How EU AI Act Fines Are Calculated: Tiers, Factors, and SME Rules

How Are Fines Calculated Under the EU AI Act

Article 99 of Regulation (EU) 2024/1689 sets the penalty framework, and its first principle is often missed: penalties must be effective, proportionate, and dissuasive, and the headline numbers are ceilings, not tariffs. No breach automatically costs the maximum. Member state authorities, and the European Commission for general-purpose AI model providers, work within three tiers defined by the type of violation, then calibrate the actual amount using a list of mandatory considerations. Understanding the mechanism matters for two reasons: it tells you where your worst exposure sits, and it shows that the factors you control, cooperation, prompt correction, documented good faith, directly move the number.

The Three Penalty Tiers

Worldwide turnover means the whole undertaking's global revenue, not just EU sales and not just the product line at issue, which is what gives the percentages their force for large groups.

The SME Reversal: Whichever Is Lower

For small and medium-sized enterprises, including startups, the Act flips the comparison: each fine is capped at whichever of the fixed amount or the turnover percentage is lower. The effect is substantial. A company with 5 million euros of annual turnover that breaches a high-risk obligation faces a ceiling of 3 percent of turnover, 150,000 euros, rather than the 15 million euro figure, because the percentage is the lower of the two. For a large multinational, the same breach is capped by whichever is higher, so the percentage of global revenue dominates. The SME definition follows the standard EU recommendation: fewer than 250 employees and turnover up to 50 million euros or balance sheet up to 43 million euros. This reversal is the single most important penalty fact for small businesses, and it was placed in the Act precisely so that proportionality is structural rather than discretionary.

The Factors Authorities Must Weigh

Article 99 requires authorities, when deciding whether to fine and how much, to take all relevant circumstances into account, including a specific list. The nature, gravity, and duration of the infringement and of its consequences, including the purpose of the system and the number of people affected and the level of damage. Whether other authorities have already fined the same operator for the same infringement. The size, annual turnover, and market share of the operator. Any financial benefit gained or loss avoided through the breach. The degree of cooperation with authorities to remedy the violation and mitigate effects. The degree of responsibility, taking into account the technical and organisational measures the operator had implemented. How the authority learned of the breach, in particular whether the operator self-reported. Intent or negligence. And any action taken to mitigate harm to affected persons. Read as a checklist, this list is effectively the regulator's published guide to reducing your own penalty: measures in place before the breach, fast cooperation afterward, and self-reporting all sit on the mitigation side of the scale.

Who Enforces and Who Fines Whom

Enforcement is split. National market surveillance authorities, designated by each member state, supervise AI systems and impose the tiered fines above on providers, deployers, importers, distributors, and authorised representatives; member states had to lay down their penalty regimes and could add further measures within the Act's frame. The European Commission, through the AI Office, exclusively supervises providers of general-purpose AI models, with fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher, for violations of the model obligations, for failing to provide requested information, or for non-compliance with requested measures. EU institutions themselves fall under a separate, lower scale enforced by the European Data Protection Supervisor, with fines up to 1.5 million euros for prohibited practices. The penalties framework attached to the governance chapter became applicable with it from August 2, 2025, while the tiers tied to obligations that begin in August 2026 bite from that date.

Fines Are Not the Only Cost

The monetary tiers sit inside a wider enforcement toolkit that often hurts more. Market surveillance authorities can require corrective action within a deadline, restrict or prohibit a system's availability, and order withdrawal or recall. For a software business, an order to stop offering a flagship product to the EU market converts directly into lost revenue that dwarfs most realistic fines. Non-compliance findings are also discoverable by customers: business deployers have their own duties and contractual reasons to drop non-compliant vendors, and procurement processes increasingly ask about regulatory history. Add private consequences, contract claims from customers whose own compliance was undermined, and the conclusion is that the fine schedule, dramatic as the numbers look, is the smaller half of the risk picture.

Stacking with GDPR and Other Regimes

The same incident can violate several laws at once, and the penalties accumulate across regimes. A discriminatory hiring system that also processed personal data unlawfully exposes the operator to AI Act fines from the market surveillance authority and GDPR fines, up to 20 million euros or 4 percent of worldwide turnover, from the data protection authority. Sector regulators in finance or healthcare may add their own measures. The Article 99 factor about fines already imposed by other authorities for the same infringement tempers pure double-counting within the AI Act's own scope, but it does not erase parallel liability under separate laws protecting separate interests. Organisations should therefore treat AI incidents as multi-regime events from the first hour, coordinating their response across privacy, AI, and sector compliance rather than handling them in separate lanes.

