The Case That Was Withdrawn: How the First AI Act Referral Reached the CJEU and Left Without an Answer

A man in Sofia stopped paying his mobile phone bill. His operator sued him. In the course of that ordinary consumer dispute, his case put a question nobody had put to the Court of Justice of the European Union before: under the AI Act, does a consumer have the right to know how the algorithm that generated his invoice actually works? On 25 November 2024 the Sofia District Court referred that question — and sixteen others — to Luxembourg¹. It was widely discussed as the first reference to put the AI Act’s substantive provisions before the Court¹. On 3 June 2026 it was closed without a single one of those questions being answered — litige au principal devenu sans objet, non-lieu à statuer².

The reason it closed is the part worth your attention. The main proceedings did not fade away on their own. After the reference was notified, the operator withdrew its own claim against the consumer³. The referring judge said out loud what that looked like: a possible abuse of right, calculated to stop the Court from ruling on whether the company’s billing practices comply with EU law⁴.

Key Takeaways

• The first AI Act case ended on procedure, not substance. C-806/24 asked whether Article 86(1) gives a consumer the right to an explanation of an automated invoice, and whether a court may demand the black box data, the source code and the algorithm⁵. None of it was answered².
• The claimant withdrew after the reference was notified. The Court’s President suspended C-806/24 on 5 May 2025; the referring court said the withdrawal could amount to an abuse of right aimed at avoiding an answer³, ⁴.
• A second referral was declared manifestly inadmissible. In Yettel Bulgaria II (C-502/25, order of 18 December 2025) the Court held that the fact an answer would be useful in many parallel cases has no bearing on admissibility⁶.
• Article 86 became applicable on 2 August 2026 — into an uncertain landscape. The Omnibus did not expressly defer Article 86, but it moved the Annex III obligations to 2 December 2027, leaving open how much Article 86 can bite in the interim⁷.
• The enforceable route today is the GDPR, not the AI Act. Article 15(1)(h) GDPR has applied since 2018 and the Court has already ruled on what “meaningful information about the logic involved” means, and on how to handle a trade-secret objection⁸.

What the Bulgarian court actually asked

The facts are mundane, which is precisely why they matter. Yettel Bulgaria sued a customer for unpaid subscription fees, for usage outside the bundle, and for a termination penalty calculated on the standard monthly fee rather than the promotional fee the parties had agreed. Invoicing ran automatically through an internal system. During the proceedings an expert examined it and found discrepancies in how calls and data volumes were counted — apparently undisclosed rounding rules and unit conversions. The consumer’s position was that he had never been told the parameters, the logic, or the purpose of those calculations¹.

Out of that, the referring court built seventeen questions. Three carry the weight.

Question 1 asks whether Article 86(1) of the AI Act means a consumer may require the trader to disclose how and with what elements and parameters the automated invoices were generated, and — read with Article 38 of the Charter — what algorithm calculates them. It closes with the sharpest formulation of all: must Article 86(1) be interpreted as applying to consumer contracts at all?⁵

Question 4 goes further than any deployer would like. Read with Articles 47 and 38 of the Charter and the effectiveness principle, does Article 86(1) permit the court to demand from the trader “the black box data, the source code and the algorithm relating to the way in which automated decisions are made under the consumer contract”?⁵

Question 5 asks whether such an automated decision is subject to review by a human being — specifically, by a judge in real judicial proceedings⁵.

Set aside the drafting. What the Bulgarian judge was reaching for is the question every DPO and CISO eventually meets: when a system produces a number that costs someone money, how much of the machine has to be opened, and who gets to look inside?

Article 86: a right with a narrow gate

It is worth being exact about what Article 86 does and does not promise, because it is routinely oversold.

Article 86 gives an affected person 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. But the gate is narrow. It applies only to decisions taken on the basis of output from a high-risk AI system listed in Annex III — excluding point 2 — which produce legal effects or similarly significantly affect the person in a way they consider to have an adverse impact on their health, safety or fundamental rights. It also applies only to the extent the right is not already provided for elsewhere in Union law⁹.

Three consequences follow, and they are why the Yettel questions were always going to be hard.

First, an automated billing engine is not obviously an Annex III high-risk system. Nothing in Annex III covers telecoms invoicing. The referring court’s premise — that Article 86 reaches ordinary consumer contracts — was itself one of the questions, and it was the one most likely to be answered “no.”

Second, the right is a right to an explanation, not to the artefact. The wording is “the role of the AI system in the decision-making procedure and the main elements of the decision taken.” That is a considerable distance from source code. Question 4 was asking the Court to bridge that distance using Article 47 of the Charter — a serious argument, and precisely the kind of argument that only a judgment can settle.

Third, and most simply: on 25 November 2024, when the reference was lodged, Article 86 was not yet applicable. Under the original Article 113 timeline it was to apply from 2 August 2026¹⁰. A referring court may ask about a provision that is in force but not yet applicable; whether the Court will engage is another matter.

