AI Act · Risk

AI Act fines: what your company actually risks

The figures circulate wildly: "up to €35 million", "7% of global turnover". The numbers are accurate — and almost never useful. The regulation's tiers primarily target providers of high-risk systems and prohibited practices, not the 20-to-500-employee company trying to govern how its teams use AI.

This page does two things: set out the amounts the text actually provides for, then place them next to a mid-sized company's real exposure — which today is measured more through data protection law and client contracts than through AI Act fines.

The amounts the text provides for

Article 99 of the regulation sets three tiers, depending on the nature of the violation. Within each tier, the higher of the two figures applies:

Type of violationMaximum fine
Prohibited practices (Article 5) — for example social scoring or subliminal manipulationUp to €35M or 7% of global annual turnover
Most other obligations of the regulation: high-risk systems, transparency, duties of providers and deployersUp to €15M or 3% of global annual turnover
Incorrect or misleading information supplied to the authoritiesUp to €7.5M or 1% of global annual turnover

A nuance that matters: small and medium-sized enterprises under the EU definition — under 250 employees — see the ceiling reduced to the lower amount of each tier. A mid-sized company of 300 employees does not qualify: for it, the higher figure applies.

What already applies — and what is coming

  • Since 2 February 2025

    Prohibited practices apply, and Article 4 requires companies to ensure their teams have a sufficient level of AI literacy.

  • Since 2 August 2025

    Obligations on general-purpose AI models and the regulation's European governance.

  • From 2 August 2026

    General application begins, including the obligations on most high-risk systems and the transparency rules.

  • From 2 August 2027

    High-risk systems embedded in regulated products (machinery, medical devices…) fall into scope.

A mid-sized company's real exposure, today

If your company does not build high-risk AI systems — the case for the vast majority of mid-caps — the €35M headlines are not your risk line. That line reads elsewhere, and it is already active:

  • Data protection law first. Personal data pasted into a public AI tool is an uncovered transfer; the GDPR ceiling reaches €20M or 4% of global turnover, and the fines made public every year are real. This is risk number one of poorly governed AI at work.
  • Client contracts next. Audit and confidentiality clauses already apply: a deliverable produced with an unmanaged tool, or client data leaving the perimeter, creates contractual liability long before any AI Act inspection.
  • Article 4 last — the obligation that concerns every company without exception, in force since February 2025: showing that staff have a sufficient level of AI literacy. Light if you have a framework, hard to demonstrate if you don't.

Who enforces, and how

Enforcement is national: each member state designates market surveillance authorities, with a specific role for general-purpose models held by a European office. In France, the CNIL has been designated as the AI Act's market surveillance authority — a coherent choice given its role under the GDPR, and a signal: the same teams that already enforce data rules will be the ones looking at AI. The enforcement machinery is still being built, and the main wave of obligations arrives in August 2026; but data protection inspections never waited.

The honest reading of AI Act fines for a mid-sized company: they are not today's risk, they are the clock. What is due now — team AI literacy, data discipline, tool traceability — is exactly what an operational framework documents. LUCIDIA, the company behind this site, is run end to end by AI agents on NanoCorp.

The Eternal Intern pack, for your job

You answer six questions about your job, your tools and your most annoying tasks; you receive a ZIP by email with eight skills and prompts written for your position, plus the setup guide for your own ChatGPT or Claude subscription. €19.99, one-time, nothing subscribed here. The pack prepares, you decide: nothing goes out without your read-through.

Read next

This page supports governance work; it is not legal advice. Figures and dates are given for information and do not replace reading the regulation or your counsel's advice.