Skip to main content
Praxikon

Technical documentation for high-risk AI: what Article 11 and Annex IV of the AI Act ask of providers

···30 min read

From 2 December 2027, a provider of a high-risk AI system listed in Annex III must have drawn up technical documentation before the system is placed on the market or put into service, and must keep that documentation up to date (Article 11(1) AI Act). The documentation must demonstrate that the system complies with the requirements of Chapter III, Section 2, and contains at least the elements of Annex IV. The provider ensures that its system meets those requirements and keeps the documentation (Article 16, points (a) and (d)). An organisation that only uses a system it has bought, the deployer, does not draw up technical documentation. Its own obligations are in Article 26.

Suppose a municipality develops, together with an external IT firm, a model that pre-sorts applications for social support under the Dutch Social Support Act (Wmo), such as help with household tasks. The model reads the application form and earlier case files, and gives the caseworker advice for each application: straightforward to handle, or an in-depth conversation first. To read the free text, the IT firm uses a pre-trained language model from a third party. The system runs under the municipality's name. The project lead asks: "We have a DPIA, an architecture diagram and the IT firm's manual. Isn't that the technical documentation already?" The answer: no. The municipality is the provider here, and Article 11 asks for one coherent file in which the supervisor can read how the system was built, tested and monitored. The example is fictitious.

In short

Article 11 has three paragraphs. Paragraph 1 requires the provider of a high-risk system to draw up technical documentation before placing it on the market or putting it into service and to keep it up to date, in such a way that compliance with Section 2 is demonstrated and that national competent authorities and notified bodies receive the information in a clear and comprehensive form. The documentation contains at least the nine points of Annex IV. SMEs, start-ups and, since the Digital Omnibus (Regulation (EU) 2026/1744), also small mid-cap enterprises may provide those elements in a simplified manner, using a form the Commission establishes for that purpose. Paragraph 2 requires a single set of documentation for products under Section A of Annex I. Paragraph 3 lets the Commission amend Annex IV by delegated act. The provider keeps the documentation at the disposal of the national competent authorities for ten years (Article 18). Under the AI Act, a deployer has no right of its own to that file; it receives the instructions for use under Article 13 and can agree the rest by contract. For Annex III, Article 11 applies from 2 December 2027; for Annex I, Section A, from 2 August 2028.

This piece is about Article 11, Annex IV and what a provider does with them. The process around it is covered in the quality management system of Article 17. How the documentation is used in the conformity assessment is covered in internal control or a notified body. When you become a provider is covered in when do you become a provider.

Who draws up the technical documentation

The obligation rests on the provider. That includes a public authority that develops an AI system, or has it developed, and puts it into service under its own name (Article 3, point (3)). The municipality in the example is therefore the provider, even though the IT firm writes the code. Who does the work does not matter; who puts the system into service under its own name does.

An organisation that buys a system and uses it in accordance with the instructions for use is a deployer and does not draw up technical documentation. If it puts its name on the system, substantially modifies it or changes its intended purpose so that it becomes high-risk, it becomes the provider (Article 25(1)). The initial provider must then cooperate closely with it. Since the Digital Omnibus, the text expressly states that this includes, where relevant, making available technical documentation sufficient to assess compliance with the requirements of Article 16, informing the new provider about known limitations and failure modes, and providing targeted technical access for testing and validation (Article 25(2), as amended by Article 1, point (12), of Regulation (EU) 2026/1744). That does not apply where the initial provider has clearly specified that its system is not to be changed into a high-risk system.

When and how: drawn up before the market, kept up to date afterwards

Article 11(1) has two moments. The documentation is drawn up before the system is placed on the market or put into service, and is kept up to date afterwards. Recital 71 explains why: comprehensible information on how a system was developed and how it performs is needed for traceability, for verifying compliance and for post-market monitoring, and the documentation should be kept up to date appropriately throughout the lifetime of the system.

Keeping it up to date is more than incrementing a version number. Annex IV asks for a description of relevant changes made by the provider through the lifecycle (point 6). A substantial modification requires a new conformity assessment. For systems that continue to learn after being placed on the market or put into service, changes that the provider pre-determined at the initial conformity assessment and described in the technical documentation (Annex IV, point 2(f)) are not a substantial modification (Article 43(4)). In our reading, good documentation in advance therefore limits the risk that a planned change later counts as a substantial modification requiring a new conformity assessment.

