The EU AI Act fine depends on the breached provision, the liable actor, and the facts used to set the penalty. Article 99 has three main administrative-fine tiers: EUR 35 million or 7%, EUR 15 million or 3%, and EUR 7.5 million or 1%. The percentage is based on an undertaking’s total worldwide annual turnover for the preceding financial year. Each figure is a ceiling. The authority still has to apply the Regulation’s proportionality factors and national procedure. This reference reflects Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, which entered into force on 27 July 2026. General information. This article does not provide legal advice.
Article 99 penalty tiers at a glance
The fixed amount applies to a person or body that is not treated as an undertaking for the turnover calculation. For an undertaking, the authority uses the higher of the fixed amount and the turnover percentage, subject to the special caps for smaller enterprises below.
| Violation | Maximum fine | Liable roles | Smaller-enterprise cap |
|---|---|---|---|
| Non-compliance with Article 5 prohibited AI practices | EUR 35 million or 7% of worldwide annual turnover, whichever is higher | The actor responsible for the prohibited practice | SMEs, including start-ups: the lower amount. No SMC reduction for this tier. |
| Non-compliance with the duties listed in Article 99(4) | EUR 15 million or 3% of worldwide annual turnover, whichever is higher | Providers, authorised representatives, importers, distributors, deployers, notified bodies, and providers or operators subject to Article 25(2) or (4) | SMEs, including start-ups, and SMCs: the lower amount |
| Incorrect, incomplete, or misleading information supplied to a notified body or national competent authority in reply to a request | EUR 7.5 million or 1% of worldwide annual turnover, whichever is higher | The person or undertaking replying to the request | SMEs, including start-ups, and SMCs: the lower amount |
Maximum fine means a ceiling
Article 99 requires penalties to be effective, proportionate, and dissuasive. Member States may also use warnings and non-monetary measures. National law supplies the detailed penalty rules, enforcement powers, and procedural safeguards.
When an authority chooses the amount, it considers the nature, gravity, and duration of the infringement; its consequences and the number of people affected; the organisation’s size, turnover, and market share; benefits gained or losses avoided; cooperation; responsibility and corrective measures; disclosure; intent or negligence; and mitigation. It also considers fines already imposed by other authorities for the same infringement or activity.
The administrative-fine schedule does not determine civil liability, criminal liability, compensation, or exposure under another law. Public authorities and bodies may be subject to national penalty rules to the extent each Member State decides.
EUR 35 million or 7%: prohibited AI practices
The highest Article 99 tier covers non-compliance with Article 5. It reaches the actor that carries out or places on the market a prohibited practice. Examples in the original Act include harmful manipulation, exploitation of vulnerability, social scoring, certain predictive policing, untargeted facial-image scraping, emotion recognition in workplaces and schools, certain biometric categorisation, and real-time remote biometric identification in public spaces outside the stated exceptions.
The original Article 5 prohibitions have applied since 2 February 2025. Regulation (EU) 2026/1744 adds prohibitions concerning AI systems intended to generate child sexual abuse material and AI systems intended to generate or manipulate intimate content without consent. Those additions apply from 2 December 2026.
A contractual label does not settle which statutory actor is responsible. Teams should identify the real activity, the party performing it, and any provider, deployer, importer, distributor, or product-manufacturer role created by the Act.
EUR 15 million or 3%: listed duties and liable roles
Article 99(4) is an enumerated list. It assigns the middle tier to the duties below. A provider and deployer can each face enforcement for their own obligations in the same system. Liability follows the breached duty and the actor that owed it.
| Liable role | Provision | Covered duty |
|---|---|---|
| Provider | Article 16 | The provider obligations for high-risk AI systems |
| Authorised representative | Article 22 | The mandate, document, cooperation, and verification duties assigned to the representative |
| Importer | Article 23 | Importer checks and corrective-action duties |
| Distributor | Article 24 | Distributor checks and corrective-action duties |
| Providers and operators subject to Article 25 | Article 25(2) and (4) | The cooperation and information duties added to this tier by Regulation (EU) 2026/1744 |
| Deployer | Article 26 | The deployer obligations for high-risk AI systems |
| Notified body | Articles 31, 33(1), 33(3), 33(4), and 34 | Eligibility, information, document, access, and operational duties |
| Provider or deployer | Article 50 | Applicable transparency duties for interactive and synthetic-content AI |
Article 27 and duties outside the middle-tier list
The Article 27 fundamental rights impact assessment is not named in the Article 99(4) list. Article 99(1) still requires Member States to set penalties and other enforcement measures for infringements of the Regulation. The applicable national measure needs to be checked in the relevant jurisdiction. The same care applies to any duty that is outside the enumerated middle tier.
Use the Article 26 deployer obligations guide and the FRIA template to separate the underlying duties before assigning a penalty tier.
EUR 7.5 million or 1%: information supplied in reply to a request
The lowest Article 99 tier is precise. It covers supplying incorrect, incomplete, or misleading information to a notified body or national competent authority in reply to a request. The maximum is EUR 7.5 million or 1% of worldwide annual turnover, whichever is higher for an undertaking.
The request condition is part of the infringement. Teams should preserve the incoming request, scope, deadline, submitted response, supporting records, reviewer, and any correction. That record helps establish what the authority asked and what the organisation supplied.
Regulation (EU) 2026/1744 also gives the AI Office direct supervisory powers in specified cases. Its new framework applies the Article 99(3) to (7) rules with necessary adaptations and brings incorrect, incomplete, or misleading information supplied in reply to an AI Office request within the requested-information tier.
Provider, deployer, GPAI, and public-body exposure
A single AI service can create duties for several actors. The table separates the main penalty routes.
| Actor | Article 99 exposure | Separate route to check |
|---|---|---|
| Provider of a high-risk AI system | Article 5 if responsible for a prohibited practice; Article 16 in the middle tier; requested-information tier when its reply meets Article 99(5) | A downstream modification may create or transfer a provider role under Article 25 |
| Deployer of a high-risk AI system | Article 5 if responsible for a prohibited practice; Article 26 in the middle tier; requested-information tier when its reply meets Article 99(5) | Article 27 national enforcement where a fundamental rights impact assessment is required |
| Provider or deployer subject to Article 50 | Applicable Article 50 transparency duties sit in the middle tier | Article 50 applies from 2 August 2026 |
| Provider of a general-purpose AI model | Article 99 can apply to conduct within its scope | Article 101 separately lets the Commission fine GPAI model providers up to EUR 15 million or 3% from 2 August 2026 |
| EU institution, body, office, or agency | Article 100 has a separate schedule administered by the European Data Protection Supervisor | Up to EUR 1.5 million for Article 5 and EUR 750,000 for other infringements |
| National public authority or body | Member State law determines the extent to which it can receive an administrative fine | Check the implementing and procedural law in the relevant Member State |
How Regulation (EU) 2026/1744 changes Article 99
The 2026 amendment entered into force on 27 July 2026. It keeps the three headline fine tiers and makes targeted changes to their coverage.
- Article 99(1) now expressly describes administrative fines, warnings, and non-monetary measures as available penalty forms for any infringement.
- Selected cooperation and information duties under Article 25(2) and (4) enter the EUR 15 million or 3% tier.
- The penalty-setting rule now names small mid-cap enterprises, or SMCs, alongside SMEs and start-ups when authorities consider economic viability.
- An SMC receives the lower of the fixed amount and percentage for the middle tier and the requested-information tier. The SMC reduction does not extend to the Article 5 tier.
- The amendment connects the three-tier schedule to specified AI Office supervision and to requested information supplied to the AI Office.
