When do the AI Act obligations start to apply to a public-law body?

government and public services

Awarding social assistance in a municipality: the FRIA and the notification

A municipality wants to deploy an AI system that sorts applications for social assistance benefits and indicates which files merit extra scrutiny before a case worker decides. The application is already listed in the public algorithm register. The question is what has to be in place before the first citizen passes through this system.

Provenance: Article 27(1) requires deployers which are bodies governed by public law, or private entities providing public services, to perform an assessment of the impact on fundamental rights that the use of a high-risk AI system referred to in Article 6(2) may produce, prior to deploying it, with the exception of systems intended to be used in the area listed in point 2 of Annex III. That assessment covers, among other elements, the categories of natural persons and groups likely to be affected, the specific risks of harm to those categories, the implementation of human oversight measures, and the measures to be taken if those risks materialise, including the arrangements for internal governance and complaint mechanisms. Article 27(3) provides that once the assessment has been performed, the deployer shall notify the market surveillance authority of its results and submit the filled-out template referred to in paragraph 5 as part of that notification, and that in the case referred to in Article 46(1) deployers may be exempt from that obligation to notify. Article 27(5) provides that the AI Office shall develop a template for a questionnaire, including through an automated tool, to facilitate deployers in complying with their obligations under this Article in a simplified manner.

We read Article 27 as making the municipality the most obvious deployer here, and as requiring the assessment to be complete before the first application runs through the system, not as an account rendered afterwards. That duty does depend first on whether this system is high-risk at all: does it help decide entitlement to social assistance, or does it stay within a preparatory or narrowly procedural task under Article 6(3), which closes its own exception again once the system profiles citizens? Answer that question before you start on paragraph 1. An entry in the public algorithm register is on our reading something different from the assessment under paragraph 1, and it does not replace notifying the market surveillance authority of the results. In practice it pays to record the citizen's complaint route and the case worker's room to depart from the signal in the same file, because paragraph 1 asks for precisely those two elements.

Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.

Article 27(1), (3) and (5)

law enforcement

Recidivism scoring in police work: when the assessment must be redone

A police service deploys an AI system that estimates the recidivism risk of a suspect, as an aid to the judgements later made by the prosecution service and the court. The model is subsequently retrained on newer investigative data and use is extended to a second region. The question is whether the assessment made for first use remains adequate.

Provenance: Article 27(1) requires deployers which are bodies governed by public law to perform, prior to deploying a high-risk AI system referred to in Article 6(2), an assessment of the impact on fundamental rights that its use may produce, with the exception of systems intended to be used in the area listed in point 2 of Annex III. That assessment consists of a description of the processes in which the system will be used, of the period and frequency of use, of the categories of natural persons and groups likely to be affected, of the specific risks of harm taking into account the information given by the provider pursuant to Article 13, of the implementation of human oversight measures according to the instructions for use, and of the measures to be taken if those risks materialise. Article 27(2) provides that the obligation applies to the first use, that previously conducted impact assessments or existing assessments carried out by the provider may be relied on in similar cases, and that a deployer who considers during use that any element listed in paragraph 1 has changed or is no longer up to date shall take the necessary steps to update the information. Article 27(3) provides that the results are notified to the market surveillance authority together with the filled-out template, and that in the case referred to in Article 46(1) an exemption from that notification duty may apply.

We read Article 27 as covering a police service as a body governed by public law under paragraph 1, and as an assessment that does not stop at first use: retraining on newer investigative data or extending use to a second region touches the elements of paragraph 1 and, on our reading, calls for updating the record. That retraining also raises a question Article 27 itself does not answer, namely whether the change goes far enough to count as a substantial modification, which would make you a provider in your own right under Article 25. Bear in mind as well that the exception in paragraph 3 concerns, on our reading, the notification and not the assessment itself. Finally, start that assessment only after establishing that the deployment as such is permitted, because Article 5 rules out certain predictive applications in criminal investigation.

Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.

Article 27(1)-(3)

education

Selection at student admission: a DPIA is not yet a FRIA

A university of applied sciences has an AI system rank applications for a vocational programme, using the exam results of earlier students to calibrate that ranking. A data protection impact assessment already exists for this processing. The question the school asks is whether that also covers the fundamental rights side of admission.

Provenance: Article 27(1) provides that deployers which are bodies governed by public law, or private entities providing public services, and deployers of high-risk AI systems referred to in points 5(b) and (c) of Annex III, shall perform an assessment of the impact on fundamental rights prior to deploying a high-risk AI system referred to in Article 6(2), with the exception of systems intended to be used in the area listed in point 2 of Annex III. That assessment consists of a description of the deployer's processes in which the system will be used in line with its intended purpose, of the period and frequency of use, of the categories of natural persons and groups likely to be affected, of the specific risks of harm to those categories, of the implementation of human oversight measures, and of the measures to be taken if those risks materialise, including the arrangements for internal governance and complaint mechanisms. Article 27(2) provides that the obligation applies to the first use, that the deployer may in similar cases rely on previously conducted impact assessments or existing assessments carried out by the provider, and that a deployer who considers during use that any element listed in paragraph 1 has changed or is no longer up to date shall take the necessary steps to update the information. Article 27(4) provides that where an obligation under this Article is already met through the data protection impact assessment conducted pursuant to Article 35 of Regulation (EU) 2016/679 or Article 27 of Directive (EU) 2016/680, the assessment under paragraph 1 complements that data protection impact assessment.

We read Article 27 as placing the education institution that runs this selection itself in the deployer role, but that alone does not settle the duty. Paragraph 1 names bodies governed by public law and private entities providing public services, and whether a state-funded or a private university of applied sciences answers to either description is the question you have to settle first. If it does, an existing data protection impact assessment is on our reading the starting point rather than the last word: paragraph 4 has the fundamental rights assessment sit alongside it, and since Regulation (EU) 2026/1744 that assessment may incorporate or cross-refer to relevant parts of it, so the real question is which elements of paragraph 1 are still missing. For you that means recording which groups of students may be affected, how the admissions committee or the teacher can correct an outcome, and where a rejected applicant can lodge a complaint. If the selection rule or the assessment component underpinning the ranking changes, that is on our reading the moment to update the record, rather than the start of the next academic year.

Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.

Article 27(1), (2) and (4)

financial services

Health insurer using AI for risk assessment: public or private makes no difference

A health insurer uses AI for risk assessment and pricing of health and life insurance. The question is whether this falls under point 5(c) of Annex III, and with that whether the Article 27 FRIA duty comes into play.

Provenance: The Commission draft guidelines of 19 May 2026 state that the health and life insurance in point 5(c) may be offered on a private or a public basis: a health insurer governed by public law also falls within it, so long as the system is intended for risk assessment or pricing with regard to natural persons. Privately serviced health insurance counts as an essential private service, even in a Member State with a public healthcare system. Unlike point 5(b), point 5(c) provides no exception for fraud detection. The document is a consultation version: non-binding and not yet final.

For the FRIA question point 5(c) counts twice: it makes the system high-risk and it makes you, as deployer, one of the parties Article 27 names. A public-law form or a public healthcare system in your Member State changes nothing there. Do not count on the fraud-detection exception from point 5(b) either: it does not apply here, and a fraud feature alongside risk assessment does not take the system out.

Draft guidelines on high-risk AI classification, 19 May 2026, annex on Annex III, paragraphs (319) to (321)

Annex I lists legislation, not products

The draft guidelines of 19 May 2026, published for consultation and expressly non-binding, clarify that Annex I AI Act does not list individual products to be classified as high-risk, but Union harmonisation legislation regulating the safety aspects of certain products. Whether an AI system falls within the scope of Annex I therefore depends on whether the system, or the product of which it is a safety component, falls within the material scope of one of the listed legislative acts. According to the draft guidelines the list in Annex I is exhaustive; products can only be added or removed by amending the scope of the harmonisation legislation itself or by adding new harmonisation legislation to Annex I. The draft guidelines also state that through Article 6(1) the AI Act does not itself extend the scope of harmonisation legislation to new or additional products, and that the AI Act does not determine or change the risk profile of a product but builds on the sectoral risk classification. Products mentioned include machinery, toys, lifts, equipment and protective systems for potentially explosive atmospheres, radio equipment, pressure equipment, recreational craft, cableway installations, appliances burning gaseous fuels, medical devices, in vitro diagnostic medical devices, and products in the automotive and aviation sectors.

