Skip to main content
Praxikon

When a high-risk AI system does not comply: corrective actions, documentation keeping and cooperation under Articles 18, 20 and 21 of the AI Act

··35 min read

Much of what is written about the AI Act treats it as a gate: a high-risk AI system goes through the conformity assessment, gets a declaration and a CE marking, and may then go to market. But a system that has passed the gate can still turn out not to comply later. An update changes its behaviour, practice differs from the test data, or monitoring reveals an error nobody foresaw. For that moment the Regulation has three short articles for the provider: Article 18 on documentation keeping, Article 20 on corrective actions and the duty to inform others, and Article 21 on cooperation with the authorities.

Suppose a Dutch municipality uses a system that helps caseworkers assess applications for special assistance benefits. One Monday morning a letter arrives from the supplier: "We have found that version 4.2 misjudges the income of applicants with fluctuating income. We have disabled the income module and are rolling the system back to version 4.1. We have informed the supervisor." The team lead asks: "May we still use the system? Do we have to report anything ourselves? And what do we do about the decisions of the past few weeks?" The example is fictitious.

In short

For ten years after a high-risk AI system has been placed on the market or put into service, the provider keeps five items at the disposal of the national competent authorities: the technical documentation, the documentation of the quality management system, the changes approved by notified bodies and their decisions (where applicable) and the EU declaration of conformity (Article 18(1)). Member States decide what happens on bankruptcy or cessation of activity (paragraph 2); financial institutions keep the technical documentation as part of their documentation under financial services law (paragraph 3). Where the provider considers or has reason to consider that its system does not comply with the Regulation, it immediately takes the necessary corrective actions: bring the system into conformity, withdraw it, disable it or recall it. It informs the distributors and, where applicable, the deployers, the authorised representative and the importers (Article 20(1)). Where the system presents a risk within the meaning of Article 79(1), to health, safety or fundamental rights, it immediately investigates the causes and informs the market surveillance authority and, where applicable, the notified body (paragraph 2). Upon a reasoned request it gives the competent authority all information and documentation needed to demonstrate conformity, and access to the logs under its control; the authority treats these confidentially under Article 78 (Article 21). A deployer that has reason to consider that the system presents a risk informs the provider or distributor and the market surveillance authority and suspends use (Article 26(5)). The Digital Omnibus (Regulation (EU) 2026/1744) did not amend Articles 18 to 21. The obligations apply to Annex III systems from 2 December 2027 and to systems linked to products under Section A of Annex I from 2 August 2028.

This piece is about what the provider must do when things go wrong, and what that means for the deployer. How the provider detects errors is covered in post-market monitoring under Article 72. How a serious incident is reported is covered in reporting serious incidents under Article 73. The deployer's obligations are set out in deployer obligations under Article 26.

Where these articles sit

Articles 18 to 21 are in Section 3 of Chapter III, the section with the obligations of providers, deployers and other parties. They elaborate the list in Article 16. The provider of a high-risk AI system keeps the documentation referred to in Article 18 (point (d)) and, when under its control, the logs referred to in Article 19 (point (e)), takes the necessary corrective actions and provides the information required in Article 20 (point (j)), and, upon a reasoned request of a national competent authority, demonstrates that the system complies with the requirements of Section 2 (point (k)).

All four articles are addressed to the provider. The authorised representative, the importer and the distributor have their own, comparable obligations in Article 22, 23 and 24; see importers and distributors under Articles 23 and 24. The deployer has its own reporting and log-keeping obligations in Article 26(5) and (6). Anyone who puts their name or trademark on a high-risk AI system already placed on the market or put into service, substantially modifies the system or changes its intended purpose so that it becomes high-risk becomes a provider and takes on these obligations (Article 25(1)).

Article 18: ten years of documentation keeping

Five items. For a period ending ten years after the system has been placed on the market or put into service, the provider keeps the following at the disposal of the national competent authorities (paragraph 1):

  1. the technical documentation referred to in Article 11 (point (a));
  2. the documentation concerning the quality management system referred to in Article 17 (point (b));
  3. the documentation concerning the changes approved by notified bodies, where applicable (point (c));
  4. the decisions and other documents issued by the notified bodies, where applicable (point (d));
  5. the EU declaration of conformity referred to in Article 47 (point (e)).

