Skip to main content
Praxikon

Instructions for use of a high-risk AI system: what Article 13 of the AI Act requires and what you do with them

··17 min read

A high-risk AI system comes with instructions for use that are concise, complete, correct and clear, with at least six elements: who the provider is, what the system can and cannot do, which changes have been pre-determined, which human oversight measures belong to it, what maintenance requires, and how you collect and read the logs (Article 13 AI Act). That duty rests on the provider. For the deployer, the organisation that uses the system, the instructions are the yardstick of its own duties: use in accordance with the instructions, human oversight, monitoring, logs and the DPIA (Article 26 AI Act) and, where one is required, the FRIA (Article 27 AI Act). For Annex III systems those duties apply from 2 December 2027. Anyone procuring now should therefore secure good instructions in the contract now.

Picture this: a Dutch municipality buys a system that analyses applications for social assistance and advises the caseworker whether the applicant is likely to be entitled. The vendor's manual covers logging in and reports; on accuracy it says only "high reliability". Enough?

In short

Article 13(3) AI Act prescribes six elements. The deployer uses the system in accordance with the instructions, monitors on the basis of them, keeps the logs for at least six months and uses the information for the DPIA (Article 26(1), (5), (6) and (9) AI Act). The instructions are something other than the technical documentation (Article 11 AI Act) and than the information given to data subjects (Articles 13 and 14 GDPR). For Annex III the duties apply from 2 December 2027; for systems already placed on the market or put into service earlier, Article 111(2) AI Act applies.

What instructions for use are

Instructions for use are the information provided by the provider to inform the deployer of, in particular, an AI system's intended purpose and proper use (Article 3(15) AI Act). They must enable the deployer to interpret the output and use it appropriately (Article 13(1) and (2) AI Act). According to Recital 72 they also help the deployer choose the right system and know the intended and precluded uses, where appropriate with examples, in a language the deployer can easily understand, as determined by the Member State concerned. That is a recital, not an operative provision, and as far as we know the Netherlands has not yet set that language. A manual that only explains the buttons does not meet the standard.

What they must contain at a minimum (Article 13(3) AI Act)

Article 13(3) AI ActWhat the provider statesWhat you use it for
(a)Identity and contact details of the provider and any authorised representativeReporting risks and serious incidents (Article 26(5) AI Act)
(b)(i)The intended purposeChecking that your use stays within it
(b)(ii)The level of accuracy, including its metrics, robustness and cybersecurity against which the system was tested, and what may affect itDeciding whether performance fits your decisions
(b)(iii)Foreseeable circumstances, including reasonably foreseeable misuse, that may lead to risks to health, safety or fundamental rightsRisk analysis in the DPIA and FRIA
(b)(iv)Where applicable: the capabilities to provide information that explains the outputExplanations to data subjects
(b)(v)When appropriate: performance regarding specific persons or groupsTesting for unequal outcomes
(b)(vi)When appropriate: specifications for the input data and information on the data setsChecking that your input data fit (Article 26(4) AI Act)
(b)(vii)Where applicable: information to interpret the outputWork instructions and training
(c)Changes pre-determined at the initial conformity assessmentKnowing which updates are expected
(d)The human oversight measures referred to in Article 14 AI ActSetting up oversight (Article 26(2) AI Act)
(e)Computational and hardware resources, expected lifetime and maintenance, including software updatesOperations and update policy
(f)Where relevant: how to collect, store and interpret the logs (Article 12 AI Act)Log management (Article 26(6) AI Act)

In our reading, "where applicable", "when appropriate" and "where relevant" mean that an item drops out if it does not apply to the system, not that the provider may freely choose. And "high reliability" without a metric and margin of error is not a level of accuracy within the meaning of point (b)(ii).

What you as deployer do with the instructions

These duties apply once Chapter III applies to your system, for Annex III from 2 December 2027 (see "From when").

