We are a public-sector organisation using AI. What needs to be in place?

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.

Artikel 27, leden 1, 3 en 5

government and public services

AI text from a municipality: when a final check counts as editorial control

A municipality has an AI system write the web pages about a changed scheme for social assistance and allowances, meant to explain to citizens what they are entitled to. A communications officer reads the text for style and spelling and publishes it. The question is whether this public service thereby falls under the exception to the labelling duty.

Provenance: The Commission guidelines of 20 July 2026 count public administration and public services among the matters of public interest covered by the disclosure duty for published AI text in Article 50(4). For the exception, the guidelines require two cumulative conditions: the AI-generated or manipulated text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for the publication. They add that human review means a deliberate examination of the substance with fact-checking as a minimum requirement, and that superficial, purely formal or procedural checks such as spell-checking or grammatical correction do not meet it. For the second condition they expect the identity or the function carrying editorial responsibility to be publicly available in an easily findable place.

Our reading is that many municipalities do have a publication process but no identifiable person carrying substantive final responsibility, and that is exactly the hinge of this exception. Assign that role explicitly, make it findable who holds it, and record per page who checked the facts, or otherwise take the simpler route and label the text. Watch the order of steps in your workflow, because an AI tool that substantively rewrites the text after human sign-off removes the ground from under that sign-off.

Commission Guidelines C(2026) 5054 final, 20.7.2026, point (131), points (133) to (136) and point (138)

government and public services

An induction call with the customer at the moment of go-live

Asimov AI is a micro organisation of at most fifteen people that supplies AI services for legislative work to government institutions and companies. With every new contract it holds one or more induction calls with the team leads and officials who will use the platform, explaining how the platform and the underlying models work and how hallucinations arise in this domain and can be mitigated.

Provenance: This practice was submitted by the organisation itself to the Commission living repository. The repository collects and shares practices; it does not approve them or set them as a standard.

This practice puts literacy where the risk arises: with the people who will operate the system, at the moment they start. For a small provider that is also the only workable moment, because there is no training department to redo it later. Anyone adopting it should record who attended and what was explained, because otherwise the effort survives only in the participants memory a year on.

Levend repository van AI-geletterdheidspraktijken, ingediende praktijk van de betrokken organisatie

government and public services

Chatbot answering factual questions from a benefits case handler

A chatbot answers a case handler's factual questions relating to the evaluation of a natural person's application for healthcare benefits, for instance the applicant's age. The case handler can grant or deny the benefits based on those answers.

Provenance: The Commission draft guidelines of 19 May 2026 address this case when determining whether an application falls under Annex III. The document is a consultation version: non-binding and not yet final.

A chatbot that only returns existing facts in structured form falls under the exemption, but once it answers case-specific legal questions it steers the decision and is high-risk.

Draft guidelines on high-risk AI classification, 19 May 2026, annex on Annex III

No mandatory course format, no certificate, no exam and no AI officer

The Commission Q&A on AI literacy states that there is no one size fits all when it comes to AI literacy and that no strict requirements or mandatory trainings are imposed. On certification, the Q&A states literally that there is no need for a certificate and that organisations can keep an internal record of trainings or other guiding initiatives. On assessment, it states that Article 4 of the AI Act does not entail an obligation to measure the AI knowledge of employees. On governance, it states that no specific governance structure is mandated to comply with Article 4, so that unlike the data protection officer under the GDPR, no AI officer needs to be appointed. On the level, the Q&A states that following the Digital Omnibus amendment AI literacy remains an obligation for providers and deployers of AI systems, but that no specific or sufficient level is mandated and that the Regulation does not require guaranteeing any specific level of AI literacy of any individual. Against that, the Q&A states that simply relying on the AI systems' instructions for use or asking staff to read them might be ineffective, and that organisations should take into account general AI understanding within the organisation, whether they are a provider or a deployer, the risks associated with the systems deployed, staff knowledge gaps considering technical knowledge, experience, education and training, and contextual factors such as sector, purpose and affected populations. The Q&A further states that organisations may implement different levels of training or learning approaches depending on knowledge, experience, education and role, and that staff with a degree or experience in AI development are normally considered AI literate, while the organisation must still verify that those persons understand the specific AI systems of the organisation, know how to deal with them and are aware of all risks.

Commission Q&A on AI literacy, sections on required level, training formats, certificates, assessment of knowledge and governance structures (consulted 9 August 2026)

