Market Surveillance Under the EU AI Act: Who Investigates, and What to Expect
How national market-surveillance authorities investigate AI systems under the EU AI Act: information requests, corrective action, Article 99 fines.
A market-surveillance authority can ask you to prove an AI system meets the rules — in writing, with documentation, on a deadline. For AI systems, that authority is a national body each member state designates under Article 70 of the EU AI Act, Regulation (EU) 2024/1689; its investigative powers run from Article 74 onward. The general-purpose AI (GPAI) model layer is supervised separately, by the European AI Office. This guide covers what the national authority can do, what an investigation looks like, and how to be ready before the letter arrives.
Who the authority is, and what it covers
Each member state names one or more national competent authorities under Article 70, including the market-surveillance authority that polices AI systems placed or used on its territory. In Spain that lead role sits with AESIA (with AEPD, the CGPJ and others competent in their own domains); in other member states it is the body each government has designated. The single point of contact under Article 70 tells you who to deal with.
Market surveillance reuses the EU's existing product-safety machinery (Regulation (EU) 2019/1020) and layers the AI Act's specifics on top. The authority's remit is the system on the market — your obligations as a provider (Article 16) or deployer (Article 26), the high-risk requirements, the Article 50 transparency duties, and the Article 5 prohibitions. It does not supervise GPAI models; that is the AI Office's job, and fines on GPAI model providers are imposed by the Commission under Article 101, not by the national authority.
The powers: what an authority can actually do
From Article 74 onward, a market-surveillance authority can:
| Power | What it means in practice |
|---|---|
| Request information and documentation | Demand the technical documentation (Article 11 / Annex IV), the declaration of conformity (Article 47 / Annex V), logs (Article 12), and records of your assessment. |
| Access source code (conditionally) | Where testing and documentation are insufficient and on a reasoned request, gain access to training data and, exceptionally, source code — subject to confidentiality safeguards. |
| Evaluate the system | Test or inspect the system to check conformity, including in real-world conditions. |
| Require corrective action | Order you to bring a non-compliant system into conformity within a set period. |
| Restrict, withdraw or recall | Restrict availability, withdraw the system from the market, or order a recall where the risk warrants it. |
| Impose penalties | Apply the Article 99 fines, as laid down in national law. |
The authority can act on its own initiative, on a complaint, or after a serious-incident report reaches it.
The penalty tiers (Article 99)
National authorities apply three tiers, each "whichever is higher" of a fixed sum or a percentage of total worldwide annual turnover:
| Breach | Maximum fine | Paragraph |
|---|---|---|
| Article 5 prohibited practices | €35 million or 7% | 99(3) |
| Most other obligations (high-risk requirements, provider/deployer duties, Article 50) | €15 million or 3% | 99(4) |
| Incorrect, incomplete or misleading information to authorities or notified bodies | €7.5 million or 1% | 99(5) |
Two things SMEs should note. First, the third tier — supplying wrong or misleading information to the authority during an investigation — is its own breach: how you answer the request matters. Second, under Article 99(6) an SME or start-up is capped at the lower of the percentage or the fixed amount, not the higher.
How serious-incident reports feed in (Article 73)
Market surveillance is not only reactive paperwork. A provider of a high-risk system must report a serious incident to the relevant authority under Article 73 — death, serious harm to health, serious and irreversible disruption of critical infrastructure, breach of fundamental-rights obligations, or serious property/environmental harm — within tight deadlines. That report is often what opens a market-surveillance file. Post-market monitoring by the provider (Article 72) is the upstream duty that should catch the problem first; Article 73 is the notification when it does. Keeping the two straight matters — they are different articles with different triggers.
What an investigation looks like — and how to be ready
Most investigations begin with a written information request, not an inspector at the door. You are given a deadline to produce specified documents. The authority reviews them, may ask follow-ups or test the system, and then either closes the file, requires corrective action, or escalates to a restriction and a fine.
Readiness is mostly about having the record already assembled:
- A register of every AI system you provide or deploy, each with its Article 6 classification and the reason.
- The technical documentation (Article 11 / Annex IV) and declaration of conformity (Article 47) for any high-risk system.
- Logs (Article 12) and your post-market monitoring records (Article 72).
- Evidence of Article 50 transparency where it applies, and of Article 4 AI-literacy measures.
- A named owner who can answer the authority accurately — because a wrong or misleading answer is itself an Article 99(5) breach.
The companies that struggle are not usually the ones with a gap in the system; they are the ones who cannot, on a two-week deadline, find the document that shows the gap was assessed and handled.
Worked example: an SME facing an information request
Lumen Analytics, a 70-person analytics firm in Lisbon (€9M turnover), deploys a third-party CV-screening tool to shortlist candidates for its own hiring. A rejected applicant complains to the Portuguese authority. A market-surveillance information request lands: produce the system's classification, the deployer records, the human-oversight arrangements, and the transparency information given to candidates.
Because Lumen had registered the tool as high-risk (Annex III, employment), kept its Article 26 deployer records, logged the human review of each shortlist, and could show the Article 50 notice to candidates, it answered fully within the deadline. The file closed with no corrective action. Had Lumen instead improvised answers, the misleading-information tier (99(5)) would have been live on top of any underlying breach. The lesson: the register and the records, prepared in calm, are what turn a request into a non-event.
How Confir helps
Confir keeps each AI system classified, role-assigned and documented in one register, so an information request is a retrieval task rather than a scramble — the classification, the cited Article, the rule that fired, and the supporting documents are already in place. The engine is deterministic and rule-based: the same intake always produces the same finding, and the reason is human-readable, which is exactly what an authority's reasoned request is testing.
Related guides
- Who enforces the EU AI Act?
- Market-surveillance authority, defined
- EU AI Act penalties (Article 99)
- Reporting serious incidents (Article 73)
- The EU AI Act in Spain
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →