Most Articles of the AI Act say what an organisation must do. Articles 79 to 82 say what happens when the regulator thinks something has gone wrong. Article 79 sets out the procedure at national level for AI systems presenting a risk. Article 80 deals with a system the provider classified as non-high-risk when it is high-risk. Article 81 governs what happens when another Member State or the Commission disagrees with a national measure. Article 82 covers the hardest case: a high-risk system that complies with the Regulation and still presents a risk.
Suppose it is spring 2028. A Dutch municipality has recently started using a system that gives every application for social assistance a score for the likelihood that a review is needed. The supplier placed the system on the market in January 2028; a caseworker decides. An advocacy group lodges a complaint with the market surveillance authority: applicants with a part-time job and irregular hours receive a high score strikingly often. Two weeks later the authority asks the municipality for information. The team lead wants to know what happens next, what the municipality has to do itself and what the supplier is facing. The example is fictitious.
In short
An AI system presents a risk if it is a "product presenting a risk" within the meaning of Article 3, point 19, of Regulation (EU) 2019/1020, in so far as it presents risks to the health or safety, or to fundamental rights, of persons (Article 79(1)). Where a market surveillance authority has sufficient reason to consider that this is the case, it evaluates whether the system complies with all the requirements and obligations of the Regulation, giving particular attention to vulnerable groups (paragraph 2). If it does not, the authority requires the relevant operator to bring the system into compliance, withdraw it or recall it, in any event within the shorter of 15 working days or the period in sectoral legislation. If that does not happen, the authority prohibits or restricts the system on its market and notifies the Commission and the other Member States (paragraph 5). Article 80 lets the authority test a classification as non-high-risk under Article 6(3), with fines if the provider meant to circumvent the requirements (paragraph 7). Article 81 gives the Commission the final say when a Member State objects. Article 82 lets the authority act even against a high-risk system that complies with the Regulation but still presents a risk. The four Articles sit in Chapter IX and have applied since 2 August 2026; the high-risk requirements the authority tests against apply to Annex III systems from 2 December 2027. The Digital Omnibus on AI left Articles 79 to 82 untouched. The Netherlands has not yet designated its market surveillance authorities by law.
This piece is about the procedure on the regulator's side, and what it means for the provider and the deployer. What the provider must do itself when it discovers a defect is covered in corrective actions under Article 20. When a deployer must suspend use is set out in the obligations under Article 26. Reporting a serious incident is a separate route: see Article 73. Which fines exist and whether a municipality can be fined is covered in fines under the AI Act.
Where these Articles sit
Articles 79 to 82 sit in Chapter IX, Section 3, "Enforcement". The Section opens with Article 74: Regulation (EU) 2019/1020 on market surveillance and compliance of products applies to AI systems, with references to an economic operator read as references to all operators and references to a product read as references to all AI systems within the scope of the AI Act (paragraph 1). An operator is a provider, product manufacturer, deployer, authorised representative, importer or distributor (Article 3, point 8). A municipality using a system is therefore an operator too.
According to recital 156, the market surveillance system of Regulation (EU) 2019/1020 applies in its entirety, and authorities may take measures in relation to all AI systems that present a risk, even though most AI systems are not subject to specific requirements. Recital 157 says which systems the procedure is meant for: high-risk AI systems presenting a risk, prohibited systems placed on the market, put into service or used in violation of Article 5, and AI systems made available in violation of the transparency requirements that present a risk. After that come Article 83, on formal non-compliance such as a missing CE marking, declaration of conformity, registration or technical documentation, and Article 84 on Union AI testing support structures.
Article 74(4) contains an exception. For AI systems related to products covered by the harmonisation legislation in Annex I, Section A, where that legislation already provides for procedures ensuring an equivalent level of protection and having the same objective, the sectoral procedures apply instead of Articles 79 to 83.
When does an AI system present a risk?
Article 79(1) borrows the definition from Regulation (EU) 2019/1020. A product presenting a risk is a product with the potential to affect adversely, among other things, the health and safety of persons, the protection of consumers and other public interests protected by the harmonisation legislation, "to a degree which goes beyond that considered reasonable and acceptable" in relation to its intended purpose or under the normal or reasonably foreseeable conditions of use (Article 3, point 19, of Regulation (EU) 2019/1020). The AI Act narrows this to risks to the health or safety, or to fundamental rights, of persons.