Article 4 reaches beyond your own staff, and the national supervisor enforces it

The Commission Q&A on AI literacy states that Article 4 applies to providers and deployers of AI systems and in addition to other persons dealing with the operation and use of AI systems on their behalf, covering persons broadly within the organisational remit, with a contractor, a service provider and a client given as examples. On clients, the Q&A states that they may need AI literacy depending on the specific risk, reasoning that affected persons should understand how decisions taken with the assistance of AI will have an impact on them. On geographic scope, the Q&A states that the AI Act's legal framework applies to both public and private actors inside and outside the EU as long as the AI system is placed on the Union market, used in the Union, or its use has an impact on people located in the EU. On supervision, the Q&A states that the supervision and enforcement of Article 4 is not with the AI Office but under the remit of national market surveillance authorities, and that supervision and enforcement began on 2 August 2026, while Article 4 itself entered into application on 2 February 2025. On sanctions, the Q&A states that national market surveillance authorities could impose penalties and other enforcement measures for infringements of Article 4, that this will be based on national laws that Member States were due to adopt by 2 August 2025, that any sanction must be proportionate and based on the individual case taking into account factors such as the nature and gravity of the infringement and its intentional or negligent character, and that sanctions are more likely if there is proof of an incident due to a lack of appropriate training and guidance. Article 4 is not listed in the enumeration in Article 99(4) of the AI Act, which covers only Articles 16, 22, 23, 24, 26, 31, 33(1), (3) and (4), 34 and 50, so the level of any penalty for Article 4 follows from national law rather than from the Regulation's own ceilings. The Q&A further states that Article 4 reinforces the transparency provisions of Article 13 and the human oversight provisions of Article 14 and indirectly contributes to the protection of affected persons, and that for deployers of high-risk systems the Article 26 obligation to ensure staff are trained to ensure human oversight is a distinct requirement; that requirement becomes applicable on 2 December 2027 for standalone Annex III systems and on 2 August 2028 for Annex I systems.

Commission Q&A on AI literacy, sections on target groups, geographic scope, supervision and enforcement, and sanctions (consulted 9 August 2026)

AI agents must disclose both their AI nature and on whose behalf they act

Point (31) of the guidelines of 20 July 2026 states that AI agents are covered by Article 50(1) if they are capable of interacting with the persons instructing them or with other natural persons in the execution of their tasks, citing as examples making bookings, managing correspondence, negotiating or concluding contracts and executing purchases. That same point requires AI agents to be designed and developed so that they disclose both their artificial nature and the person on whose behalf they are acting, given the need for transparency of the origin and of the delegation of authority and accountability for the consequences of their actions. This also applies in complex multi-agent architectures in which other agents interact directly with natural persons. Where the provider cannot reliably determine before placing on the market or putting into service whether the agent will directly interact with a natural person, the agent should be designed at the architecture level and instructed to disclose itself in every situation where it is reasonably likely to interact with a natural person, including where that person represents a legal entity. Agents should also disclose themselves to the persons instructing them at key steps such as authorisation, reporting and validation, including where the agent receives, processes or relies upon outputs generated by other AI systems rather than by a natural person, and at every new interaction. Point (63) adds that Article 50(2) may apply to AI agents where the agent takes an action whose output is AI-generated or manipulated content perceptible by natural persons, while intermediate processing steps such as reasoning and chain of thought and non-perceptible actions such as a web request or browser action fall outside that scope.

Commission Guidelines C(2026) 5054 final, Section 3.1.1 point (31) and Section 4.1.2 point (63)

Artistic or satirical work is not exempt but attenuated, and the informative character always prevails

Point (119) of the guidelines of 20 July 2026 describes an attenuated transparency obligation for deep fakes forming part of evidently artistic, creative, satirical, fictional or analogous works or programmes, where the obligation is limited to disclosure in an appropriate manner that does not hamper the display or enjoyment of the work. Point (120) describes the categories: artistic works are created for the purpose of art, including music, cinematographic works and visual arts; creative works involve creative choices, while works mainly motivated by functional or technical considerations cannot be regarded as creative; satirical works are intended to criticise society, politics, business or public figures through humoristic techniques; fictional works involve persons, objects, places, entities or events in an imaginary but verisimilitude setting; analogous works share core traits with those categories without fitting neatly into one. Point (122) states that it must be evident to the natural persons exposed to it that the content falls within one of those categories, that the categories must therefore be interpreted strictly given the lighter disclosure regime and the interests of freedom of expression and freedom of the arts and sciences, and that content whose nature is potentially unclear or ambiguous to the audience falls outside this lighter regime. Relevant factors, per that same point, are whether the content displays formats or styles characteristic of the category, the context in which it is presented, and audience expectations. That same point excludes content whose nature is exclusively informative or commercial and recognisable as such, citing news reporting, notes that advertisements or documentaries may be regarded as evidently creative or fictional in certain specific situations but not in others because the assessment is case-specific, and states that where the deep fake combines multiple characters, for example informative and creative, the informative character should always prevail and the standard labelling requirements apply. Point (123) stresses that these deep fakes are not excluded from the obligation: the deployer must still disclose the AI origin or manipulation, but may do so in an appropriate manner, and must in any case comply with Article 50(5). Point (124) states that reliance on the attenuated obligation cannot justify failing to respect the fundamental rights of individuals or the rights of rightsholders under Union intellectual property or data protection law. As examples within the categories the document cites movies featuring AI de-aged existing actors or digital replicas of deceased actors, AI-generated music in the style of existing artists, and an AI-manipulated image of an existing politician in a scene clearly meant as humorous criticism. Outside the categories the document places among others an AI-manipulated video in the style of a teleshopping channel, AI-generated images of celebrities implying involvement in activities that never happened, and an AI-manipulated video featuring a realistic synthetic influencer focused solely on displaying a sponsored product's functionalities.

Commission Guidelines C(2026) 5054 final, Section 6.1.3, points (119) to (124) and the accompanying example lists