"Keep at the disposal" is more than storing somewhere. In our reading it means the provider can produce the documents within a reasonable time when an authority asks, in a form the authority can read. That ties in with Article 21(1), discussed below. What goes into the technical documentation is covered in technical documentation under Article 11 and Annex IV; the quality management system in the quality management system under Article 17; the declaration in the EU declaration of conformity and the CE marking.

From when does the period run, and for which version? In EU product law the Commission counts per individual product: "placing on the market" refers to each individual product, not to a type of product, and the retention period for the technical documentation is calculated from the moment the individual product covered by that documentation is placed on the market (Blue Guide 2022, sections 2.3 and 4.3). The AI Act defines placing on the market in the same way, as the first making available in the Union (Article 3, point (9)). The technical documentation must also be kept up to date (Article 11(1)). How that works for software that is updated often or delivered as a service, Article 18 does not say. In our reading the provider keeps each version of the documentation that belongs to a version of the system placed on the market, and the ten-year period for that documentation runs from the last supply of that version. Otherwise, when asked about a decision taken three years ago, it cannot show how the system worked at the time. A substantial modification triggers a new conformity assessment (Article 43(4)) and a new declaration, and in our reading also a new start of the period for that version.

Bankruptcy or cessation. Each Member State determines the conditions under which the documentation remains at the disposal of the national competent authorities for the full ten years where a provider or its authorised representative established on its territory goes bankrupt or ceases its activity before the end of that period (paragraph 2). The Netherlands has not yet laid this down in law. The draft bill for the Dutch AI Act Implementation Act (Uitvoeringswet AI-verordening), in public consultation from 20 April to 1 June 2026 (in Dutch), provides in Article 7.1 that the provider ensures that the documentation referred to in Article 18(1) remains at the disposal of the national competent authorities for the whole period, also if it is declared bankrupt or ceases its activities before the end of that period. The draft contains no separate rule for the authorised representative, because the retention duty of Article 18(1) applies to providers only (draft explanatory memorandum, section 4.9.2). It is a draft and not yet law: submission to the House of Representatives is planned for the second quarter of 2027 (Parliamentary Paper 26643, no. 1554, in Dutch). In our reading a duty resting on a bankrupt provider gives a deployer little to rely on. That is a reason to agree in the contract on access to the documentation if the supplier disappears, for example through an escrow arrangement.

Financial institutions. Providers that are financial institutions subject to requirements regarding their internal governance under Union financial services law maintain the technical documentation as part of the documentation they already keep under that law (paragraph 3). The paragraph says where the documentation is kept. In our reading it leaves the ten-year period of paragraph 1 untouched.

Who else keeps records? The Regulation does not leave record keeping to the provider alone.

PartyWhatHow longProvision
ProviderTechnical documentation, quality management system, approvals and decisions of notified bodies, EU declaration of conformityTen years after placing on the market or putting into serviceArticle 18(1)
ProviderAutomatically generated logs, to the extent under its controlAppropriate to the intended purpose, at least six monthsArticle 19(1)
Authorised representativeProvider contact details, copy of the declaration, the technical documentation and, where applicable, the certificateTen yearsArticle 22(3)(b)
ImporterCopy of the certificate, the instructions for use and the declarationTen yearsArticle 23(5)
DeployerAutomatically generated logs, to the extent under its controlAppropriate to the intended purpose, at least six monthsArticle 26(6)

Article 19 in brief: the logs

Besides the documentation, the provider keeps the logs automatically generated by its system (Article 12(1)), to the extent such logs are under its control, for a period appropriate to the intended purpose and of at least six months, unless Union or national law, in particular on the protection of personal data, provides otherwise (Article 19(1)). Financial institutions keep the logs as part of their documentation under financial services law (paragraph 2). How that period relates to the GDPR is covered in logs for high-risk AI. For this piece one thing matters: without logs a provider cannot investigate the cause of an error, and cannot give an authority access under Article 21(2).

Article 20: corrective actions and the duty of information

When. Article 20(1) starts with a provider that "consider[s] or ha[s] reason to consider" that a system it has placed on the market or put into service is not in conformity with the Regulation. It covers the whole Regulation, not only the requirements of Section 2. An established infringement is not needed; reason to consider it is enough. Such reasons can come from the provider's own monitoring (Article 72), from a deployer's report, from complaints or from a question by an authority.

