ignoring Article 93 risk-mitigation measures, or blocking Commission access for model evaluations. Beam.ai’s enforcement analysis explicitly flags the €35M/7% figure as commonly mis-attributed to GPAI providers — the higher tier belongs to prohibited practices, not to model-provider obligations.

Two structural softeners are worth knowing. For SMEs and startups, each Article 99 fine is capped at the lower of the percentage or the fixed euro amount — the inverse of the rule for larger undertakings. And Article 101 fines come with procedural safeguards: the Commission must communicate preliminary findings and give the provider a chance to respond before finalizing a decision, the EU Court of Justice holds unlimited jurisdiction to cancel, reduce, or increase a fine, and imposed fines are reported to the European Artificial Intelligence Board. National fining powers carry their own gate — each member state had to lay down proportionate, dissuasive penalty rules by August 2, 2025 and must report annually to the Commission on fines imposed.

Enforcement authority splits three ways by system and actor type, per the Commission’s enforcement-framework page. The EU AI Office, inside the Commission, enforces GPAI model rules, AI systems built by GPAI providers on their own models, and AI integrated into very large online platforms and search engines. National market surveillance authorities enforce everything else — including most high-risk systems once their compliance dates arrive, and Article 50 transparency for non-GPAI deployers. The European Data Protection Supervisor covers AI systems used by EU institutions themselves. The table below maps obligation to enforcer, penalty basis, cap, and — critically — whether it is actually live.

One actor is missing from the table because it polices a different population: the European Data Protection Supervisor, which supervises AI systems used by the EU’s own institutions. For businesses, the operative split is AI Office versus national authority — and which one knocks depends on whether you are a GPAI provider, a platform at very-large scale, or an ordinary deployer.

For general-purpose model providers, the AI Office’s toolkit is now fully exercisable, per Wilson Sonsini and the Commission’s framework page:

• Information requests (Article 91) — demand documentation and answers from providers.
• Model evaluations (Article 92) — obtain access to the model itself to conduct evaluations.
• Corrective measures (Article 93) — require risk-mitigation steps, and in serious cases request that a provider restrict, withdraw, or recall a model from the EU market.
• Fines — up to the Article 101 cap of €15 million or 3% of global turnover.

The subtlety practitioners keep underlining is that the process itself is the exposure. Elisabetta Righini, a partner at Sidley Austin, told CNBC that “A U.S. address does not put a lab outside the EU regulator’s reach” — any company offering a GPAI model in the EU is in scope regardless of where it is based, and non-EU providers must appoint an EU-based authorised representative as the regulator’s point of contact.

In practice, the AI Office says its preferred opening move is not a fine at all. Its stated initial tool is the technical compliance dialogue — informal engagement to assess compliance and resolve questions — which the Office says will continue and may intensify after August 2, with formal powers reserved for cases dialogue does not resolve. Two more scoping details matter. Models placed on the market before August 2, 2025 have a longer runway — until August 2, 2027 — to reach full compliance, which shapes which models are genuinely exposed today. And the named labs have stayed conciliatory in public: OpenAI’s VP of EMEA policy, Tom Duff Gordon, told CNBC the company has collaborated closely with the Commission on implementing the Act and its Codes of Practice, while a Google spokesperson said the company remains dedicated to meeting all applicable rules as the Act and its codes take effect.

The transparency tier deserves its own playbook — and has one, in the Article 50 checklist for agencies — so here is only the enforcement-relevant shape. Four duty categories applied from August 2, 2026: AI systems that interact directly with people (chatbots, voice assistants, AI agents) must disclose the interaction is with AI unless obvious from context; systems generating synthetic audio, image, video, or text must mark outputs in a machine-readable, detectable way; deployers of deepfakes and AI-generated text on public-interest matters must disclose the AI origin; and emotion-recognition or biometric-categorization deployers must inform exposed individuals.

The duties apply to in-scope systems regardless of when they were placed on the market — content published before August 2 does not need retroactive labeling, and the only transition is for the machine-readable-marking duty on systems already on the market, which have until December 2, 2026. Breaches sit in Article 99’s €15M/3% tier. On the compliance-support side, the Commission’s enforcement announcement notes that over 180 organizations have signed the voluntary Code of Practice on transparency of AI-generated content — a separate instrument from the GPAI Code above — which the Commission endorses as an adequate way to demonstrate marking and labeling compliance. For the implementation mechanics in an ad pipeline, see our provenance-marking guide for ad creative.

Member states were legally required to designate their national competent authorities and single points of contact by August 2, 2025 — a full year before enforcement day. They did not all make it, and the two independent trackers that count disagree on how short the field fell, because they use different completeness criteria. The honest summary: somewhere between a third and half of member states had their enforcement side in place as of mid-2026.

Two things stop this from being a free pass. First, the Act applies to providers and deployers regardless of national enforcement readiness — the gap creates no exemption, only uneven near-term enforcement intensity. A company operating from Spain or Italy faces a materially different practical environment than one in Austria or Greece, for now. Second, the Commission’s own channels are already open: an AI Act Complaint Tool for alleged infringements by supervised providers, an AI Act Whistleblower Tool for professionals connected to providers, and a Downstream Provider Complaints Channel for issues with integrated GPAI models. Under Article 85, any natural or legal person — not only those directly harmed — may lodge a complaint with a market surveillance authority. And the same August 2 deadline required every member state to have at least one operational AI regulatory sandbox — a compliance-support mechanism running in parallel with enforcement, not an enforcement power.

The right response depends entirely on which enforcement track you are on. Four postures cover most businesses:

Reading the machinery as a whole, the near-term pattern is dialogue-first, process-offence-second: the AI Office says it prefers technical compliance dialogues, which means the earliest real tests are likely to be about how completely and honestly providers answer Article 91 requests — not about headline substantive breaches. Layer on the staffing gap and the picture is an enforcement regime that is legally live everywhere but practically concentrated: strongest at the Commission level over a small set of frontier labs, patchy at national level, and varying sharply by member state.

Looking forward from here, three dates structure the next eighteen months: December 2, 2026, when the marking transition for pre-existing systems closes; August 2, 2027, when pre-Aug-2025 GPAI models must reach full compliance; and December 2, 2027, when the Annex III high-risk tier finally lands. Enforcement day also reframes procurement — “where does my AI run and who is liable” is now a purchasing question, which is exactly the calculus behind the sovereign-AI decision framework. If your systems touch scoring or screening, start with the lead-scoring risk-classification walkthrough; US-based teams should read what US businesses need to do. And if you want the classification and compliance work folded into how your AI systems are actually built, that is what our AI transformation engagements are for.