Two things follow. Not every defect is a risk in this sense: the threshold is what is reasonable and acceptable for the intended purpose. And a risk to fundamental rights counts as much as a risk to safety. A system that structurally disadvantages a group of applicants can therefore present a risk without anyone being in physical danger.
The same notion of "a risk within the meaning of Article 79(1)" recurs elsewhere. The provider of a high-risk system that becomes aware of such a risk immediately investigates the causes and informs the market surveillance authority (Article 20(2)). A deployer with reason to consider that use in accordance with the instructions may result in such a risk informs the provider or distributor and the market surveillance authority without undue delay, and suspends use (Article 26(5)). Importers and distributors also have duties when this risk arises; see Articles 23 and 24.
Article 79: the procedure at national level
The trigger. A market surveillance authority has "sufficient reason" to consider that an AI system presents a risk (Article 79(2)). That reason can come from its own investigation, from a notification by a provider or deployer, or from a complaint. Any natural or legal person with grounds to consider that the Regulation has been infringed may submit a complaint to the relevant market surveillance authority, which takes it into account for its market surveillance activities (Article 85). Regulation (EU) 2019/1020 lists complaints by consumers and other information that may indicate non-compliance as a factor when an authority decides which checks to carry out, and requires authorities to establish procedures for handling complaints or reports on issues relating to risks or non-compliance (Article 11(3)(e) and (7)(a) of that Regulation). In our reading a complaint does not give the complainant a right to a particular outcome; it is, however, a common trigger for an evaluation.
The evaluation. The authority evaluates the system for compliance with "all the requirements and obligations" of the Regulation, giving particular attention to AI systems presenting a risk to vulnerable groups (paragraph 2, first subparagraph). Where it identifies risks to fundamental rights, it informs and fully cooperates with the national public authorities or bodies referred to in Article 77(1). Those are the bodies that supervise or enforce Union law protecting fundamental rights, including the right to non-discrimination. The relevant operators cooperate as necessary with the authority and with those bodies.
The corrective action. If the authority finds that the system does not comply with the Regulation, it requires the relevant operator without undue delay to take all appropriate corrective actions to bring the system into compliance, to withdraw it or to recall it. It may prescribe a period, and the action must in any event be taken "within the shorter of 15 working days" or the period provided for in the relevant Union harmonisation legislation (paragraph 2, second subparagraph). It informs the notified body, if there is one (third subparagraph). The operator ensures that the corrective action covers all the AI systems concerned that it has made available on the Union market (paragraph 4), so also those at other customers and in other Member States. Where the non-compliance is not restricted to its national territory, the authority informs the Commission and the other Member States (paragraph 3).
The operator's rights. Article 18 of Regulation (EU) 2019/1020 applies to the measures under paragraph 2 (paragraph 2, third subparagraph). Every measure states the exact grounds on which it is based, is communicated without delay together with the remedies and time limits under national law, and before the measure the operator is given the opportunity to be heard within an appropriate period of not less than 10 working days, unless urgency makes that impossible; in that case the hearing follows as soon as possible afterwards (Article 18 of Regulation (EU) 2019/1020).
If the operator does not correct. If the operator does not take adequate corrective action within the period, the authority takes all appropriate provisional measures to prohibit or restrict the system's being made available on its national market or put into service, or to withdraw or recall it. It notifies the Commission and the other Member States without undue delay (paragraph 5). The notification includes all available details, the arguments of the operator, and the reason for the non-compliance: a prohibited practice (Article 5), a high-risk system failing to meet the requirements of Chapter III, Section 2, shortcomings in harmonised standards or common specifications, or non-compliance with Article 50 (paragraph 6).