Use and oversight. You take appropriate technical and organisational measures to use the system in accordance with the instructions for use (Article 26(1) AI Act), and assign oversight to people with the necessary competence, training and authority (paragraph 2). The provider determines the oversight measures; you decide who carries them out and how (paragraph 3). The provider must supply the system so that those people, as appropriate and proportionate, understand its limitations, remain aware of over-reliance on its output, can disregard or reverse the output and can stop the system (Article 14(4) AI Act). See human oversight under Article 14.

Input data. To the extent you exercise control over them, you ensure the input data are relevant and sufficiently representative (paragraph 4).

Monitoring and intervening. You monitor on the basis of the instructions for use (paragraph 5). If use in accordance with the instructions may result in a risk within the meaning of Article 79(1) AI Act, you inform the provider or distributor and the market surveillance authority without undue delay, and suspend the use. Where you identify a serious incident, you immediately inform first the provider, and then the importer or distributor and the market surveillance authorities; if you cannot reach the provider, Article 73 AI Act applies mutatis mutandis.

Logs. You keep the logs under your control for a period appropriate to the intended purpose, of at least six months, unless other law, in particular data protection law, provides otherwise (paragraph 6). Which events are logged follows from Article 12 AI Act.

DPIA. Where applicable, you use the information provided under Article 13 AI Act for your DPIA under Article 35 GDPR (paragraph 9).

For informing affected persons (paragraph 11) and for the explanation under Article 86 AI Act, you need points (b)(iv) and (vii).

Not the same: instructions, technical documentation and Article 13 GDPR

Technical documentation. The provider draws it up to demonstrate that the system complies with the requirements, for authorities and notified bodies (Article 11(1) AI Act), and keeps it at the disposal of the national competent authorities for ten years (Article 18 AI Act). The instructions for use are part of it (Annex IV, point 1(h), AI Act). The AI Act gives the deployer no right to that documentation. Only a party that itself becomes the provider under Article 25(1) can require the initial provider to make available technical documentation sufficient to assess compliance with Article 16, unless that provider has clearly specified that its system is not to be changed into a high-risk AI system (Article 25(2) AI Act). What you need must therefore be in the instructions or the contract.

Articles 13 and 14 GDPR. These are the information duties of the controller towards the data subject (Article 13 GDPR where you collect the data from the data subject, Article 14 GDPR where they come from elsewhere), including, for automated decision-making within the meaning of Article 22 GDPR, meaningful information about the logic involved (Article 13(2)(f) GDPR and Article 14(2)(g) GDPR). The direction is reversed: Article 13 AI Act runs from provider to deployer, Article 13 GDPR from controller to citizen, customer or applicant.

The link is in the content. A credit score established automatically by a credit information agency can itself be an automated decision where a third party draws strongly on it to decide whether to enter into a contract (C-634/21, SCHUFA, 7 December 2023). In a case on the right of access in Article 15(1)(h) GDPR, which according to the Court forms a whole with Article 13(2)(f) and Article 14(2)(g) GDPR, the Court held that meaningful information about the logic involved requires an explanation of the procedure and principles actually applied; a formula or a description of every step is not enough (C-203/22, Dun & Bradstreet Austria, 27 February 2025, ECLI:EU:C:2025:117, paragraphs 46, 58, 59 and 61). That explanation requires information from the provider.

The DPIA and the FRIA

A DPIA contains, among other things, a description of the processing, the risks to data subjects and the measures to address them (Article 35(7) GDPR). The instructions fill exactly those parts. See DPIA for AI systems: when it is required.

The FRIA applies only to high-risk systems under Annex III (Article 27(1) refers to Article 6(2)), not to Annex I, and only to bodies governed by public law, private entities providing public services and deployers of systems for creditworthiness and for life and health insurance (Annex III, point 5(b) and (c)); critical infrastructure (point 2) is excluded. In it you take into account the information under Article 13 AI Act and describe oversight according to the instructions for use (Article 27(1)(d) and (e) AI Act). Where the DPIA already covers part of it, you may cross-refer (paragraph 4). See the FRIA template and DPIA and FRIA side by side.