What. The provider then immediately takes the necessary corrective actions, as appropriate:

  • to bring the system into conformity, for example with a fixed version or amended instructions for use;
  • to withdraw it: any measure aiming to prevent a system in the supply chain from being made available on the market (Article 3, point (17)), so no new supplies;
  • to disable it, for example by switching off a function remotely;
  • to recall it: any measure aiming to achieve the return to the provider, taking out of service or disabling the use of a system already made available to deployers (Article 3, point (16)).

Which action is needed, the Regulation does not say; "as appropriate" leaves that to the provider. In our reading the choice depends on severity: an error in the instructions for use calls for something different than a model that systematically gives wrong results for a group of people.

Whom it informs. The provider informs the distributors of the system and, where applicable, the deployers, the authorised representative and the importers (paragraph 1). The word "immediately" in paragraph 1 belongs to the corrective actions. For the notice, paragraph 1 sets no deadline of its own, and on content it only says that the provider informs the others "accordingly". In our reading its purpose implies that the recipient must be able to understand which version is concerned, what is wrong, which action was taken and what it must do itself. A deployer that only hears that "an update is coming" cannot assess its own obligation under Article 26(5).

Where there is a risk: investigate and inform the supervisor. Where the system presents a risk within the meaning of Article 79(1) and the provider becomes aware of it, the provider immediately investigates the causes, in collaboration with the reporting deployer where applicable. It informs the market surveillance authorities competent for the system and, where applicable, the notified body that issued a certificate in accordance with Article 44, in particular of the nature of the non-compliance and of any relevant corrective action taken (paragraph 2).

An "AI system presenting a risk" is a "product presenting a risk" as defined in 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)). For a system that supports decisions on benefits, credit or job applicants, in practice the concern is mainly fundamental rights, such as equal treatment and non-discrimination.

How this differs from Article 73 and Article 79. Article 20 is the provider's own route. Two other routes exist alongside it, and they often run at the same time.

Article 20(1)Article 20(2)Article 73Article 79(2)
TriggerReason to consider the system does not complyThe system presents a risk (Article 79(1))A serious incident (Article 3, point (49))The market surveillance authority has sufficient reason to consider the system presents a risk
Who actsProviderProviderProviderMarket surveillance authority
WhatTake corrective action, inform the supply chainInvestigate causes, inform market surveillance authority and notified bodyReport to the market surveillance authority, investigate, take corrective actionEvaluation; on non-compliance, require the operator to take corrective action
DeadlineAction immediately; no deadline of its own for the noticeImmediatelyImmediately and no later than 15 days; two days for a widespread infringement or critical infrastructure, ten days in the event of deathSet by the authority, in any event within 15 working days

If the system leads to a serious incident, the provider reports it within the deadlines of Article 73; a serious incident includes an infringement of obligations under Union law intended to protect fundamental rights (Article 3, point (49)(c)). Note the second subparagraph of Article 73(6) in that case: the provider shall not perform any investigation which involves altering the AI system concerned in a way which may affect any subsequent evaluation of the causes of the incident, prior to informing the competent authorities of such action. In our reading a quick rollback to an earlier version can be such an alteration, and the provider then informs the authorities beforehand. If the market surveillance authority acts itself, it can require the operator to take corrective action (Article 79(2)) and, if that does not happen, prohibit or restrict the making available or putting into service of the system on its national market, or withdraw or recall it (paragraph 5). And a system that does comply but still presents a risk can also lead to an order to take measures (Article 82(1)). In our reading a provider that acts in time under Article 20 keeps more control over the measure than one that waits for the supervisor to impose it.

Article 21: cooperation with the authorities

Information and documentation on request. Upon a reasoned request by a competent authority, the provider supplies all information and documentation necessary to demonstrate that the system complies with the requirements of Section 2, in a language easily understood by the authority, in one of the official languages of the institutions of the Union as indicated by the Member State concerned (paragraph 1). The documentation the provider keeps for ten years under Article 18 is exactly what it must then be able to produce.

Access to the logs. Upon a reasoned request the provider also gives the requesting authority, as applicable, access to the automatically generated logs referred to in Article 12(1), to the extent they are under its control (paragraph 2). With a system delivered as a service, the logs are often held by the provider. With a system running in the deployer's own environment, they may sit with the deployer. That is a reason to agree in the contract who keeps which logs.