The other Member States. Other authorities report their own measures and information, and their objections if they disagree (paragraph 7). If no objection is raised by another Member State or the Commission within three months, the provisional measure is deemed justified; for a prohibited practice that period is 30 days (paragraph 8). All authorities then ensure that appropriate restrictive measures are taken on their market without undue delay (paragraph 9). In practice, a national measure thus becomes a European one.
| Step | Who | What | Provision |
|---|---|---|---|
| Trigger | Authority | Sufficient reason to consider a risk, for example after a complaint | Article 79(2), Article 85 |
| Evaluation | Authority, fundamental rights bodies, operators | Check against all requirements; attention to vulnerable groups | Article 79(2), first subparagraph |
| Corrective action | Operator | Bring into compliance, withdraw or recall, within 15 working days at most | Article 79(2) and (4) |
| Provisional measure | Authority | Prohibition, restriction, withdrawal or recall; notification to Commission and Member States | Article 79(5) and (6) |
| Response of Member States | Other authorities, Commission | Objection within three months (30 days for Article 5), otherwise deemed justified | Article 79(7) to (9) |
| On objection | Commission | Evaluation and decision within six months (60 days for Article 5) | Article 81 |
Article 80: wrongly classified as non-high-risk
A system listed in Annex III is in principle high-risk (Article 6(2)), but not where it does not pose a significant risk of harm to health, safety or fundamental rights, for example because it only performs a narrow procedural task or a preparatory task (Article 6(3), with four conditions in points (a) to (d)). Where the system performs profiling of natural persons, it is always high-risk (final subparagraph of paragraph 3). A provider that considers its Annex III system not to be high-risk documents that assessment before the system is placed on the market, registers it in the EU database (Article 49(2)) and provides the documentation to the competent authorities on request (Article 6(4)).
Article 80 is the check on that self-classification. Where an authority has sufficient reason to consider that a system so classified is in fact high-risk, it evaluates the classification against the conditions of Article 6(3) and the Commission guidelines (Article 80(1)). Under Article 6(5) the Commission was to provide those guidelines by 2 February 2026. On 19 November 2026 they were still listed as a draft on the Commission's page. What the draft guidelines say is covered in the Article 6(3) filter.
If the system is high-risk after all, the authority requires the provider to bring it into compliance with the Regulation and to take appropriate corrective action within a period it may prescribe (paragraph 2). If the provider does not do so within that period, it is subject to fines in accordance with Article 99 (paragraph 4), and Article 79(5) to (9) applies (paragraph 6), meaning the provisional measures and the European coordination. Where the authority establishes that the provider misclassified the system as non-high-risk "in order to circumvent the application of requirements in Chapter III, Section 2", the provider is also subject to fines for that (paragraph 7). In its checks, the authority may take into account in particular the information stored in the EU database referred to in Article 71 (paragraph 8).
In our reading, the difference between paragraph 4 and paragraph 7 is this: paragraph 4 penalises failing to remedy after the finding, paragraph 7 the intent behind the classification itself. A provider that documents its assessment carefully and honestly runs the risk of paragraphs 2 and 4 if it gets it wrong, but in our reading makes it hard to assume that it meant to circumvent the requirements (paragraph 7). That makes the documentation under Article 6(4) the most important document in such a procedure.
Article 81: the Union safeguard procedure
Where an authority of another Member State objects to a national measure within three months of the notification under Article 79(5), or within 30 days for a prohibited practice, or where the Commission considers the measure contrary to Union law, the Commission enters into consultation with the authority and the operator and evaluates the measure (Article 81(1)). It decides within six months of the notification, or within 60 days for a prohibited practice, whether the measure is justified. If it is, all Member States take appropriate restrictive measures; if it is not, the Member State withdraws the measure (paragraph 2).
Where the national measure is considered justified and the non-compliance is attributed to shortcomings in a harmonised standard or common specification, the Commission applies the procedure of Article 11 of Regulation (EU) No 1025/2012 (paragraph 3). That is the procedure for formal objections to a standard. An individual case can thus end up affecting the standard itself. What a standard does and does not prove is covered in Articles 40 to 42.
Article 82: compliant, and still a risk
Article 82 goes further than the rest. Where an authority, having performed an evaluation under Article 79 and after consulting the fundamental rights body referred to in Article 77(1), finds that "although a high-risk AI system complies with this Regulation", it nevertheless presents a risk to the health or safety of persons, to fundamental rights, or to other aspects of public interest protection, it requires the operator to take all appropriate measures to remove that risk within a period it may prescribe (Article 82(1)). The provider or other operator takes those measures for all the systems concerned on the Union market (paragraph 2). The Member State immediately informs the Commission and the other Member States (paragraph 3); the Commission consults, evaluates and decides whether the measure is justified, and proposes other measures where necessary (paragraphs 4 and 5).