Procurement: what to require, and what if the instructions fall short

The provider ensures that its system complies with Article 13 AI Act (Article 16(a) AI Act); importers and distributors verify that the instructions accompany it (Article 23(1)(c) and Article 24(1) AI Act). In the contract you lay down:

  1. Complete instructions before you choose, in your own language (a contractual arrangement of your own).
  2. Measurable accuracy: metrics, margins of error and performance per group.
  3. Updated instructions with every update.
  4. Oversight and logs: disregarding output, stopping the system, exporting logs.
  5. Support for the DPIA and FRIA, and a processing agreement under Article 28 GDPR where the vendor processes personal data on your behalf.
  6. Information for explanations to data subjects (Article 15(1)(h) GDPR and Article 86 AI Act).
  7. A fixed contact point for reports under Article 26(5) AI Act.

Public buyers can start from the non-binding EU model contractual AI clauses on the European Commission's Public Buyers Community platform, with a version for high-risk systems. They date from 5 March 2025, before the Digital Omnibus, so check them against the current text. See also assessing AI vendors.

Instructions missing or inadequate? From 2 December 2027 (Annex III), or once the duties apply to your system, you cannot comply with Article 26(1) AI Act without instructions, and the system does not comply with Article 13 AI Act; for systems already placed on the market or put into service earlier, Article 111(2) applies. Do not put such a system into use then, and ask in writing for the gaps to be filled; a provider that has reason to consider its system not in conformity immediately takes corrective action (Article 20(1) AI Act). Article 85 AI Act gives the right to lodge a complaint with the market surveillance authority, but in the Netherlands that authority has not yet been designated: the government has proposed supervision per sector, with the Dutch Data Protection Authority where no supervisor exists yet (Rijksoverheid, 20 April 2026, in Dutch). Until the duties apply, the contract is your main lever.

Nor should you fill the gaps by using the system outside its intended purpose. Once Article 26(1) applies, that breaches it. A change of the intended purpose for which the system was assessed can also be a substantial modification (Article 3(23) AI Act); if the system remains high-risk, you become its provider (Article 25(1)(b) AI Act). See when do you become a provider. The Commission must develop guidelines on, among other things, Articles 8 to 15, 25 and 26 and on substantial modification (Article 96(1)(a) and (c) AI Act); check whether they are available.

From when

Articles 13 and 26 AI Act are in Chapter III, Sections 2 and 3. These apply to Annex III from 2 December 2027 and to Annex I from 2 August 2028 (Article 113, third paragraph, point (c), AI Act, as amended by Regulation (EU) 2026/1744). If a high-risk system was placed on the market or put into service earlier, the duties apply only after significant changes in its design; systems intended for use by public authorities must comply by 2 August 2030 at the latest (Article 111(2) AI Act). Whether that also covers an organisation that takes a licence after 2 December 2027 on a system the provider placed on the market earlier is unsettled. In EU product law each individual product is placed on the market (Blue Guide 2022, section 2.3); how that works for software under the AI Act has not been settled. Do not count on it. The DPIA duty under Article 35 GDPR already applies.

Three examples

The examples are fictitious.

Municipality. A system used by or on behalf of a public authority to evaluate whether people are eligible for essential public assistance benefits is listed in Annex III, point 5(a), AI Act. It predicts per applicant, based on personal data, whether they are entitled and thereby evaluates their economic situation; that is profiling (Article 3(52) AI Act and Article 4(4) GDPR), so the exception in Article 6(3) AI Act does not apply (last subparagraph). The municipality asks for margins of error and performance per group; caseworkers may disregard the advice. It carries out a DPIA and, as a body governed by public law, a FRIA, notifying its results to the market surveillance authority (Article 27(3) AI Act). It registers its use in the EU database (Article 26(8) and Article 49(3) AI Act) and informs applicants (Article 26(11) AI Act), in addition to the information under Article 13 GDPR and, because part of the data comes from other sources, Article 14 GDPR. If the system was already running before 2 December 2027, Article 111(2) AI Act applies. On selection for fraud checks: AI for welfare fraud signals.