Confidentiality. The authority treats what it obtains in accordance with the confidentiality obligations of Article 78 (paragraph 3). Article 78(1) protects, among other things, intellectual property rights, confidential business information and trade secrets, including source code (access to the source code itself is subject to the conditions of Article 74(13), see below), and the effective implementation of inspections, investigations and audits. The authorities request only data strictly necessary to assess the risk and exercise their powers, protect that data and delete it as soon as it is no longer needed (Article 78(2)). In our reading a provider therefore cannot answer a request by invoking trade secrets: Article 21 has no such exception, and the protection lies in how the authority handles the data.

More powers in Article 74. Article 21 is not the supervisor's only tool. Market surveillance authorities are granted, where relevant and limited to what is necessary for their tasks, full access to the documentation and to the training, validation and testing data sets, including remotely (Article 74(12)). They get access to the source code only upon a reasoned request, where that is necessary to assess conformity with Section 2 and testing or auditing procedures and verifications based on the data and documentation provided by the provider have been exhausted or proved insufficient (paragraph 13).

Which authority? A national competent authority is a notifying authority or a market surveillance authority (Article 3, point (48)). Since the Digital Omnibus, the Commission's AI Office has exclusive competence for supervision and enforcement for AI systems based on a general-purpose AI model where the model and the system are developed by the same provider or the same undertaking, and for AI systems that constitute or are integrated into a very large online platform or very large online search engine. The first group has exceptions, among others for systems related to products under Annex I, for critical infrastructure (Annex III, point 2) and for systems of financial institutions covered by Article 74(6) (Article 75(1), as replaced by Article 1, point (31)(b), of Regulation (EU) 2026/1744). For a provider in that situation, the AI Office is the point of contact; a provider of a high-risk AI system subject to the AI Office's competence also reports serious incidents to the AI Office (Article 75(1a)). In the Netherlands no market surveillance authority for the AI Act has yet been designated; the supervisors are to be designated in the AI Act Implementation Act, whose submission to the House of Representatives is planned for the second quarter of 2027 (Parliamentary Paper 26643, no. 1554).

What the Digital Omnibus changed

Regulation (EU) 2026/1744, the Digital Omnibus on AI, lists in Article 1 the provisions of the AI Act it amends. Articles 18, 19, 20 and 21 are not among them, nor are Articles 78 and 79. In the consolidated text of 27 July 2026 these articles carry no amendment markers. No SME simplification was added to them: a small provider also keeps records for ten years and takes corrective action immediately.

Amended, and relevant to what the provider keeps and whom it deals with:

  • Simplified technical documentation. SMEs, including start-ups, and small mid-caps (SMCs) may provide the elements of Annex IV in a simplified manner, using a form the Commission establishes (Article 11(1), second subparagraph, as replaced by Article 1, point (10)). Article 18 makes no distinction; in our reading the ten-year period also applies to that simplified documentation.
  • Proportionate quality management. Implementation of the quality management system is proportionate to the size of the provider's organisation, in particular for SMEs, start-ups and SMCs (Article 17(2), as replaced by Article 1, point (11)). What is kept of that system under Article 18(1)(b) may therefore be smaller; the obligation to keep it remains.
  • Annex I systems. Under the new Article 2(13), the application of specific requirements or obligations in Articles 9 to 15 and 17 to 25, and so also in Articles 18 to 21, may be limited for Annex I systems, where the Union harmonisation legislation in Section A of Annex I provides an equivalent or higher level of protection and the limitation does not affect the overall level of protection. The Commission adopts delegated acts for this by 2 August 2027 (Article 1, point (3)). For Annex III systems nothing changes.
  • The AI Office as supervisor for part of the systems (Article 75, as amended by Article 1, point (31)), see above.
  • The dates of application in point (c) of the third paragraph of Article 113 (Article 1, point (40)), and the fines rule for SMCs in Article 99(6a) (Article 1, point (38)), see below.

If you come across the words "in Article 20, the following paragraph is added" in the Omnibus: that concerns Article 20 of the Machinery Regulation (EU) 2023/1230, which gets a new paragraph 10 on the presumption of conformity (Article 3, point (2), of Regulation (EU) 2026/1744). It is not an amendment of Article 20 of the AI Act.

