On 2 February 2025, the first legal teeth of the European Union's Artificial Intelligence Act became operational. That was the date the regulation's ban on unacceptable AI practices started to apply, putting social scoring, exploitative manipulation, untargeted facial-image scraping, and most real-time remote biometric identification in public spaces off-limits. On 2 August 2025, a second layer switched on: the governance machinery and transparency obligations for general-purpose AI models, along with the fine schedule. The next milestone is 2 August 2026, when the full regulation applies except for AI systems that are safety components in regulated products, which get until 2 August 2027.
The first compliance actions now being announced by European and national authorities are not the dramatic public fines that the regulation's penalty clauses make easy to imagine. They are formal requests for information, documentation reviews, public inspection priorities, and registration checks. A provider gets a dated letter, a list of technical documentation categories, and a reminder that supplying incorrect or incomplete information carries its own penalty. That sequence is deliberate: Article 99's largest fine, up to €35 million or 7 percent of total worldwide annual turnover, whichever is higher, sits at the top of a ladder whose first rungs are administrative.
A risk-based rulebook, not a blanket ban
The AI Act divides systems into four categories: unacceptable, high-risk, limited, and minimal risk. Most systems in use are expected to fall into the low and minimal risk bands. The Commission's guidance says systems in those bands face transparency duties rather than the full conformity machinery: a chatbot must disclose that a machine is answering, and deep fakes must be labelled, but the provider is not required to run a notified-body assessment.
What counts as high-risk is specific. Annex III lists sectors including biometrics, critical infrastructure, education, employment and worker management, access to essential private and public services, law enforcement, migration and border control, and administration of justice. Annex I captures AI systems used as safety components in products such as medical devices, machinery, toys, and lifts. Those are the two routes into the high-risk category, and each one changes the compliance burden dramatically.
The penalties that focus attention
EU law rarely separates fine tiers this cleanly. For placing a prohibited AI system on the market, the European Commission and national authorities can pursue penalties up to €35 million or 7 percent of worldwide annual turnover. For most other violations, including high-risk system breaches, the ceiling is €15 million or 3 percent. Supplying incorrect, incomplete, or misleading information to regulators sits in a third band at up to €7.5 million or 1.5 percent. For a large enterprise, the turnover-based figure quickly dwarfs the flat euro amount because the regulation specifies whichever figure is higher.
The European AI Office coordinates enforcement for general-purpose AI models. National market surveillance authorities enforce product rules inside member states. Spain's supervisory agency, AESIA, was set up in A Coruña in 2023, before the final text was adopted, and is one of the national bodies now receiving inspection and enforcement powers. The institutional wiring matters because a company can face an information request from the AI Office, a separate inquiry from a national authority, and a product recall in an individual member state from the same underlying model.
What the first enforcement wave targets
Providers of general-purpose AI models must document how their models were trained, including a summary of content used for training and a policy to comply with EU copyright law. Under the Act, those providers may have to supply that documentation to the European AI Office on request. The first wave of action is concentrated there: requests for technical documentation, clarifications on data governance, and checks on whether providers have registered in the relevant EU database.
For high-risk systems, enforcement will look more like traditional product compliance. A high-risk model used in recruitment sorting, exam scoring, or medical triage has to carry risk management documentation, technical logs, human oversight measures, and a conformity assessment before it can carry a CE mark. National market surveillance authorities can suspend a system or order its withdrawal while an assessment is under way. The full text of Regulation (EU) 2024/1689 spells out each of those steps, and the Commission's AI Act implementation page tracks the timeline.
The August 2026 cliff and the 2027 carve-out
Most Annex III high-risk obligations become applicable on 2 August 2026, which is why regulators are issuing first compliance actions now. Organisations that have waited for the final date are discovering that compiling the required documentation for even one production system is not a two-week exercise. The carve-out is narrower than it sounds: AI systems already regulated as safety components under existing EU product legislation, such as medical devices and machinery, get an additional year, until 2 August 2027. That delay applies to the classification route in Article 6(1), not to every AI system those companies use.
The same distinction shows up in current export-control debates. AI chip restrictions are forcing companies to track which models can be trained where, a parallel compliance burden to the documentation duties under the AI Act. AcademicJobs has covered the U.S.-China AI chip export restrictions and the capital spending scrutiny hitting large AI infrastructure.
What this means for teams building or deploying AI
One product lead at a German medical-device supplier, call her Dr. M, spent eight weeks mapping every production model against the Act's classification rules. Three of her eleven models sat inside Annex I as safety components and had never been through a notified-body assessment. She had treated 2 August 2027 as a distant concern. The inventory showed the exposure had already begun, because the documentation and risk-management habits needed by 2027 take longer to build than the legal deadline suggests.
The base rate is reassuring: most AI systems are not high-risk. The exception is common enough, though, because safety components hide inside existing products and because some general-purpose model providers do not yet have the documentation habit. What regulators are announcing is not a campaign against AI in general; it is a campaign against undocumented risk.
The practical first step is an inventory. List every AI system and general-purpose model your organisation builds, buys, or integrates. Classify each one against the prohibited practices list and the two high-risk routes. Check whether the provider has filed technical documentation and, where required, registered the system. Do that before the end of the month, while authorities are still sending letters rather than imposing fines.
Photo by Guillaume Périgois on Unsplash