Draft guidelines Annex I, points (23) to (26)

Section A and Section B of Annex I trigger different requirement sets

The Commission draft guidelines of 19 May 2026, which are non-binding as long as the final version has not been adopted, draw a distinction that is often missed in practice. AI systems classified as high-risk under Article 6(1) in respect of products covered by the harmonisation legislation in Section A of Annex I are subject to the requirements for high-risk systems in Section 2 of Chapter III AI Act. By contrast, for AI systems classified as high-risk under Article 6(1) in respect of products covered by the harmonisation legislation in Section B of Annex I, only Article 6(1), Articles 102 to 109 and Article 112 AI Act apply. The draft guidelines refer to Article 2(2) AI Act for this. Section A contains harmonisation legislation based on the New Legislative Framework, Section B the other Union harmonisation legislation.

Draft guidelines Annex I, point (60), referring to Article 2(2) AI Act

Two cumulative conditions for high-risk under Annex I

The European Commission draft guidelines of 19 May 2026, which are expressly non-binding and not final, read Article 6(1) as two cumulative conditions. First, the AI system must be intended to be used as a safety component of a product, or the AI system must itself be a product, covered by the Union harmonisation legislation listed in Annex I. Second, that product, or the AI system itself where it is the product, must be required to undergo a third-party conformity assessment. The draft guidelines state explicitly that not all AI systems that are components of regulated products are high-risk, but only the subset that satisfies both criteria.

Draft guidelines Annex I, points (27) and (21)

The Article 6(3) filter: four exhaustive grounds, to be read narrowly

According to the non-binding draft guidelines of 19 May 2026 on the classification of high-risk AI, Article 6(3) sets out four grounds on which a provider may exempt a system from high-risk classification: performing a narrow procedural task, improving the result of a previously completed human activity, detecting decision-making patterns or deviations from prior patterns without replacing or influencing the previously completed human assessment absent proper human review, and performing a preparatory task. Paragraph (88) of this draft states these grounds are exhaustive but alternative, that there is no separate independent risk test, and that they must be interpreted narrowly because Article 6(3) is an exception to rules that among other things protect fundamental rights. Paragraph (87) states the filter applies only to systems under Article 6(2) and not to systems under Article 6(1). Paragraph (89) states a system always remains high-risk where it performs profiling. Paragraph (90) adds that the filter does not apply where the system forms part of a complex system whose combined intended purpose or joint outputs materially influence an individual decision, including agentic AI. Paragraphs (113) to (116) of this draft describe that this is a self-assessment by the provider, that Article 6(4) requires documenting the assessment before placing on the market and registering in the Article 71 EU database, and that the assessment must contain at least the intended purpose, why the system falls under Article 6(2), which Article 6(3) condition applies and why, and why the system does not perform profiling. Paragraph (117) of these draft guidelines points to Articles 80 and 99 where an authority finds a system was misclassified as non high-risk to circumvent the rules.

Draft guidelines on high-risk AI classification (19 May 2026), Annex III chapter, sections 2.7, 2.7.1, 2.7.3 and 2.7.4, paragraphs (84) to (90) and (113) to (117)

prEN 18285: conformity assessment framework for AI systems

prEN 18285 (Conformity assessment framework) is the JTC 21 deliverable under M/613 covering the conformity assessment of high-risk AI systems under Article 43 and Annex VII of the AI Act. As at June 2026 the deliverable was at the drafting stage. It has not yet been published as an EN and is not cited in the Official Journal. Standardisation request M/613 was amended by Implementing Decision C(2025)3871 of 23 June 2025 and expires on 28 February 2027.