How the case ended, and the second attempt

After the reference was notified, Yettel filed a declaration withdrawing its action for a declaration that the debt existed. By decision of 5 May 2025 the President of the Court suspended the proceedings in C-806/24³.

The Sofia court did not accept this quietly. It made a second reference — registered as C-502/25, Yettel Bulgaria II — expressly intended to complete the first. In it, the court recorded that the withdrawal was said to be motivated by the cost of enforcing the debt no longer being worth it, and then stated its own reading: the withdrawal could constitute an abuse of right, reflecting Yettel’s wish to end the main proceedings so that the questions in C-806/24 — questions designed to establish whether its telecoms billing practices are contrary to EU law — would go unanswered. The court added that the Court’s answer would apply equally to all the other cases Yettel brings before it, which it does in very large numbers each year⁴.

On 18 December 2025 the Court dismissed the second reference as manifestly inadmissible under Articles 53(2) and 94 of the Rules of Procedure: the referring decision did not adequately state the reasons why an answer was necessary or the link between the EU provisions invoked and the national law applicable⁶. On the substance of the complaint, the Court was terse. The fact that its answer might be relevant in other proceedings of the same type “is without bearing” on the admissibility of the present request; the existence of other national cases in which the answer would be useful cannot justify the Court answering additional questions in the same main proceedings⁶. It did add that the referring court remains free to submit a fresh reference containing everything the Court needs⁶.

Then, on 3 June 2026, C-806/24 itself was disposed of, the main proceedings having ceased to exist. No substantive ruling².

What this actually leaves you with

It is tempting to read this as a story about one telecoms company. It is better read as a story about an enforcement architecture.

Article 86 has no supervisory backstop of its own. It is a right an individual asserts against a deployer, and in practice it will be asserted inside private litigation the individual usually did not start. The AI Act does provide a complaint route: under Article 85, any natural or legal person having grounds to consider that there has been an infringement may complain to the relevant market surveillance authority¹¹. That route matters — but it presupposes an obligation already applicable, and it has none of the institutional depth a data protection authority has built over eight years. Where a right depends on a single consumer dispute reaching Luxembourg, the counterparty can end the dispute for the price of one unpaid phone bill. That is not a hypothetical; it is what the record shows happened here.

And the ground under the right has shifted. Article 86 applies from 2 August 2026. The Digital Omnibus, published in the Official Journal on 24 July 2026 and in force since 27 July 2026, did not expressly defer Article 86 itself; what it deferred was the stand-alone Annex III high-risk regime, from 2 August 2026 to 2 December 2027⁷. Since Article 86 only engages where a decision rests on the output of an Annex III high-risk system, a right that has now formally switched on will in practice have very little to attach to for another sixteen months. That is an inference from the structure of the two provisions rather than a stated amendment — but it is the inference a compliance programme has to plan around. The opposite reading — that Article 86 attaches from 2 August 2026 to any decision resting on a system falling within an Annex III category, deferral or not — remains arguable until the amended Article 113 is judicially construed; a deployer of an Annex III-category system should not treat the right as dormant. Either way, anyone waiting for judicial clarity on explainability under the AI Act should plan for a long wait.

Meanwhile the enforceable right is the one you already have. Article 15(1)(h) GDPR entitles a data subject to meaningful information about the logic involved in automated decision-making, and unlike Article 86 it has been litigated. In Dun & Bradstreet the Court held that the controller must explain the procedure and principles actually applied, that handing over a complex mathematical formula or an exhaustive description of every step does not satisfy the obligation, and that where a trade secret is invoked the protected material should be put before the competent supervisory authority or court to balance the competing interests — not simply withheld⁸. That is a working answer to the trade-secrets question the Bulgarian court was still asking under the AI Act.

The practical instruction for a DACH deployer is therefore not “wait and see.” Build the explanation now, in a form that satisfies Article 15(1)(h) today and will satisfy Article 86 when it arrives: what the system does in the decision, which inputs it used, which principles it applied, expressed so an affected person can understand it — and document the reasoning wherever you decline to go further on trade-secret grounds, because that reasoning is what a court will weigh. If you cannot produce that document for your own automated processes, the deferral has not helped you; it has only postponed the day you find out.

The right to an explanation is only as strong as the procedure that carries it. In Yettel Bulgaria, the procedure did not survive contact with a claimant who preferred not to be answered.

The next candidate is already pending. In Rowicz (C-159/25), a Polish referral asks whether a court can be considered independent and impartial when the algorithm that randomly allocates cases to judges cannot be inspected — neither its source code nor any means of verifying its operation or its vulnerability to error and manipulation — invoking Recital 61 of the AI Act and Article 47 of the Charter¹². That one puts the court system itself in issue, not one company’s invoice — a harder case to make disappear.

Which of your automated decisions could you explain to the person on the receiving end — not in code, but in a paragraph they would accept?