Three things stand out. Article 82 applies only to high-risk systems. It requires a prior evaluation under Article 79, so the system must first have come into view as a risk. And it means compliance with the requirements is not a free pass: a system that meets every requirement can still be subjected to measures. That is also why the presumption of conformity from a harmonised standard is not the end of the story.
What the Digital Omnibus on AI changed
Nothing in Articles 79 to 82. Article 1 of Regulation (EU) 2026/1744 does not mention these Articles, and in the consolidated text of 27 July 2026 they appear as original text. Two amendments in the same Section do have an effect.
- The fundamental rights bodies. Article 77 received a new title and a new paragraph 1, and paragraphs 1a and 1b were inserted (Article 1, point 34, of Regulation (EU) 2026/1744). Those bodies may request and access information and documentation electronically and in a machine-readable format from the market surveillance authority, which grants them access and where necessary requests the documents from the provider or deployer without delay; both cooperate closely. Because Article 79(2) and Article 82(1) refer to these bodies, in our reading their role in the procedure becomes firmer.
- The AI Office. The AI Office has exclusive competence for supervision and enforcement for AI systems based on a general-purpose AI model of the same provider or undertaking, with exceptions including Annex I, critical infrastructure and systems of financial institutions falling under Article 74(6), and for systems that constitute or are integrated into a very large online platform or very large online search engine (Article 75(1), as replaced by Article 1, point 31). For those systems the AI Office has all the powers of a market surveillance authority provided for in this Section, and it may also act following a complaint under Article 85 (Article 75a(1) and (2), inserted by Article 1, point 32). More in the AI Office.
Who is the market surveillance authority in the Netherlands?
That has not yet been laid down in law. On 20 April 2026 the Dutch government proposed a structure in which existing supervisors supervise AI within their own domain, with the Dutch Data Protection Authority (AP) where there is no clear supervisor, and with the AP and the Dutch Authority for Digital Infrastructure (RDI) as coordinators (Government of the Netherlands, 20 April 2026, in Dutch). The draft bill for the Dutch AI Act Implementation Act (Uitvoeringswet AI-verordening) was in public consultation until 1 June 2026. According to the Planningsbrief Digitale Zaken, submission to the House of Representatives is planned for the second quarter of 2027 (Parliamentary Paper 26643, no. 1554, in Dutch).
What the draft provides is relevant to this procedure:
- Designation. For Chapter IX, Sections 1 to 4, and therefore also for Articles 79 to 82 and complaints under Article 85, the draft designates a market surveillance authority per area of Annex III, for high-risk AI systems referred to in Article 6(2): the AP for, among others, point 5(a) (public assistance benefits and services), and the Dutch Authority for the Financial Markets (AFM), De Nederlandsche Bank (DNB) and the AP for points 5(b) and (c) (creditworthiness and life and health insurance), with a ministerial regulation limiting their designation to categories of operators to be specified (Article 2.2(1)(c) and (2) of the draft).
- Measures. The authority is empowered to take the measures under Article 79(2), (5) and (9), Article 80(2), Article 81(2), Article 82(1) and Article 83(2), and whoever is subject to such a measure must comply with it (Article 3.5 of the draft); an order subject to administrative enforcement is added (Article 3.6).
- Fines. Among other things, the draft makes Article 79(2), first subparagraph, last sentence (the duty to cooperate, which the Dutch language version addresses to deployers), Article 79(4), Article 80(5) and Article 82(2) subject to fines up to the maximum of Article 99(4) (Article 3.7(2)(b) of the draft).
That is a draft and not yet law. Until the Implementation Act applies, the Netherlands has not designated market surveillance authorities by law, and the national powers and rules on fines are missing. Only for high-risk systems of financial institutions and for products covered by Annex I does the AI Act itself designate an authority (Article 74(3) and (6)). In our reading, a national procedure under Articles 79 to 82 can therefore for now mainly be started by those authorities, alongside the AI Office where it is competent.