Enforcement timeline after the 2026 amendment
A fine provision can be in force before the underlying duty applies. An authority needs an applicable substantive obligation and an infringement of it. The dates below keep those two questions separate.
| Date | What applies |
|---|---|
| 2 February 2025 | The original Article 5 prohibited practices |
| 2 August 2025 | Chapter XII penalties, including Article 99, except the deferred Article 101 GPAI fining power |
| 27 July 2026 | Regulation (EU) 2026/1744 enters into force and amends Article 99 |
| 2 August 2026 | Article 50 transparency duties and the Article 101 Commission fining power for GPAI model providers |
| 2 December 2026 | The new Article 5 prohibitions and the transition date for machine-readable marking of certain generative-system outputs |
| 2 December 2027 | Chapter III, Sections 1 to 3 duties for stand-alone Annex III high-risk AI systems |
| 2 August 2028 | Chapter III, Sections 1 to 3 duties for high-risk AI systems embedded in Annex I products |
Build the evidence before an authority asks
Start with a role and duty register. Record whether the organisation is provider, deployer, importer, distributor, authorised representative, product manufacturer, or another Article 25 actor for each system. Link each applicable duty to an owner, control, operating record, and review cadence. Preserve regulator and notified-body requests as governed cases with their submitted responses and supporting records.
The KLA Control Plane can operationalise part of that control design for AI-agent actions. The KLA Policy Engine evaluates an action and returns allow, warn, require approval, or block. A require-approval result creates a Decision Request in Decision Desk. Each governed run produces a Lineage Record with the policy version, verdict, matched rules, reason, and action context. Evidence Room and Sealed Evidence Bundles support review and export of those records, while Control Mapping links controls to framework requirements.
KLA does not determine an organisation’s legal role, decide whether a duty applies, set a fine, or guarantee compliance. Legal analysis and the organisation’s wider compliance records remain necessary.
Official sources and related AI Act guides
Primary legal texts: Regulation (EU) 2024/1689 in the Official Journal and the amending Regulation (EU) 2026/1744 in the Official Journal. Official supporting material: Article 99 at the Commission AI Act Service Desk and the Council adoption record for the 2026 amendment.
For classification and scope, use the EU AI Act hub, 2026 requirements guide, provider and deployer guide, and high-risk classification guide.
For duty-level work, use the Annex IV documentation guide, Article 17 checklist, Article 26 deployer guide, FRIA template, registration guide, Article 50 checklist, and post-market monitoring plan.
For the amended dates, read the Digital Omnibus deadline guide and 2 August 2026 guide.
Frequently Asked Questions
What is the maximum fine under the EU AI Act?
The highest Article 99 ceiling is EUR 35 million or 7% of an undertaking’s total worldwide annual turnover for the preceding financial year, whichever is higher. It applies to non-compliance with Article 5 prohibited AI practices.
When does the 7% EU AI Act fine apply?
The 7% tier applies to non-compliance with Article 5 prohibited AI practices. The authority still applies the proportionality factors in Article 99 and the relevant national procedure.
Can a provider and deployer both be fined?
Yes. A provider and deployer can each face enforcement for duties assigned to their role. The authority must identify the applicable obligation, the actor that owed it, and that actor’s infringement.
What triggers the EUR 7.5 million or 1% tier?
It covers incorrect, incomplete, or misleading information supplied to a notified body or national competent authority in reply to a request. Regulation (EU) 2026/1744 extends the framework to information supplied in reply to specified AI Office requests.
Do SMEs and small mid-cap enterprises have lower caps?
SMEs, including start-ups, receive the lower of the fixed amount and percentage in all three Article 99 tiers. Small mid-cap enterprises receive that lower cap for the EUR 15 million or 3% tier and the EUR 7.5 million or 1% requested-information tier. They do not receive it for the Article 5 tier.
Are GPAI model provider fines part of Article 99?
Article 101 creates a separate Commission fining power for providers of general-purpose AI models, capped at EUR 15 million or 3% of worldwide annual turnover. It applies from 2 August 2026.
When did Article 99 start to apply?
Chapter XII, including Article 99, applied from 2 August 2025. Regulation (EU) 2026/1744 amended Article 99 and entered into force on 27 July 2026. A fine still depends on the underlying duty being applicable when the infringement occurs.
Key Takeaways
Treat the Article 99 table as the start of the analysis. Confirm the actor, applicable duty, application date, Member State procedure, and facts that affect proportionality. Keep the evidence that shows how each control operated and how every requested response was prepared.