Last updated: August 27, 2026
The EU AI Act sets three main tiers of administrative fines. Prohibited AI practices under Article 5 can draw up to 35,000,000 EUR or 7 percent of total worldwide annual turnover for the preceding financial year, whichever is higher. Most other obligations on providers, deployers, importers, distributors and notified bodies top out at 15,000,000 EUR or 3 percent, and supplying incorrect, incomplete or misleading information to authorities tops out at 7,500,000 EUR or 1 percent. Separate and much lower caps apply to EU institutions, and providers of general purpose AI models face a dedicated regime enforced directly by the European Commission.
This post is part of our AI governance guide. For the wider regulatory picture, see our overview of EU AI Act compliance.
Key takeaways
- Three headline tiers apply: 35,000,000 EUR or 7 percent, 15,000,000 EUR or 3 percent, and 7,500,000 EUR or 1 percent of worldwide annual turnover.
- For undertakings, the regulator applies whichever figure is higher. For SMEs and startups, Article 99(6) flips that and applies whichever is lower.
- The penalty framework has been applicable since August 2, 2025. It is not a future problem.
- From August 2, 2026 the European Commission can fine providers of general purpose AI models up to 15,000,000 EUR or 3 percent under Article 101.
- The Digital Omnibus on AI, Regulation (EU) 2026/1744, deferred several high-risk deadlines but left the fine amounts in Article 99 untouched.
- Fines are issued by national market surveillance authorities, except for EU institutions (the EDPS, under Article 100) and general purpose AI models (the Commission).
What are the EU AI Act penalty tiers?
Article 99 of the AI Act creates a graduated structure. The severity of the fine tracks the severity of the conduct, with prohibited practices at the top and administrative failures at the bottom.
| Tier | What triggers it | Maximum administrative fine |
|---|---|---|
| Tier 1 | Non-compliance with the prohibited AI practices in Article 5, such as untargeted scraping of facial images, social scoring, or manipulative techniques that cause significant harm | 35,000,000 EUR or 7 percent of total worldwide annual turnover, whichever is higher |
| Tier 2 | Non-compliance with most other obligations placed on providers, authorised representatives, importers, distributors, deployers and notified bodies, including the requirements attached to high-risk systems and the transparency duties in Article 50 | 15,000,000 EUR or 3 percent of total worldwide annual turnover, whichever is higher |
| Tier 3 | Supplying incorrect, incomplete or misleading information to notified bodies or national competent authorities in reply to a request | 7,500,000 EUR or 1 percent of total worldwide annual turnover, whichever is higher |
| EU institutions | Union institutions, bodies, offices and agencies, fined by the European Data Protection Supervisor under Article 100 | 1,500,000 EUR for prohibited practices, 750,000 EUR for other non-compliance |
| General purpose AI models | Intentional or negligent infringement of GPAI obligations, misleading information, or failure to comply with Commission requests, evaluations or corrective measures, under Article 101 | 15,000,000 EUR or 3 percent of total worldwide annual turnover, whichever is higher |
Two details matter more than the headline numbers. First, the percentage is calculated on total worldwide annual turnover, not EU turnover, so a company with modest European revenue can still face a fine sized against its global books. Second, the caps are maximums, not tariffs. Most enforcement is expected to land well below them.
Who actually issues the fines?
Enforcement is decentralised. Each Member State designates national competent authorities, including at least one market surveillance authority, and lays down its own rules on penalties within the ceilings the Act sets. That means the procedure, the appeal route and the practical appetite for large fines will vary between Member States even though the ceilings do not.
There are two carve-outs. The European Data Protection Supervisor fines EU institutions, bodies, offices and agencies under Article 100, with substantially lower caps of 1,500,000 EUR and 750,000 EUR. And the European Commission, acting through the AI Office, enforces directly against providers of general purpose AI models under Article 101 rather than leaving that to national regulators.
When did EU AI Act penalties start applying?
The Act entered into force on August 1, 2024 and applies in stages. Prohibited practices and AI literacy duties have applied since February 2, 2025. Obligations on general purpose AI models and the penalty framework itself became applicable on August 2, 2025. From August 2, 2026 the transparency obligations in Article 50 apply and the Commission gained its enforcement and fining powers over general purpose AI model providers.