Main example: the municipality and the complaint about the social assistance system
The roles. The system helps assess whether someone is eligible for social assistance, an essential public assistance benefit; that falls under Annex III, point 5(a). It gives each applicant a score, which in our reading is profiling; the system is then always high-risk (final subparagraph of Article 6(3)). The supplier is the provider, the municipality the deployer. Because the supplier placed the system on the market in January 2028, after 2 December 2027, the high-risk requirements apply in full and the transitional rule in Article 111(2) does not come into play. Had the system already been running before that date, the requirements would only apply in the event of a significant change in design, and by 2 August 2030 at the latest.
What the authority does. In our reading, the complaint gives sufficient reason for an evaluation under Article 79(2): applicants with irregular income are a group that deserves particular attention, and a systematically higher score can be a risk to the right to non-discrimination. The authority checks the system against all requirements, such as data (Article 10), accuracy (Article 15) and human oversight (Article 14), and involves the fundamental rights body. It can request documentation and datasets from the supplier (Article 74(12)), and asks the municipality, which must cooperate, for the logs and work instructions. If it finds a shortcoming, it requires the supplier to correct the system within 15 working days at most, after first hearing it. That correction covers all municipalities using the system (paragraph 4).
What the municipality must do itself. In our reading, these four things:
- Cooperate. The municipality cooperates as necessary with the authority and the fundamental rights body (Article 79(2), first subparagraph). It provides the logs it keeps (Article 26(6)), the instructions for caseworkers and figures on how often caseworkers depart from the score. Under the draft Dutch Implementation Act, refusing is subject to fines; whether a public authority can be fined is for each Member State to decide (Article 99(8)).
- Assess whether to suspend use. If the municipality has reason to consider that use in accordance with the instructions may result in a risk within the meaning of Article 79(1), it informs the supplier and the authority and suspends use (Article 26(5)). In our reading, a substantiated complaint with figures is such a reason as soon as the municipality recognises those figures itself; it does not wait for the outcome of the evaluation.
- Look at decisions already taken. The AI Act does not say what must happen to earlier decisions; that is administrative law. In our reading, the municipality checks which decisions followed a high score and whether they need to be reviewed. An applicant can also ask for an explanation of the role the system played in their decision; see the right to explanation under Article 86.
- Do not forget the GDPR route. The AI Act complaint runs alongside the GDPR. An applicant who considers that the processing of their personal data infringes the GDPR can lodge a complaint with the AP as supervisory authority (Article 77 GDPR), and Article 85 AI Act applies "without prejudice to other administrative or judicial remedies". In our reading, the DPIA and the fundamental rights impact assessment (Article 27) are the first documents an authority will want to see in such a procedure; see the FRIA for municipalities.
If the system complies and still presents a risk. Suppose the supplier shows that the system stays within its declared accuracy and meets every requirement, but that precisely this accuracy leads to many unnecessary reviews for this group. The authority can then, after consulting the fundamental rights body, require measures under Article 82(1) to remove that risk, for example an adjustment of the model or of the instructions for use. The Commission then evaluates that measure (paragraph 4).
In the contract. Because the measures land with the supplier but the use sits with the municipality, in our reading it is sensible to agree in the contract that the supplier informs the municipality without delay of an evaluation, a required corrective action or a provisional measure, and cooperates in reviewing earlier outputs. See AI procurement contracts.
Example 2: a bank and a document check that would not be high-risk
Suppose it is 2028, so after 2 December 2027, the date from which Article 6(3) applies to Annex III. A bank uses a system from an external supplier for consumer credit applications that checks whether the documents submitted are complete and legible. The supplier classified the system under Article 6(3) as non-high-risk, because it would perform a narrow procedural task, and registered it in the EU database. In practice the system gives each application a label "needs attention", and the credit department rejects applications with that label more often. A consumer organisation lodges a complaint. This example is fictitious too.
The authority. For high-risk systems used by financial institutions, the market surveillance authority is the national authority responsible for the financial supervision of that institution, in so far as the use is directly connected with the provision of financial services (Article 74(6)); a Member State may designate another authority in justified circumstances (paragraph 7). Recital 158 explains this. In the Dutch draft, the AFM, DNB and the AP are designated for Annex III, point 5(b). In our reading the AI Office plays no role here: its exclusive competence under Article 75(1) covers systems based on a general-purpose AI model of the same provider or undertaking and very large online platforms and search engines, and neither applies in this example. Even for a system based on such a model, Article 75(1)(a), point (iii), excludes AI systems provided by financial institutions, insofar as they fall under Article 74(6).