The documentation has two readers: national competent authorities and notified bodies. They must receive the information "in a clear and comprehensive form" (paragraph 1). In our reading, a folder of raw exports and notebooks does not meet that standard, even if all the information is in there somewhere.

Annex IV: the nine points

Annex IV lists the information the technical documentation contains at least, "as applicable to the relevant AI system". In our reading, that qualification is not a free pass: a point only drops out if it genuinely does not arise for this system, such as the external features of a product (point 1(f)) for a system that is not a component of a product.

PointWhat Annex IV requiresWhat it means for the municipality (our translation)
1A general description: intended purpose, name of the provider, version and relation to previous versions, interaction with other hardware, software and AI systems, software versions, forms in which the system is placed on the market, hardware, user interface and instructions for use for the deployerWhat the model is and is not intended for, which version runs, how it is connected to the case management system and what the caseworker sees on screen
2A detailed description of the elements and the development process: methods and steps, including pre-trained systems or tools from third parties ((a)), design specifications and key design choices with their rationale and assumptions ((b)), system architecture and computational resources ((c)), data requirements and datasheets on training and training data ((d)), the assessment of human oversight ((e)), pre-determined changes ((f)), validation and testing procedures with metrics for accuracy, robustness and potentially discriminatory impacts, and dated and signed test reports ((g)), and cybersecurity measures ((h))Which language model the IT firm uses and how it was modified, why certain case file fields do not count, which historical applications were used and how they were labelled, and who signed the test reports
3Detailed information about monitoring, functioning and control: capabilities and limitations, accuracy for specific persons or groups and overall, foreseeable unintended outcomes and risks to health, safety, fundamental rights and discrimination, human oversight and input data specificationsWhether the advice works equally well for older people, people with a mild intellectual disability and applicants who fill in the form in limited Dutch
4A description of the appropriateness of the performance metrics for this systemWhy the municipality chooses this particular metric, and not only an average across all applications
5A detailed description of the risk management system under Article 9The risk analysis and the measures, not only the DPIA
6A description of relevant changes through the lifecycleA change log per model version
7The harmonised standards applied or, where none were applied, a detailed description of the solutions adopted to meet the requirements of Section 2, with the other standards and technical specifications appliedWhich standards the IT firm follows, and how the municipality substantiates the requirements where there is no standard
8A copy of the EU declaration of conformity under Article 47The signed declaration of the municipality itself
9A detailed description of the system for evaluating performance in the post-market phase (Article 72), including the post-market monitoring plan under Article 72(3)How often the advice is compared with the final decisions, and what happens when the two drift apart

The right-hand column is our practical translation, not text from the Regulation. It does show where the project lead in the example goes wrong. The DPIA touches on point 2(d) and part of points 3 and 5. The architecture diagram touches on point 2(c). The IT firm's manual is at most a start on point 1. What is missing is, we expect, exactly what a supervisor asks for first: the test reports, the accuracy per group, the change log and the monitoring plan.

Two points deserve extra attention for the municipality. Point 2(a) asks how the provider used, integrated or modified pre-trained systems or tools from third parties. For that, the municipality needs information from the supplier of the language model. Article 25(4) requires the provider of a high-risk system and the third party that supplies an AI system, AI model, tools, services, components or processes to specify by written agreement the information, capabilities, technical access and other assistance the provider needs to comply with the Regulation. That does not apply to third parties making tools, services, processes or components, other than general-purpose AI models, accessible to the public under a free and open-source licence. The AI Office may recommend voluntary model terms for such contracts. If the language model is a general-purpose AI model, its provider must moreover itself draw up, keep up to date and make available information and documentation to providers that intend to integrate the model into their AI system, containing at a minimum the elements of Annex XII (Article 53(1)(b)). That obligation does not apply to certain models released under a free and open-source licence, unless the model has systemic risk (Article 53(2)). Point 8, the declaration of conformity, comes only at the end: the file is complete only after the conformity assessment.

SMEs and small mid-caps: the simplified form

SMEs, including start-ups, could already provide the elements of Annex IV in a simplified manner. Originally, the Commission was to establish a simplified form for this "targeted at the needs of small and microenterprises". The Digital Omnibus replaced that subparagraph (Article 1, point (10), of Regulation (EU) 2026/1744). Since 27 July 2026, small mid-cap enterprises (SMCs) may also choose the simplified route, and the form is targeted at the needs of SMEs, start-ups and SMCs. On 18 November 2026 we could not yet find that form from the Commission; check whether it is available by now. The notion of SME follows Recommendation 2003/361/EC, the notion of SMC Recommendation (EU) 2025/1099 (Article 3, points (14a) and (14b)). Recital 6 of the Omnibus explains why: enterprises that just outgrow the SME threshold face the same administrative burden.

Three things did not change. Whoever chooses the simplified route must use the Commission's form. Notified bodies must accept the form for the purposes of the conformity assessment. And the elements of Annex IV stay the same: the Regulation simplifies the manner of providing them, not what must be demonstrated. In our reading, this also means that the requirements of Section 2 themselves, such as accuracy per group and risk management, do not become lighter for a small provider.

Do not confuse this with Article 63. That article lets certain SMEs comply with elements of the quality management system in a simplified manner, and provides in paragraph 2 that this is no exemption from other obligations, expressly including Article 11. How that works is covered in our piece on Article 17.

A municipality is not an enterprise within the meaning of the SME Recommendation. In our reading, the simplified form is therefore not intended for it. That the IT firm building for it is itself an SME does not change this, in our reading: the firm is not the provider here, and the choice of the simplified route lies with the provider.

Paragraph 2: a single set of documentation for products under Section A of Annex I

If the high-risk system is related to a product covered by the harmonisation legislation in Section A of Annex I, a single set of technical documentation is drawn up containing all the information of paragraph 1 and the information required under that product legislation (paragraph 2). A manufacturer of a medical device with an AI component therefore does not create a second file next to the technical documentation under the MDR, but one file containing both. This fits Article 43(3): the requirements of Section 2 are assessed within the conformity assessment of the product legislation. The declaration of conformity then also becomes a single declaration for all applicable Union law (Article 47(3)). How that works out for medical AI is covered in medical AI under the AI Act and the MDR.

The Omnibus deleted the Machinery Directive (Directive 2006/42/EC) from Section A and added the Machinery Regulation (Regulation (EU) 2023/1230) to Section B of Annex I (Article 1, point (41)). For systems under Section B, only Article 6(1), Article 60a and Articles 102 to 112 apply, and Articles 57, 58 and 59 in so far as the AI Act requirements have been integrated in that product legislation (Article 2(2), as amended by Article 1, point (2)). Article 11(2) therefore does not come into play there. For systems under Section A, Article 11 applies from 2 August 2028.

For high-risk systems referred to in Article 6(1), the application of specific requirements or obligations of Articles 9 to 15 and 17 to 25, and therefore also of Article 11, may moreover be limited, where and to the extent that the harmonisation legislation in Section A of Annex I lays down requirements providing an equivalent or higher level of protection and the limitation does not reduce the overall level of protection of the AI Act (Article 2(13), added by Article 1, point (3), of the Omnibus). By 2 August 2027, the Commission adopts delegated acts specifying the systems and requirements concerned, the conditions under which the limitation applies and its scope.

Retention: ten years, and who may look

The provider keeps the technical documentation at the disposal of the national competent authorities for a period ending ten years after the system has been placed on the market or put into service (Article 18(1)). The same applies to the documentation concerning the quality management system, the changes approved by notified bodies and their decisions, and the EU declaration of conformity. Paragraph 1 names the two starting moments side by side without choosing. In our reading, counting from the later of the two is the safe choice, because you then do not fall short under either reading.

Three additions are worth noting. Paragraph 2 leaves it to the Member States to determine what happens to the documentation if a provider or its authorised representative goes bankrupt or ceases its activity. If the provider is established outside the EU, its authorised representative also keeps the technical documentation available for ten years (Article 22(3)(b)). And the logs that the system generates automatically are not covered by Article 18 but by Article 19, with its own period of at least six months.

Who may look? Upon a reasoned request by a competent authority, the provider supplies all the information and documentation necessary to demonstrate conformity with Section 2, in a language which can be easily understood by the authority, in one of the official languages of the institutions of the Union as indicated by the Member State concerned (Article 21(1)). Where relevant and limited to what is necessary to fulfil their tasks, market surveillance authorities are granted full access to the documentation and the training, validation and testing data sets, where appropriate remotely (Article 74(12)). They are granted access to the source code only upon a reasoned request, where that is necessary and where testing or auditing based on the documentation has been exhausted or proved insufficient (paragraph 13). What the authority obtains in this way is covered by the confidentiality obligations of Article 78. Good technical documentation is therefore also a way of preventing a supervisor from having to dig deeper.

How Article 11 connects with Articles 9, 10, 12, 13, 15, 17, 43, 47 and 72

Article 11 adds few new substantive requirements. It asks the provider to show how it meets the other requirements. That is why Annex IV is best read as a table of contents of Section 2 and the obligations around it.

Article 9: risk management (point 5). The risk management system is a continuous iterative process throughout the lifecycle (Article 9). Annex IV asks for a detailed description of it, and point 3 asks for the foreseeable unintended outcomes and sources of risk. See the Article 9 risk management system.

Article 10: data (point 2(d)). The datasheets on training and training data, the provenance and selection of data, labelling and cleaning are the record of the data governance under Article 10. The Omnibus moved the exception for special categories of personal data to detect bias to Article 4a. See also data governance under Article 10.

Article 12: logging (point 2(g)). Annex IV asks for test logs and signed test reports. Those are different logs from the automatic recording of events during use that Article 12 must technically allow for. In our reading, the description of that logging capability does belong in the file, under point 2(c) (architecture) and point 3 (functioning and control). See logs for high-risk AI.

Article 13: instructions for use (point 1(h), point 2(e) and point 3). The instructions for use for the deployer are themselves part of the technical documentation, and point 2(e) refers to Article 13(3)(d) for the measures that facilitate the interpretation of the output. The difference: the technical documentation is for the supervisor, the instructions for use are for the user. See instructions for use under Article 13. For human oversight itself: Article 14.

Article 15: accuracy, robustness and cybersecurity (point 2(g) and (h), and points 3 and 4). The metrics for accuracy and robustness, the cybersecurity measures and the appropriateness of the performance metrics substantiate Article 15. See our piece on Article 15.

Article 17: quality management. The quality system is the process; the technical documentation is what that process delivers for each system. Article 17(1) requires, among other things, systems and procedures for record-keeping of all relevant documentation (point (k)), and Annex IV, point 7, ties in with the standards and specifications of point (e). See the quality management system.

Article 43: conformity assessment. Under internal control, the provider examines the information in the technical documentation to assess whether the system complies with Section 2, and verifies that the design and development process and the post-market monitoring are consistent with that documentation (Annex VI, points 3 and 4). Under the route with a notified body, that body assesses the technical documentation (Annex VII, point 4). For the municipality, internal control applies: Annex III, point 5 falls under Article 43(2).

Article 47: the EU declaration of conformity (point 8). A copy of the declaration belongs in the file. The declaration itself has its own ten-year retention period and must be submitted to the authorities upon request (Article 47(1)).

Article 72: post-market monitoring (point 9). The monitoring plan is part of the technical documentation, and the Commission adopts guidance, including a template, for that plan by 2 September 2027 (Article 72(3), as amended by Article 1, point (30), of the Omnibus). See monitoring under Article 72.

What a deployer can and cannot ask of its vendor

Suppose a neighbouring municipality buys a comparable social support system from a vendor. It is the deployer. Can it request the technical documentation?

What not. The AI Act gives the deployer no right to the Annex IV file. Article 11(1) names national competent authorities and notified bodies as its readers, and Articles 18 and 21 govern its availability to authorities. The obligation to cooperate, including handing over technical documentation, in Article 25(2) only applies once the deployer itself becomes the provider.

What yes, under the Regulation. The system is accompanied by instructions for use with concise, complete, correct and clear information (Article 13(2) and (3)). They contain much of what is also in Annex IV: the accuracy and the metrics against which the system was tested, where appropriate the performance for specific persons or groups, the pre-determined changes, the human oversight measures, the computational resources needed and expected lifetime, and how the logs can be collected and interpreted. The system bears the CE marking, and importers and distributors verify that it is accompanied by (a copy of) the EU declaration of conformity (Article 23(1)(c), Article 24(1)). If the deployer is a public authority, it does not use a system that is not registered in the EU database (Article 26(8)).

What it must do itself. The deployer uses the system in accordance with the instructions for use, assigns human oversight, ensures, to the extent it controls the input data, that the data is relevant and sufficiently representative, and monitors the operation (Article 26(1), (2), (4) and (5)). It uses the information under Article 13 for its DPIA (Article 26(9)). As a public authority, it registers itself and its use of the system in the EU database (Article 26(8) and Article 49(3)). As deployer of an Annex III, point 5 system, a municipality also carries out a fundamental rights impact assessment (Article 27); see the FRIA template. The municipality in the first example is both provider and deployer: as provider it registers itself and the system (Article 49(1)), as deployer it registers its use (Article 49(3)) and carries out its own fundamental rights impact assessment (Article 27).

What it can agree by contract. Anything beyond that is a procurement question. In our reading, these are reasonable questions to a vendor: a summary of the test reports with accuracy per relevant group, advance notice of new versions and substantial modifications, cooperation when a supervisor asks the municipality questions about the system, and an arrangement for the documentation if the vendor stops or goes bankrupt. In that case, Article 18(2) only lets the Member States ensure that the documentation remains at the disposal of the national competent authorities, not of the deployer. A vendor may protect its trade secrets; access under confidentiality is then a middle way. A vendor that cannot answer any of these questions tells you something, in our reading, about the state of its file. See also assessing AI vendors in procurement.

The GDPR: accountability, records and DPIA

The AI Act does not affect the GDPR, without prejudice to Articles 4a and 59 (Article 2(7), as amended by the Omnibus). The municipality in the example processes personal data, including health data, to develop and use the model. Three GDPR obligations then run parallel to Article 11.

  • Article 5(2) GDPR, the accountability principle: the controller is responsible for, and must be able to demonstrate, compliance with the principles (Article 5 GDPR). The same idea as in Article 11: it is not only about doing, but about being able to show.
  • Article 30 GDPR, the record of processing activities, including the purposes and the categories of data subjects and personal data (Article 30 GDPR). In our reading, the municipality can align the description of the training data under Annex IV, point 2(d), with its record.
  • Article 35 GDPR, the data protection impact assessment. It is required in particular for a systematic and extensive evaluation of personal aspects based on automated processing, including profiling, on which decisions are based that produce legal effects or similarly significantly affect the person (Article 35(3)(a) GDPR), and for processing on a large scale of special categories of personal data, such as health data (point (b)). A DPIA is not technical documentation, and technical documentation is not a DPIA. The DPIA can, however, draw on points 2 and 3 of Annex IV, and the risk management of point 5 can take the outcomes of the DPIA into account.

The training itself also requires a GDPR assessment. In our reading, training the model on earlier social support case files is further processing for a purpose other than that for which the data were collected. If that processing is not based on consent or on a legal provision, the municipality assesses whether the new purpose is compatible with the original one (Article 6(4) GDPR). For the health data in those files, the prohibition in Article 9(1) GDPR also applies, so the municipality needs an exception under Article 9(2). Article 4a(1) AI Act exceptionally allows providers to process special categories of personal data, but only to the extent strictly necessary for detecting and correcting bias, and subject to the conditions of that paragraph. In our reading, that is not a legal basis for the training as a whole.

If the caseworker follows the advice almost always in practice, the advice itself can, in our reading, fall under Article 22 GDPR. According to the SCHUFA judgment of the Court of Justice (C-634/21, 7 December 2023), an automatically established score can already be a decision where someone else draws strongly on it in deciding.

In our reading, the municipality therefore does much of the work once, but fills two different files for two different supervisors.

From when?

Article 11 is in Chapter III, Section 2. For high-risk systems under Annex III that Section applies from 2 December 2027, for systems under Section A of Annex I from 2 August 2028 (Article 113, third paragraph, point (c), as amended by Regulation (EU) 2026/1744). The municipality's system assesses whether residents are eligible for an essential public service and, in our reading, falls under Annex III, point 5(a). Because the model profiles applicants, in our reading the municipality cannot rely on the derogation in Article 6(3) (last subparagraph). Even a provider that considers an Annex III system not to be high-risk must document: it records that assessment before the system is placed on the market or put into service, and provides it upon request (Article 6(4)).

If the system was placed on the market or put into service before 2 December 2027, the high-risk obligations only apply if it undergoes significant changes in its design from that date; for systems intended to be used by public authorities, providers and deployers must comply with the Regulation by 2 August 2030 at the latest (Article 111(2), as amended by Article 1, point (39), of Regulation (EU) 2026/1744). For a municipality building now, that is no reason to wait. Test reports you do not date and sign today cannot be reconstructed in 2030.

Example 2: the bank

A bank that develops its own model to assess the creditworthiness of individuals is the provider of a high-risk system under Annex III, point 5(b). It draws up the full technical documentation of Annex IV. The Regulation has a specific rule for financial institutions on retention, not on content: the bank keeps the technical documentation as part of the documentation it keeps under Union financial services law (Article 18(3)). The financial institutions rule in Article 17(4) concerns the quality management system, not the technical documentation.

In our reading, the bank can reuse much of the model documentation it already has, but Annex IV asks for things that are not automatically in it: the accuracy for specific persons or groups (point 3), the metrics for potentially discriminatory impacts (point 2(g)) and a description of human oversight with the measures that facilitate the interpretation of the output (point 2(e)). If the bank uses an external model or a general-purpose AI model as a building block, point 2(a) asks how it integrated or modified it, and it agrees in writing with the supplier which information it receives for that purpose (Article 25(4)). As deployer of its own model, the bank also carries out a fundamental rights impact assessment (Article 27(1)).

A fintech that builds the same kind of model and is an SME or SMC may provide the elements of Annex IV in a simplified manner using the Commission's form, once the Commission has published that form. A fintech that is not subject to governance requirements under Union financial services law keeps the documentation, in our reading, under the general rule of Article 18(1). How existing banking rules otherwise relate to the AI Act is covered in the EBA mapping for the financial sector.

What to do now

  1. Establish whether you are the provider. Do you build a high-risk system yourself, have it built and put it into service under your name, or substantially modify it? Then you draw up the technical documentation. If you only use a system you bought, you do not.
  2. Map your existing documents against Annex IV. Mark for each of points 1 to 9 what is already there (DPIA, architecture, test reports, model documentation) and what is missing.
  3. Arrange the information from third parties. Record in writing with the supplier of a pre-trained model or component which information you receive (Article 25(4)), because without that information you cannot complete point 2(a).
  4. Document pre-determined changes. If your system continues to learn after being placed on the market or put into service, describe at the initial conformity assessment which changes to the system and its performance you pre-determine (point 2(f)). Such pre-determined changes are not a substantial modification (Article 43(4)); any other substantial modification requires a new conformity assessment.
  5. Determine your size and status. Are you an SME, start-up or small mid-cap? Record that assessment. Include partner and linked enterprises in it (Recommendation 2003/361/EC). Check whether the Commission has already published the simplified form before you choose that route.
  6. Arrange retention for ten years. Designate a place of retention, record the date of placing on the market and of putting into service, and count from the later moment.
  7. Are you a deployer? Ask for the instructions for use, the declaration of conformity and the registration in the EU database, and agree in the contract what else you need.

Unsure whether your system is high-risk and whether you are provider or deployer? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.

Frequently asked questions about Article 11 and Annex IV 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

The quality management system for high-risk AI: what Article 17 of the AI Act asks of providers

A provider of a high-risk AI system must have a quality management system that ensures compliance with the AI Act, documented in written policies, procedures and instructions (Article 17). That obligation rests on the provider, not on the organisation that only uses a system it has bought. What the thirteen aspects of paragraph 1 require, what the Digital Omnibus changed for SMEs and small mid-caps, how sectoral quality systems and bank governance count, what EN 18286:2026 does and does not prove, and how it connects with Articles 9, 10, 11, 12, 72 and 73, with a municipality that builds its own system and a bank as examples.

EU AI ActArticle 17Quality management

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

Deployer obligations for high-risk AI: what Article 26 of the AI Act requires

An organisation that uses a high-risk AI system has twelve paragraphs of obligations of its own (Article 26 AI Act): use in accordance with the instructions, human oversight by competent people, sound input data, monitoring and notification, keeping logs, informing workers and affected persons, registration by public authorities, feeding the DPIA and cooperating with the authorities. What each paragraph requires, what the Digital Omnibus did and did not change, the rules for banks and insurers, the difference with the provider, and how the FRIA and the right to explanation connect, with a municipality, a bank and an employer as examples.

EU AI ActArticle 26Deployer