Practical Conclusions for Budgeting Risk

Four planning points follow from the mechanics. First, locate your worst case by tier: anything near the Article 5 list is in the 7 percent tier and deserves immediate screening, while most operational compliance debt sits in the 3 percent tier. Second, if you are an SME, compute your actual ceilings, percentage of your real turnover, because the realistic numbers transform board conversations from panic to planning. Third, invest in the mitigating factors in advance: documented measures, monitoring that detects problems early, and a rehearsed cooperation posture are penalty reducers written into the law. Fourth, remember that the cheapest fine is the one that never matures: classification done early, vendor documents collected, and transparency notices in place cost a fraction of any tier. The penalty framework rewards exactly the organisations that engaged with the Act before being asked.

Worked Examples: What the Ceilings Mean at Three Company Sizes

Percentages become concrete the moment you run them against a turnover figure. The table below applies the Article 99 rules mechanically to three illustrative companies: a small business with 10 million euros of worldwide annual turnover that qualifies as an SME, a mid-sized firm with 300 million euros that does not, and a large group with 2 billion euros. These are maximum ceilings computed from the rules as written, not predictions of what any authority would actually impose in a given case, since the calibration factors discussed above pull real penalties below the ceiling.

Three lessons sit in this table. For genuine SMEs, the lower-of rule keeps every ceiling proportionate to actual size, and computing your own numbers usually shrinks the perceived threat to something a risk register can hold. For mid-sized companies above the SME thresholds, the fixed amounts are normally the binding figure, which means exposure stops scaling with revenue until the company is very large. For large groups, the percentages dominate, and because the base is the total worldwide turnover of the whole undertaking, a breach in one small product line is priced against the entire group's revenue.

How an Enforcement Case Typically Unfolds

A fine is the end of a process, not a lightning strike, and each stage of that process offers a chance to change the outcome. A case usually opens with a trigger: a complaint from an affected person, a referral from another regulator, a market surveillance sweep of a sector, or a serious incident report. The authority then gathers information, and this step deserves respect, because supplying incorrect, incomplete, or misleading answers is itself the third penalty tier; accuracy in replies is a compliance obligation in its own right. If problems are found, a corrective phase follows, in which the authority can require fixes within a deadline or restrict a system's availability, and full cooperation here feeds directly into the Article 99 factors. Only then comes a penalty decision, reasoned against those factors, with rights of defence and appeal governed by national procedure. Organisations that prepare for this sequence in advance, with a named contact point, a rehearsed information-response process, and clean documentation, consistently fare better at every stage than those improvising after the first letter arrives.

The Structural Difference from GDPR Penalties

Comparisons with GDPR help calibrate expectations, because most businesses already carry a mental model from data protection. GDPR works with two main ceilings, 20 million euros or 4 percent of worldwide turnover for the most serious violations and 10 million euros or 2 percent for others, in both cases whichever is higher, and it contains no structural SME reversal; proportionality for small firms under GDPR arrives only through discretionary weighing. The AI Act uses three tiers, tops out higher at 35 million euros or 7 percent, and writes the SME lower-of rule directly into Article 99. The practical reading: for a large company the AI Act's worst tier is materially heavier than GDPR's, while for a small company the AI Act's arithmetic is structurally gentler, provided SME status genuinely applies.

A Self-Check Before Any Authority Asks

Five questions test whether penalty exposure is managed rather than merely feared. First, can you produce a dated record showing every AI system was screened against the Article 5 prohibitions, the 7 percent tier? Second, do you know, per system, which tier a failure would fall into, so the risk register prices each gap correctly? Third, are the mitigating factors already documented: training records, oversight appointments, monitoring that would catch a problem early? Fourth, could you answer a detailed information request accurately within a short deadline without heroics? Fifth, do you know which national market surveillance authority supervises you, and have you read its published guidance? An organisation that can answer yes five times has, in effect, already done the work that moves a hypothetical penalty from the ceiling toward the floor.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.