The procedure. In our reading, the authority has sufficient reason for an evaluation under Article 80(1): a label that influences the outcome of the credit decision is hard to reconcile with a narrow procedural task, and if the system ranks applicants by their personal characteristics, that is profiling and always high-risk. Creditworthiness falls under Annex III, point 5(b). The authority compares the documentation under Article 6(4) with the registration in the EU database (Article 80(8)) and with how the system is actually used. If the system is high-risk, the supplier must make it meet all requirements within the prescribed period (paragraph 2), with the fines of paragraph 4 if that fails. If the classification turns out to have been meant to circumvent the requirements, fines under paragraph 7 follow.
What the bank does. The procedure targets the supplier, but the bank bears the consequences. If the system becomes high-risk, the bank as deployer has the obligations of Article 26 and carries out a fundamental rights impact assessment (Article 27(1)). In our reading, for every system offered to it as non-high-risk under Article 6(3), the bank asks for the documentation under Article 6(4) and tests it against its own use. If the bank uses the label differently from what the supplier intended, it may moreover become a provider itself if it modifies the intended purpose in such a way that the system becomes a high-risk AI system (Article 25(1)(c)). See also AI in credit decisions.
From when?
Articles 79 to 82 sit in Chapter IX. They are not among the exceptions in the third paragraph of Article 113 (from Chapter IX only Article 78 is listed there, which has applied since 2 August 2025), and have therefore applied since 2 August 2026, the general date in the second paragraph. That does not yet say what an authority can test against:
- Prohibited practices and Article 50. Article 5 has applied since 2 February 2025, except for the prohibitions in Article 5(1), first subparagraph, points (ba) and (bb), and paragraphs 1a and 1b, which apply from 2 December 2026 (Article 113, third paragraph, point (a), as replaced by the Omnibus). Article 50 has applied since 2 August 2026; providers of systems generating synthetic content that were placed on the market before 2 August 2026 have until 2 December 2026 for Article 50(2) (Article 111(4)). For those provisions the Article 79 procedure can already run, with the shorter periods of 30 days and 60 days for a prohibited practice.
- High-risk systems. The requirements of Chapter III, Sections 1 to 3, apply from 2 December 2027 to systems classified via Article 6(2) and Annex III, and from 2 August 2028 to systems classified via Article 6(1) and Annex I (Article 113, third paragraph, point (c), as replaced by Article 1, point 40, of Regulation (EU) 2026/1744). Only from those dates can an evaluation under Article 79 test against those requirements, can Article 80 test a classification under Article 6(3), and can Article 82 come into play. For systems already on the market or in service before that date, Article 111(2) applies.
- In the Netherlands. Until the Implementation Act applies, the Netherlands has not designated market surveillance authorities by law, and the national powers and rules on fines are missing. Only for high-risk systems of financial institutions and for products covered by Annex I does the AI Act itself designate an authority (Article 74(3) and (6)); the AI Office is also competent for the systems in Article 75(1).
What to do now
As a deployer, such as a municipality:
- Record when something is a risk within the meaning of Article 79(1), who decides that and who then informs the provider and the authority (Article 26(5)).
- Keep ready the documents an authority will ask for: logs, work instructions, figures on departures from the output, the DPIA and the fundamental rights impact assessment.
- Agree in the contract that the supplier informs you of evaluations and measures and cooperates in a review.
- Set up a route for complaints and questions from residents, including requests for explanation under Article 86, so that a signal reaches you before it reaches the authority.
As a provider:
- Document a classification as non-high-risk carefully (Article 6(4)) and keep the EU database registration in line with actual use; that is your defence under Article 80.
- Make sure you can correct within 15 working days, for all customers in the Union (Article 79(2) and (4)). Know which version runs where.
- Decide how you will use a hearing (Article 18 of Regulation (EU) 2019/1020): who writes the response and with which documents.
- Do not assume compliance is the end. Article 82 allows measures against a compliant system that still presents a risk; post-market monitoring (Article 72) is the way to see that before the authority does.
Want to know which AI Act and GDPR obligations come into view for your use case and what your role is? With the case check you describe the use case in a few questions.
Frequently asked questions about AI systems presenting a risk
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe