EU AI Act Fine Calculation: How Regulators Actually Set the Amount

A €35 million headline figure grabs attention. What it hides is that EU AI Act fine calculation is a structured process with multiple stages, and the inputs at each stage are things your business can actually influence before an investigation begins.

Understanding the mechanics is not just useful for lawyers. HR directors deciding whether to deploy an AI screening tool, compliance officers building an audit trail, and managing directors signing off on AI procurement all need to know how enforcement math works.

# The Three Penalty Tiers

The AI Act establishes three distinct fine ceilings in Article 99.

Tier 1 applies to the most serious violations: deploying a prohibited AI system listed in Article 5, such as social scoring or real-time biometric surveillance in public spaces. The ceiling is €35 million or 7% of total annual worldwide turnover, whichever is higher.

Tier 2 covers non-compliance with obligations that apply to high-risk AI systems. These include failures around conformity assessments, technical documentation, human oversight, or transparency requirements. The ceiling is €15 million or 3% of global turnover.

Tier 3 targets supplying incorrect, incomplete, or misleading information to regulators and notified bodies. The ceiling is €7.5 million or 1.5% of global turnover.

Note the word "or" in each tier. Regulators take whichever figure is higher, not lower. A small company with €10 million in annual revenue that commits a Tier 2 violation faces a ceiling of €15 million, not €300,000. The percentage cap only provides relief for very large multinationals where 3% of global turnover exceeds the fixed amount.

# How the Starting Point Is Set

The European Commission's approach, mirrored in guidance from national market surveillance authorities, follows a logic borrowed from GDPR enforcement. Regulators begin with the gravity and nature of the infringement, then adjust upward or downward based on specific circumstances.

Gravity assessment looks at three things:

1. What category of system was involved? A prohibited system triggers Tier 1 regardless of how briefly it was used. A high-risk system in the wrong configuration triggers Tier 2.
2. How many people were affected? A recruitment AI that screened 500 candidates unlawfully is treated differently from one that screened 50,000.
3. What actual harm occurred? Job rejections, loan denials, or medical triage errors each carry different weight.

The AI Act's recital 161 explicitly instructs supervisory authorities to consider the nature, gravity, and duration of the infringement when setting fines. Duration matters more than most deployers expect. Running a non-compliant system for 14 months is dramatically worse than running it for 6 weeks before correcting course.

# The "Turnover" Definition

One technical point that catches SMEs off guard: "total annual worldwide turnover" is measured at the group level, not the entity level. If your Dutch operating company is owned by a holding structure, regulators will look at consolidated group revenue. A subsidiary with €8 million in local revenue but a parent group with €400 million in turnover faces percentage-based fines calculated on €400 million.

This is consistent with how Dutch and Belgian data protection authorities already apply GDPR Article 83, and there is no reason to expect AI Act enforcement to diverge.

# Aggravating Factors That Push Fines Up

Regulators applying Article 99 have explicit authority to consider aggravating circumstances. Based on both the Act's text and enforcement patterns established under GDPR by authorities including the Autoriteit Persoonsgegevens (AP), the following factors reliably increase final amounts:

• Intentional or negligent conduct. Deliberately disabling a human oversight mechanism is treated far more harshly than a configuration error.
• Refusal to cooperate. Stonewalling investigators, providing incomplete documentation, or delaying responses multiplies the base amount.
• Prior violations. If your organisation has received warnings or fines under GDPR for related AI processing, that history is admissible.
• Obstructing access. Denying regulators access to technical logs or training data is an aggravating factor in its own right, separate from the original violation.
• Targeting vulnerable groups. AI systems that affect children, people with disabilities, or asylum seekers receive heightened scrutiny.

# Mitigating Factors That Push Fines Down

The same framework cuts the other way. Documented mitigation can meaningfully reduce final amounts, sometimes by half or more based on GDPR precedent from the European Data Protection Board's guidelines on calculating administrative fines.

The most effective mitigating factors are:

• Proactive disclosure. Reporting an incident before regulators discover it independently carries significant weight.
• Immediate corrective action. Disabling the non-compliant system the moment a problem is identified, with documented timestamps, demonstrates good faith.
• Existing compliance infrastructure. A written AI governance policy, completed staff training records, and a system inventory all signal organisational maturity.
• Full cooperation during investigation. Providing complete documentation promptly, making staff available for interviews, and avoiding procedural challenges each contribute.
• Limited actual harm. If no individual suffered a concrete negative outcome, regulators have grounds to reduce the amount.

This is where the preparation you do now has a direct monetary value. A compliance officer who can produce a dated training log, a risk classification decision, and a supplier contract with Article 26 responsibilities assigned is in a structurally different negotiating position than one who cannot.

# What Enforcement Is Starting to Look Like

Formal AI Act enforcement against deployers begins applying in full from 2 August 2026 for high-risk systems in Annex III, but supervisory authorities are already building capacity. The AP in the Netherlands has published its AI supervision priorities and is actively auditing algorithmic systems under existing GDPR powers in the meantime.

The pattern from GDPR enforcement is instructive. When the AP fined a Dutch insurer for algorithmic fraud detection in 2023, the final amount was reduced substantially because the company had documented its human review process, even though that process was ultimately found inadequate. Documentation of intent and effort matters even when the outcome fails.

The EDPB's coordinated enforcement actions have consistently shown that the gap between maximum and actual fines is largest for organisations that cooperate fully and demonstrate they had a compliance programme, even an imperfect one. Organisations that treat enforcement as adversarial from day one tend to end up at the higher end of the range.

# The SME Proportionality Provision

Article 99(6) gives national authorities explicit discretion to reduce fines for small and medium-sized enterprises, including startups. This is not automatic. You must demonstrate that the maximum amount would be disproportionate given your financial capacity.

In practice, this means submitting financial statements, showing that the fine would threaten the company's viability, and demonstrating that you acted in good faith. The proportionality argument is weakest when the violation is in Tier 1 (prohibited systems) and strongest when it involves a technical documentation failure in Tier 2 or 3.

For a company with 50 employees and €5 million in revenue, a well-documented proportionality submission combined with mitigating factors could realistically bring a theoretical Tier 2 maximum of €15 million down to a fraction of that. The EU legislator built this discretion in deliberately.

# The One Number You Can Control

Every factor that drives fines downward — proactive disclosure, cooperation, documented governance, limited harm — is something you build before the investigation starts, not during it. The fine calculation process rewards organisations that treated compliance as an ongoing operational function rather than a crisis response.

Start with knowing what AI systems you deploy and how they are classified. A system inventory, a risk tier for each tool, and a clear record of who is responsible for oversight costs almost nothing to build. It is the foundation every other mitigating factor rests on.

If you are not certain how your current AI deployments would be classified under Annex III or what obligations attach to them, the two-minute compliance check at comply.khairos.ai will give you a starting point — and the documentation trail that regulators look for when deciding where on the penalty scale your organisation lands.