Main example: the municipality receives a notice from its supplier

Back to the municipality. The system helps caseworkers assess whether someone is entitled to special assistance. That falls under Annex III, point 5(a): AI systems intended to be used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits and services. The supplier is a Dutch company that developed the system itself and supplies it under its own name: the provider. The municipality buys directly, without an importer or distributor, and is the deployer. To keep the transitional rule of Article 111(2) out of the picture, we assume the system was first placed on the market in 2028, so the Chapter III obligations apply.

The provider's monitoring showed that for applicants with fluctuating income, such as freelancers and flexible workers, version 4.2 overestimated income and proposed a rejection more often. The provider concluded that accuracy no longer met what it had declared in the instructions for use (Article 15) and that the system presented a risk to the fundamental rights of those applicants. It switched off the income module (disabling), rolled the system back to version 4.1 (bringing it into conformity), and informed the deployers (Article 20(1)) and the market surveillance authority (paragraph 2).

What the municipality does, in our reading:

  1. Read the notice against four questions. Which version and which period? What is wrong? Which action was taken? What must the municipality do itself? If an answer is missing, it asks straight away. Without that information it cannot assess its own obligations.
  2. Decide whether to suspend use. Where the municipality has reason to consider that use in accordance with the instructions may result in the system presenting a risk within the meaning of Article 79(1), it informs the provider or distributor and the relevant market surveillance authority without undue delay, and suspends the use of the system (Article 26(5)). A notice from the provider that the system presents a risk to fundamental rights is, in our reading, such a reason. The text speaks of the use "of that system", not of one module. In our reading the municipality suspends use until it has established that the fixed version removes the risk, and records why and when it resumes use.
  3. Report itself, even though the provider has. Article 26(5) makes no exception for the case where the provider has already informed the market surveillance authority. In our reading the municipality makes its own short report, referring to the provider's. That costs little and shows it knows its obligation.
  4. Check whether there is a serious incident. If applications were wrongly rejected because caseworkers relied on the system's output, this may amount to an infringement of obligations intended to protect fundamental rights (Article 3, point (49)(c)). Where the municipality identifies a serious incident, it immediately informs first the provider, and then the importer or distributor, if any, and the market surveillance authority (Article 26(5)). If it cannot reach the provider, Article 73 applies mutatis mutandis (Article 26(5)). How that continues is covered in reporting serious incidents.
  5. Keep the logs and cooperate. The municipality keeps the logs under its control for at least six months (Article 26(6)) and makes sure logs from the period concerned are not deleted automatically. If the municipality reported the risk itself, the provider investigates the causes in collaboration with it (Article 20(2)). Even where the provider found the problem itself, it often needs the municipality's data to see which applications were affected.
  6. Look at decisions already taken. The AI Act says nothing about decisions already taken on the system's advice; that is a question for administrative law. In addition, the GDPR requires personal data to be accurate and, where necessary, kept up to date (Article 5(1)(d) GDPR). In our reading the municipality lists the applications assessed with version 4.2 in the period concerned, so it can check which decisions may rest on a wrong assessment.
  7. Update the fundamental rights impact assessment. As a body governed by public law using an Annex III system, the municipality must carry out a fundamental rights impact assessment (Article 27). If during use it considers that any element has changed or is no longer up to date, it updates the information (Article 27(2)). A newly known risk for freelancers and flexible workers is, in our reading, such an element. See the FRIA under Article 27.
  8. Tighten human oversight. The municipality assigns human oversight to people with the necessary competence, training and authority (Article 26(2)). In our reading a notice like this is the moment to tell caseworkers what went wrong, so they know what to watch for with the fixed version.

Which market surveillance authority the municipality informs is not yet settled in the Netherlands. On 20 April 2026 the Dutch government proposed to give supervision per domain to existing supervisors, with the Dutch Data Protection Authority (AP) where there is none (Government of the Netherlands, 20 April 2026, in Dutch). In the draft bill for the AI Act Implementation Act, the AP is the market surveillance authority for, among others, systems under Annex III, point 5(a) (Article 2.2(1)(c), under 1°, of the draft, in Dutch). That is a draft and not yet law. Until the act applies, the contract is the main tool. What the municipality agrees in it, in our reading:

  1. A fixed contact point in both directions, for the provider's notices under Article 20 and for the municipality's reports under Article 26(5).
  2. The content of a notice: version, period, nature of the non-compliance, action taken, consequences for earlier outputs and what the municipality must do.
  3. A deadline. Article 20(1) requires the action to be taken immediately, but sets no deadline of its own for the notice. Agree within how many hours or working days the provider informs you.
  4. Data for remediation: which applications were affected, with the logs and outputs the municipality needs to review its own decisions.
  5. Who keeps which logs, for how long, and how an authority gets access (Article 19, Article 21(2) and Article 26(6)).
  6. Access to the documentation during the ten years of Article 18, also after the contract ends and if the supplier goes bankrupt.
  7. No silent changes. The provider reports every corrective action, including one it regards as maintenance.

More on the procurement side in AI procurement contracts. For this type of system in particular: welfare fraud risk scoring.

Example 2: the bank with its own credit model

A bank develops its own model to assess the creditworthiness of individuals and uses it in its own lending process. That is Annex III, point 5(b), with an exception for systems used to detect financial fraud. Because the bank develops the system itself and puts it into service under its own name, it is the provider (Article 3, point (3)) and at the same time the deployer. This example is fictitious too.

The special rules for financial institutions apply to the bank. It keeps the technical documentation and the logs as part of the documentation it already keeps under financial services law (Article 18(3) and Article 19(2)). For these systems the market surveillance authority is the national authority responsible for the financial supervision of the institution, in so far as the use is directly connected with the provision of financial services (Article 74(6)); a Member State may in justified circumstances designate another authority (paragraph 7). In the draft bill for the Dutch AI Act Implementation Act, the Dutch Authority for the Financial Markets (AFM), De Nederlandsche Bank (DNB) and the AP are designated as market surveillance authorities for Annex III, point 5(b) and (c), with a ministerial regulation limiting their designation to categories of operators to be specified in it (Article 2.2(1)(c), under 3°, and (2)(a), of the draft). Which of the three will send a request under Article 21 therefore only follows from that regulation. This is not yet law either.

Article 20 applies to the bank in full. If monitoring shows that the model deviates structurally for a group of customers, the bank immediately takes the necessary corrective action and, where there is a risk, informs the market surveillance authority. Because provider and deployer are the same party, informing the deployer is an internal step. In our reading the bank records it just like an external notice: the model team tells the department that takes credit decisions which version is concerned, what is wrong and which action applies. For financial institutions the monitoring obligation of Article 26(5) is deemed fulfilled by complying with the internal governance rules under financial services law (Article 26(5), second subparagraph). In our reading that rule concerns monitoring only, not the suspension of use and the reports in case of a risk or a serious incident. How the model relates to banking law is covered in AI in credit decisions.

From when?

Articles 18 to 21 are in Section 3 of Chapter III. That section applies from 2 December 2027 to AI systems classified as high-risk under Article 6(2) and Annex III, and from 2 August 2028 to AI systems classified as high-risk under Article 6(1) and Annex I (Article 113, third paragraph, point (c), as replaced by Article 1, point (40), of Regulation (EU) 2026/1744). They do not apply to AI systems related to products covered by the Union harmonisation legislation in Section B of Annex I; only the provisions listed in Article 2(2) apply there (Article 2(2), as replaced by Article 1, point (2)).

Article 78 on confidentiality has applied since 2 August 2025 (Article 113, third paragraph, point (b)). In our reading it only becomes relevant to the information under Article 21 once Article 21 itself applies.

If a system was placed on the market or put into service before that date, the Regulation applies to the operators of that system only if it is subject to significant changes in its design from that date. Providers and deployers of systems intended to be used by public authorities must comply by 2 August 2030 (Article 111(2), as replaced by Article 1, point (39), of Regulation (EU) 2026/1744). For the municipality in the example that means: if the system was already in use before 2 December 2027, provider and municipality must comply with Article 20 and Article 26(5) by 2 August 2030, and earlier if the system's design is significantly changed before then.

Fines

Non-compliance with the obligations of providers under Article 16 is subject to administrative fines of up to EUR 15 million or, for an undertaking, up to 3 percent of total worldwide annual turnover for the preceding financial year, whichever is higher (Article 99(4)(a)). Article 16 covers keeping the documentation (point (d)), the corrective actions and information of Article 20 (point (j)) and demonstrating conformity on request (point (k)). Article 21 itself is not in the list of paragraph 4; in our reading its core, demonstrating conformity on request, falls under paragraph 4(a) through Article 16(k). Supplying incorrect, incomplete or misleading information to authorities in reply to a request can lead to a fine of up to EUR 7.5 million or 1 percent (paragraph 5).

For SMEs and start-ups the lower of the amount and the percentage applies (paragraph 6), and since the Omnibus also for SMCs for the fines of paragraphs 4 and 5 (paragraph 6a). The municipality as deployer falls under paragraph 4(e) for Article 26. To what extent fines can be imposed on public authorities is for each Member State to decide (paragraph 8). According to the draft explanatory memorandum to the Dutch AI Act Implementation Act (section 4.5.4.2), the bill does not rule out fines for public authorities: an administrative fine is imposed on the offender, which under Article 5:1(2) of the Dutch General Administrative Law Act (Awb) is whoever commits or co-commits the offence, including a public authority. That is not yet law for the AI Act either. See also fines under the AI Act and the GDPR for municipalities.

What to do now

As a provider:

  1. Set up ten-year record keeping, per version of the system, with the five items of Article 18(1), and make sure they stay available on bankruptcy or takeover.
  2. Lay down a corrective action procedure in your quality management system: who decides, which action for which severity, and how fast.
  3. Keep an up-to-date list of your distributors, importers, authorised representative and deployers, with contact points, so you can inform them without delay.
  4. Create a notice template with version, period, non-compliance, action and what the recipient must do.
  5. Decide when a non-compliance is a risk within the meaning of Article 79(1), and who then informs the market surveillance authority and the notified body.
  6. Make sure you can deliver on a request under Article 21: documentation in a language the authority understands, and access to the logs under your control.

As a deployer:

  1. Agree in the contract how and how fast the provider informs you, and which data you get to review earlier outputs.
  2. Prepare a playbook for a notice under Article 20: read it, decide whether to suspend use, report yourself, check for a serious incident, secure the logs, review earlier decisions, update the FRIA.
  3. Keep your logs for at least six months and make sure they are not deleted automatically while a notice is open.

Not sure whether your system is high-risk and which role you have? With the case check you describe the use in a few questions and see which AI Act and GDPR obligations come into view.

Frequently asked questions about corrective actions and documentation keeping under the AI Act

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

Related obligations

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

When the regulator knocks: Articles 79 to 82 of the AI Act on AI systems presenting a risk

What happens when a market surveillance authority considers that an AI system presents a risk? Article 79 sets out the national procedure: evaluation, corrective action within a period of at most 15 working days, and where needed a prohibition or recall. Article 80 deals with a system the provider wrongly classified as non-high-risk, Article 81 with the Union safeguard procedure and Article 82 with a system that complies with the Regulation and still presents a risk. With who will supervise in the Netherlands, from when it applies, and a municipality using a system for social assistance applications and a bank as examples.

EU AI ActArticle 79Article 80

Can the list of high-risk AI still change? Articles 7 and 8 of the AI Act on amending Annex III and complying with the requirements

Annex III of the AI Act is not a fixed list. Article 7 lets the Commission add, modify and remove use cases by delegated act, within the eight areas and against eleven criteria. Article 8 says what complying with the requirements means: taking into account the intended purpose and the state of the art, and for products under Section A of Annex I, the option to integrate testing and documentation processes. With the Commission's first annual report on Annex III, what the Digital Omnibus on AI did and did not change, a Dutch municipality with a system that currently falls outside Annex III and an insurer as examples.

EU AI ActArticle 7Article 8

EU declaration of conformity and CE marking for high-risk AI: what Articles 47 and 48 of the AI Act require

After the conformity assessment, the provider of a high-risk AI system draws up an EU declaration of conformity for each system and affixes the CE marking (Articles 47 and 48). What the declaration must contain under Annex V, why the signature means taking responsibility, how a digital CE marking works, what authorised representatives, importers and distributors check, what a municipality verifies when it buys, when it applies and what happens when it is missing, with a vendor of a welfare model and a bank as examples.

EU AI ActArticle 47Article 48