Skip to main content
Praxikon

The right to explanation under Article 86 of the EU AI Act: what a municipality must explain

··23 min read

Yes: where a municipality takes a decision about a resident on the basis of the output of a high-risk AI system listed in Annex III, and that decision produces legal effects or similarly significantly affects the resident, the resident has the right to clear and meaningful explanations of the role of the AI system in the decision-making procedure and of the main elements of the decision (Article 86(1) AI Act). The right is supplementary: it applies only to the extent that other Union law does not already provide for it (paragraph 3), and the GDPR already does so for part of these decisions. From when a resident can actually rely on it is less simple than the text suggests. More on that below.

Picture this: a municipality has a model give every application for social assistance a risk score. One application scores high, a caseworker opens a re-examination, finds that the stated address does not add up and rejects the application. The applicant writes back: "What role did that computer play in my rejection?" Does the municipality have to explain, and what should the explanation say?

In short

Article 86 AI Act gives every affected person a right to an explanation from the deployer, here the municipality, not from the vendor. Four conditions: a decision taken by the deployer, on the basis of the output of a high-risk AI system listed in Annex III (point 2, critical infrastructure, is excluded), producing legal effects or a similarly significant effect, which the person considers to have an adverse impact on their health, safety or fundamental rights. Unlike Article 22 GDPR, Article 86 does not require the decision to be based solely on automated processing: the right can apply even when a caseworker decides. The explanation covers the role of the AI system in the decision-making procedure and the main elements of the decision. Article 86 sits in Chapter IX and falls under the general rule of Article 113 (applicable since 2 August 2026), but it hinges on the high-risk classification, which for Annex III only applies from 2 December 2027, with transitional rules for systems already in use. Where a decision falls under Article 22 GDPR, the GDPR itself grants a right to an explanation (CJEU, Dun & Bradstreet Austria, C-203/22) and, in our reading, Article 86 steps back to the extent that right reaches. In our reading the GDPR then mainly covers the logic of the system; for the other elements of the decision Article 86 remains relevant.

What Article 86 says

The provision has three paragraphs. Paragraph 1 grants the right. Any affected person subject to a decision taken by the deployer on the basis of the output from a high-risk AI system listed in Annex III, with the exception of the systems listed under point 2 thereof, which produces legal effects or similarly significantly affects that person in a way that they consider to have an adverse impact on their health, safety or fundamental rights, has the right to obtain from the deployer "clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken".

Paragraph 2 sets the right aside where Union law, or national law in compliance with Union law, provides for exceptions or restrictions. Paragraph 3 makes it supplementary: the Article "shall apply only to the extent that the right referred to in paragraph 1 is not otherwise provided for under Union law". Recital 171 states the purpose: the explanation should be clear and meaningful and should provide a basis on which affected persons are able to exercise their rights.

The Article sits in Chapter IX, Section 4 (remedies), right after the right to lodge a complaint with a market surveillance authority (Article 85). It is a right of the individual, not a design requirement for the vendor. Who built the system does not matter for Article 86: the request goes to the organisation that takes the decision.

Who holds the right

The "affected person" is whoever is subject to the decision. Article 3 does not define the term, and Article 86 does not tie the right to personal data. That is a difference with the GDPR, which revolves around the data subject whose data are processed. In the social assistance example both are usually the same person, but not always. Think of a household member who counts in a household assessment: in our reading that person can be affected too where the decision also applies to them, as with assistance granted to the household.

The obligation rests on the deployer. If the municipality uses a vendor's system, the municipality must give the explanation. It does need information from the vendor to do so, and that information belongs in the instructions for use (see below).

When the right applies: four conditions

1. A decision taken by the deployer. This is a decision the municipality takes, not the score itself. In the example, the rejection of the application is the decision.

2. On the basis of the output of a high-risk AI system listed in Annex III. A system 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, or to grant, reduce, revoke or reclaim such benefits, is listed in Annex III, point 5(a). An Annex III system is not high-risk where it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision making, and it meets one of the conditions of Article 6(3). That way out is always closed when the system performs profiling of natural persons. A risk score that predicts per applicant how likely incorrect statements are is, in our reading, profiling (Article 4(4) GDPR), so the way out is closed here. Point 2 of Annex III, critical infrastructure, is expressly excluded from Article 86.

Note a difference in wording. The Article says "on the basis of the output", while recital 171 refers to a decision that is "based mainly" on that output. In our reading the binding text is the Article and the recital is an aid to interpretation: the larger the role of the score, the more clearly the right applies. How far "on the basis of" reaches when a score was just one of many signals has not been decided by the Court. The Court has not yet ruled on Article 86. A reference is pending (C-245/25, DZI Obshto Zastrahovane), including on what Article 86(1) requires where a human verifies an algorithmic outcome, and on the temporal application of Article 113.

3. Legal effects or a similarly significant effect. Rejecting an application for social assistance produces legal effects. The re-examination itself is less straightforward: opening an investigation is usually not yet a decision with legal effects, but it can significantly affect a resident, for example where benefits are suspended in the meantime. In our reading the municipality is wise to treat the re-examination as part of the decision-making and to cover the role of the score in it when explaining the final decision.

4. Adverse in the eyes of the person concerned. The Article refers to effects that the person "consider[s]" to have an adverse impact on their health, safety or fundamental rights. That assessment lies with the affected person. In the Dutch language version that own assessment grammatically belongs to the second limb (the similarly significant effect), while the English text leaves more open; for the social assistance example it makes no difference, because a rejection produces legal effects. In our reading the municipality therefore cannot refuse a request by arguing that, in its view, the decision was not adverse.

What the Article does not say matters just as much: Article 86 does not require the decision to be based solely on automated processing. That is the key difference with Article 22 GDPR. A caseworker who genuinely decides on the merits, with the score as an aid, takes the decision outside Article 22 GDPR, but not outside Article 86. When signing off is not a genuine decision of one's own is discussed in Article 22 GDPR and rubber-stamping an algorithm.

What the explanation must contain

Article 86 names two subjects: the role of the AI system in the decision-making procedure, and the main elements of the decision taken. The explanation must be clear and meaningful. The Article says nothing about form or time limit.

What "meaningful" means has not yet been settled by the Court for Article 86. For the GDPR it has. In Dun & Bradstreet Austria the Court held that, for an automated decision, the data subject has a genuine right to an explanation as to the functioning of the mechanism involved and of the result (paragraph 57). Neither the mere communication of a complex mathematical formula nor a detailed description of all the steps will do (paragraph 59). The explanation must allow the person to understand which personal data have been used in what way (paragraph 61), and according to the Court a national court could find the explanation sufficiently transparent and intelligible where the person is told to what extent a variation in the data taken into account would have led to a different result (paragraph 62). That judgment concerns the GDPR, not Article 86. In our reading it is nonetheless the best yardstick available, because both rules serve the same purpose: enabling the person to understand the decision and to challenge it.

On that basis, a usable explanation answers six questions in our reading:

  1. Which system. Which AI system was used, for what purpose, and that it falls under Annex III.
  2. What the system said in this case. The output for this person: the score, label or signal, and what that result means.
  3. What the result mainly rested on. The data or characteristics that weighed most, to the extent the system can show them.
  4. What the human did with it. Who assessed the output, what that person checked independently, and whether they could and were allowed to disregard the output.
  5. The other elements of the decision. The facts and legal conditions on which the decision otherwise rests, independently of the system.
  6. What the person can do now. How to give their view, have data corrected and lodge an objection.

In the social assistance example that reads roughly like this: the model gave your application a high score, mainly because of a mismatch between your stated living situation and registry data; a re-examination was opened for that reason; the caseworker checked the address independently and the rejection rests on that investigation and on the legal conditions, not on the score alone; you can object and explain your situation. If the municipality cannot fill in points 3 and 4, that is a sign that the system or the work process is not built to be explained. We wrote earlier about explainability as a technical question in AI's mystery box.

From when: three readings

This is where you need to read closely. The second paragraph of Article 113 says the Regulation applies from 2 August 2026. The third paragraph lists exceptions, and Chapter IX is not among them. Formally, then, Article 86 has applied since that date. But Article 86 hinges on "a high-risk AI system listed in Annex III", and the classification rules of Article 6 sit in Chapter III, Section 1. Since Regulation (EU) 2026/1744, that Section applies to systems under Article 6(2) and Annex III only from 2 December 2027 (Article 113, third paragraph, point (c)(i)). On top of that comes the transitional rule in Article 111(2): for high-risk AI systems placed on the market or put into service before the date of application of Chapter III, the Regulation applies to operators only if, as from that date, those systems are subject to significant changes in their designs; providers and deployers of systems intended to be used by public authorities must comply by 2 August 2030.

That leaves three readings.

  • Reading 1: since 2 August 2026. Article 86 falls under the general rule, and a system that appears in Annex III is "listed in Annex III" even while the high-risk obligations do not yet apply.
  • Reading 2: in practice from 2 December 2027. Whether a system is a high-risk AI system follows from Article 6, and that Article only applies to Annex III from that date. Until then there is no high-risk Annex III system for the right to attach to.
  • Reading 3: later for systems already running. Article 111(2) speaks of "this Regulation", not only of Chapter III, and the duty to explain under Article 86 rests on an operator. If the social assistance model is already running before 2 December 2027, under this reading Article 86 applies only after a significant change in its design and, for a system intended to be used by public authorities, by 2 August 2030 at the latest.

We consider reading 2 the most defensible: Article 86 requires a system that is high-risk, and that classification only operates for Annex III from 2 December 2027. For systems already running before that date, the broad wording of Article 111(2) ('this Regulation', 'operators') points to reading 3. The pending case C-245/25 asks the Court, among other things, about the temporal application of Article 113. In practice this matters less than it seems: the GDPR applies now, so does the duty under Dutch administrative law to state reasons for a decision, and a work process that can explain the role of a score is not built in a month. So do not start from the date, start from the question of whether you could give the explanation today.

Article 86 alongside Article 22 and Article 15(1)(h) GDPR

The GDPR does not have a general right to an explanation for every decision, but it does have one for automated decision-making. Article 22(1) GDPR gives the data subject the right not to be subject to a decision based solely on automated processing which produces legal effects or similarly significantly affects them. In SCHUFA (C-634/21) the Court held that the automated establishment of a score by a credit information agency is itself such a decision where a third party draws strongly on that score to establish, implement or terminate a contractual relationship with the person. Whether a caseworker who signs off on a score takes the decision outside Article 22 GDPR is something we discussed in Article 22 GDPR and rubber-stamping an algorithm; for agents, see AI agents and Article 22 GDPR.

Article 15(1)(h) GDPR grants, for automated decision-making referred to in Article 22 GDPR, a right to meaningful information about the logic involved, as well as the significance and the envisaged consequences. In Dun & Bradstreet Austria (C-203/22, 27 February 2025) the Court turned that into the genuine right to an explanation described above. The Court added that invoking trade secrets or third-party data does not simply block the explanation: the controller must then provide that information to the supervisory authority or the court, which balances the interests (paragraph 74 and operative part, point 2).

How do the two regimes relate? Article 86(3) gives the order.

  • If the decision falls under Article 22 GDPR, Union law already provides for a right to an explanation through Article 15(1)(h) GDPR, and in our reading Article 86 steps back to the extent that right reaches. In our reading the GDPR then mainly covers the logic of the system; for the other elements of the decision Article 86 remains relevant. Note: a municipality relying on a legal basis for such a decision (Article 22(2)(b) GDPR) needs a law that itself lays down suitable measures; the right to human intervention in Article 22(3) GDPR applies only to decisions based on a contract or explicit consent.
  • If the decision falls outside Article 22 GDPR, because a caseworker genuinely decides, the GDPR explanation from Dun & Bradstreet is not automatically available: the Court gave that explanation for decisions within the meaning of Article 22(1) GDPR. This is where Article 86 does its real work. Genuine human involvement does not remove the right to an explanation; it moves it to the AI Act.
  • Dutch administrative law still applies. Article 3:46 of the Dutch General Administrative Law Act (Awb) requires a decision to be based on sound reasoning. Article 86(3) only looks at Union law, so in our reading the Awb does not set Article 86 aside. Paragraph 2 concerns exceptions and restrictions, not a parallel duty to give reasons. In practice you can meet both duties in one good statement of reasons, provided it genuinely describes the role of the system.

Two differences to remember. The GDPR sets a time limit: within one month, extendable by two further months where necessary (Article 12(3) GDPR). Article 86 sets none. And the GDPR requires personal data; Article 86 does not.

A short note on banks and credit scoring

A bank faces the same question. An AI system intended to evaluate the creditworthiness of natural persons or establish their credit score is listed in Annex III, point 5(b), with the exception of systems used to detect financial fraud. As a deployer, the bank must also carry out a fundamental rights impact assessment from 2 December 2027 (or later under Article 111(2)) (Article 27(1)), and under SCHUFA the score itself can fall under Article 22 GDPR.

For consumer credit there is also a specific right under Union law. The Consumer Credit Directive, Directive (EU) 2023/2225, provides in Article 18(8) of the Directive that, where the creditworthiness assessment involves automated processing of personal data, the consumer has the right to request and obtain human intervention, including a clear and comprehensible explanation of the assessment, covering the logic and risks involved in the automated processing and its significance and effects on the decision. Paragraph 9 requires the creditor, when rejecting an application, to inform the consumer without delay, including, where applicable, that the assessment is based on automated processing and how to contest the decision. Member States apply those rules from 20 November 2026 (Article 48 of the Directive). To the extent that right reaches, Article 86 steps back in our reading under paragraph 3. One open point: a directive works through national law, and whether "Union law" in paragraph 3 also covers that transposed law has not been decided. What the EBA says about this overlap is covered in the EBA mapping for the financial sector.

What to record: the practice

An explanation after the fact is only as good as what you recorded at the time of the decision. Five things make the difference.

The provider's instructions for use. Where applicable, the provider must describe in the instructions for use the technical capabilities and characteristics of the system to provide information relevant to explain its output, and information to enable deployers to interpret the output and use it appropriately (Article 13(3)(b)(iv) and (vii)). That is the material for points 2 and 3 of the explanation. Ask for it at procurement and agree in the contract that the vendor can deliver the main factors per decision. More in our piece on instructions for use under Article 13.

Logs. High-risk AI systems must technically allow for the automatic recording of events (Article 12), and the deployer keeps the logs under its control for a period appropriate to the intended purpose, of at least six months, unless provided otherwise in applicable Union or national law (Article 26(6)). Logs are evidence, not an explanation. In our reading the retention period should match the period in which an objection or appeal can still run, within the limits of the GDPR.

The decision file. Record per decision: which system and which version, the output for this person, what the caseworker saw, what they checked independently, whether they deviated and why, and the reasoning for the final decision. That is the material for points 4 and 5.

Human oversight worth the name. The provider must deliver the system in such a way that the person overseeing it remains aware of the tendency to over-rely on the output, can correctly interpret the output and can decide in any particular situation to disregard or override it (Article 14(4)(b), (c) and (d)). From 2 December 2027 (or later under Article 111(2)) the municipality assigns oversight to people with the necessary competence, training and authority (Article 26(2)). A caseworker who cannot interpret the score cannot explain it either.

Inform in advance, one entry point afterwards. Deployers of Annex III systems that make or assist in making decisions related to natural persons inform those persons, from 2 December 2027 (or later under Article 111(2)), that they are subject to the use of the system (Article 26(11)). Then set up one route for questions: a resident does not distinguish between an access request, a request under Article 86 and an objection, and does not need to. For the DPIA and FRIA on fraud signals in social assistance, see AI fraud risk scoring in welfare.

If a resident gets no explanation, they can lodge a complaint with a market surveillance authority (Article 85). In the Netherlands no such authority has yet been designated for the AI Act; that is for the Dutch AI Act Implementation Act; the government plans to submit the bill to the House of Representatives in the second quarter of 2027 (Planning letter on digital affairs, Parliamentary Paper 26643, no. 1554). The Dutch Data Protection Authority can already act under the GDPR.

What to do now

  1. Map which systems you use that fall under Annex III and that weigh in on decisions about residents.
  2. Establish per system whether the decisions fall under Article 22 GDPR or whether a caseworker genuinely decides. That determines which explanation route takes precedence.
  3. Write out the explanation for one real decision using the six questions above. If that fails, you know what is missing.
  4. Ask the vendor for the information in Article 13(3)(b)(iv) and (vii), and agree that it can deliver the main factors per decision.
  5. Set up the decision file and the log retention period so that you can still support an explanation after an objection procedure.
  6. Train the staff who assess the output to interpret, and where needed disregard, the score.

Unsure whether your use case falls under Annex III and which duties then apply? With the case check you describe the use case in a few questions and see which AI Act and GDPR duties come into view.

Frequently asked questions about the right to explanation

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Official Journal, recital 171 (12 July 2024, accessed 17 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 4, 12, 15 and 22 (27 April 2016, accessed 17 November 2026)
Court of Justice of the EU: Case overview C-203/22, Dun & Bradstreet Austria (accessed 17 November 2026)
EUR-Lex: Directive (EU) 2023/2225 on credit agreements for consumers, Articles 18 and 48 (18 October 2023, accessed 17 November 2026)
Overheid.nl: Dutch General Administrative Law Act (Algemene wet bestuursrecht), Article 3:46 (in force from 10 November 2026, accessed 17 November 2026)
State Secretary for Economic Affairs and Climate Policy: Planningsbrief Digitale Zaken, Parliamentary Paper 26643, no. 1554 (31 August 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

What is social scoring? The EU AI Act ban explained, and which scores are still allowed

Social scoring means rating people on their behaviour or personal characteristics, with a score that leads to worse treatment. The EU AI Act has banned it in two situations since 2 February 2025, for public authorities and companies alike (Article 5(1)(c)). What exactly is prohibited, what falls outside the ban, examples at a municipality, a bank and an employer, and what the GDPR adds.

EU AI ActSocial scoringArticle 5

Logs of high-risk AI systems: how long to keep them under the EU AI Act, and does the GDPR allow it?

From 2 December 2027 (for systems already in use, by 2 August 2030 at the latest), a municipality using a high-risk AI system keeps the automatically generated logs under its control for a period appropriate to the intended purpose, of at least six months (Article 26(6) AI Act). In principle the GDPR does not prohibit this, but it does decide how much longer is allowed and how the logs must be secured. What Articles 12, 19 and 26 require, what the logs must record, and how to justify the retention period, with a municipality and a bank as examples.

EU AI ActLoggingRecord-keeping