Bank. A bank uses a vendor's system to assess the creditworthiness of consumers (Annex III, point 5(b), AI Act) and therefore carries out a FRIA. The intended purpose is consumer credit; the bank also wants to use it for mortgages. That too falls under point 5(b), so the system remains high-risk; if mortgage credit falls outside the intended purpose for which it was assessed, that can be a substantial modification and the bank becomes the provider (Article 25(1)(b) with Article 3(23) AI Act). If the bank in effect decides on the basis of the score, without meaningful human assessment, there may be an automated decision within the meaning of Article 22 GDPR. Because it is subject to internal governance requirements under Union financial services law, its monitoring duty under Article 26(5) AI Act is deemed fulfilled; the duties to report risks and serious incidents and to suspend use where needed remain. It keeps the logs as part of the documentation under that law (paragraph 6). See from scoring algorithm to transparent credit decision.

Employer. A platform that analyses applications and ranks candidates falls under Annex III, point 4(a), AI Act. The instructions state that it is less accurate for CVs with career breaks. Recruiters assess those CVs themselves and are trained on the risk of over-relying on the ranking (Article 26(2) and Article 4 AI Act). The employer ensures that the input data it supplies, such as job requirements and vacancy data, are relevant and sufficiently representative for the intended purpose (Article 26(4) AI Act), and informs workers' representatives and affected workers beforehand (paragraph 7). As a private employer not providing public services it needs no FRIA, but a DPIA if the processing is likely to result in a high risk. See AI in recruitment and selection.

Six steps

  1. Establish whether the system is high-risk, for example with the case check.
  2. Test the instructions before you choose; see the obligation page on Article 13.
  3. Translate them into work instructions, including for departing from the output.
  4. Set up monitoring and log management.
  5. Feed the DPIA and the FRIA with the instructions.
  6. Keep everything up to date: new instructions with every update, the DPIA when the risk changes (Article 35(11) GDPR) and the FRIA when one of its elements changes (Article 27(2) AI Act).

Frequently asked questions about instructions for use under Article 13 AI Act

Sources

EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI) (8 July 2026, accessed 17 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 4, 13, 14, 15, 22, 28 and 35 (27 April 2016, accessed 17 November 2026)
Court of Justice of the European Union: Judgment of 7 December 2023, SCHUFA Holding and Others (Scoring), C-634/21, ECLI:EU:C:2023:957 (7 December 2023, accessed 17 November 2026)
Court of Justice of the European Union: Judgment of 27 February 2025, CK v Magistrat der Stadt Wien (Dun & Bradstreet Austria), C-203/22, ECLI:EU:C:2025:117 (27 February 2025, accessed 17 November 2026)
European Commission, Public Buyers Community: Updated EU AI model contractual clauses (5 March 2025, accessed 17 November 2026)
European Commission: The Blue Guide on the implementation of EU product rules 2022 (2022/C 247/01), section 2.3 (29 June 2022, accessed 17 November 2026)
Rijksoverheid (Dutch government): Kabinet zet stap met toezicht op Europese AI-regels (in Dutch) (20 April 2026, accessed 17 November 2026)

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

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

Is your AI system on the Dutch DPA's mandatory DPIA list?

An AI system is never on the list of the Dutch Data Protection Authority as AI. The list looks at what you do with which data, and for most categories only when that happens on a large scale or systematically. What the 17 categories actually say, with examples from a municipality, a bank, an insurer, an employer and a school.

DPIAGDPRDutch Data Protection Authority

Regulators will look at automated candidate screening in 2027: what to put in place now

On 11 November 2026 the European data protection authorities chose the topic of their joint enforcement action for 2027: automated decisions in recruitment, including with AI. What it means for the privacy officer at a municipality, a bank or a school, which questions to expect under Articles 15 and 22 GDPR, and how the AI Act adds to that from 2 December 2027.

GDPRArticle 22 GDPRRecruitment