The Digital Omnibus on AI was published in the Official Journal on July 24, 2026 and entered into force on July 27, 2026. It moved the application date for standalone high-risk systems under Annex III from August 2, 2026 to December 2, 2027, and for AI embedded in regulated products under Annex I to August 2, 2028. It did not reduce any fine. For the full sequence of dates, see our EU AI Act compliance timeline.
How do regulators decide the size of a fine?
Article 99 requires penalties to be effective, proportionate and dissuasive, and lists the factors authorities weigh when setting an amount. Expect a regulator to look at:
- The nature, gravity and duration of the infringement, and the number of people affected and the level of damage they suffered.
- Whether another authority has already fined the same operator for the same conduct.
- The size, annual turnover and market share of the operator.
- Whether the infringement was intentional or negligent.
- Any action taken to mitigate the harm, and the degree of cooperation with the authority.
- Whether the operator is a provider or a deployer, and whether the system had actually been placed on the market or put into service.
- Any relevant previous infringements.
In practice, documentation is what separates a cooperative operator from a negligent one. An organisation that can produce a dated risk assessment, a record of human oversight decisions and evidence of vendor due diligence is arguing from a very different position than one that cannot.
Do SMEs and startups face the same numbers?
No. Article 99(6) provides that for SMEs, including startups, each fine is capped at the lower of the two figures rather than the higher. A small company with 2,000,000 EUR of turnover facing a Tier 2 infringement is looking at a ceiling of 60,000 EUR, not 15,000,000 EUR. The obligation to comply is identical, but the financial exposure scales with the business.
Which organisations are actually exposed?
Penalty exposure follows the Act’s scope, which is broader than many teams assume and reaches operators outside the EU whose AI output is used inside it. Whether you are a provider or a deployer changes both which obligations attach and how a fine is calculated, so the classification question comes first. Our guides on who the EU AI Act applies to and the high-risk categories under Annex III walk through both questions.
How do you reduce penalty exposure?
The controls that reduce exposure are unglamorous and mostly organisational rather than technical:
- Maintain a live AI system inventory. You cannot classify or defend what you have not catalogued, and shadow AI is where prohibited-practice risk usually hides.
- Classify every system against Article 5 first. Tier 1 exposure is the one that scales to 7 percent, and Article 5 has applied since February 2025.
- Fix the transparency basics now. Chatbot disclosure, synthetic content marking and deepfake labelling under Article 50 are in scope from August 2, 2026 and are cheap to get right.
- Re-paper vendor contracts. If a supplier’s model sits inside your product, their documentation gaps become your evidence gaps.
- Keep records that survive a request. Tier 3 exists specifically for incomplete or misleading answers to authorities.
Most organisations do not want a separate control set for every AI regulation. A certifiable AI management system gives you one place to keep the risk assessments, roles, oversight records and supplier controls that regulators ask for, and it maps cleanly onto the AI Act’s expectations. If you are building that foundation, our guide to ISO 42001 covers what the standard requires and how Compyl automates the evidence collection behind it, so your AI governance program produces audit-ready records as a by-product of normal operations rather than a fire drill.
Frequently asked questions
What is the maximum fine under the EU AI Act?
35,000,000 EUR or 7 percent of total worldwide annual turnover for the preceding financial year, whichever is higher. That ceiling applies only to breaches of the prohibited practices in Article 5.
Are EU AI Act fines calculated on group revenue or entity revenue?
On the total worldwide annual turnover of the undertaking. The AI Act uses the EU competition law concept of an undertaking, which generally means the wider economic group rather than the single legal entity that committed the infringement.
Did the Digital Omnibus reduce EU AI Act fines?
No. Regulation (EU) 2026/1744 deferred several high-risk application dates and adjusted parts of the framework, but the fine tiers and amounts in Article 99 were not changed.
Can companies outside the EU be fined?
Yes. The Act applies to providers placing systems on the EU market regardless of where they are established, and to providers and deployers outside the EU where the output produced by the system is used in the EU.
Have any EU AI Act fines been issued yet?
Enforcement activity has been limited so far. The penalty framework has been applicable since August 2, 2025, and national authorities in several Member States were still standing up their supervisory structures through 2026, so early activity has focused on information requests rather than headline fines.
Do the penalties apply to internal AI tools?
Yes, if you are a deployer within scope. Deployer obligations attract Tier 2 exposure, and an internally built system that meets a prohibited-practice definition carries the same Tier 1 exposure as a commercial product.
