Case C-245/25, DZI – Obshto Zastrahovane – the expert, the algorithm and the black box
Courts decide on expert evidence they cannot themselves produce. That has always required a measure of trust. This reference asks what is left of that arrangement when the expert’s own reasoning was generated by software he did not write and cannot fully explain.
Facts
The underlying dispute is unremarkable: one Bulgarian insurer, DZI – Obshto Zastrahovane, against another, ZK Lev Ins, before the Sofiyski rayonen sad (Sofia District Court). What brought it to Luxembourg is the evidence. The expert’s report on which the court would base its final decision was produced with the assistance of a software program that generates automated outcomes or uses elements of artificial intelligence — software not intended for the Union market, which the expert adapted to European requirements and standards, and whose calculations he verified against his own professional knowledge. The referring court has asked ten questions, running from the classification of that software under the Artificial Intelligence Act to whether the Regulation can be invoked at all in proceedings that began before it applied.
Questions Referred
According to the Official Journal notice, the Sofiyski rayonen sad asks:
1. Is the term ‘high-risk AI system’ within the meaning of point 8 of Annex III to Regulation No 2024/1689 to be interpreted as including a software program that generates automated outcomes or uses elements of artificial intelligence and assists the expert in producing an expert’s report, on the basis of which the court will deliver its final judicial decision?
2. Is Article 101 TFEU to be interpreted as meaning that the court may rely on an expert’s report that has been produced with the assistance of a software product which is not intended for the Community market but which the expert has adapted to the requirements and standards of the European Community?
3. Are Article 86(1) of Regulation (EC) No 2024/1689 and the principle of human oversight/verification to be interpreted as permitting a national court to use, for the purposes of delivering the final judicial decision, an expert’s report produced by an expert but with the assistance of artificial intelligence or an algorithm for automated outcome generation, only if the expert has manually verified the outcome generated by the artificial intelligence or algorithm, taking into account the facts and their likely existence in objective reality on the basis of his or her knowledge of science and technology and his or her professional experience?
4. Is the right to a fair and independent tribunal provided for in Article 19 TEU and the principle of transparency in relation to automated decision-making under Article 86(1) of Regulation No 2024/1689 to be interpreted as permitting a national court to rely on an expert’s report that has been produced with the assistance of artificial intelligence or an automatedly generated outcome, provided that the expert’s report has been verified by the expert on the basis of his or her professional experience and knowledge?
5. Is the right to a fair and independent tribunal provided for in Article 19 TEU and the principle of traceability in relation to automated decision-making under Article 86(1) of Regulation (EC) No 2024/1689 to be interpreted as meaning that the production of an expert’s report may be assisted by AI or an automatedly generated outcome, which is not consistent with European standards, provided that the expert has taken into account the established facts and the evidence gathered and adapted these [tools] to European standards when producing the expert’s report?
6. Must the right to human oversight of important decisions, the principle of traceability and the principle of explainability under Article 86(1) of Regulation No 2024/1689 be interpreted as meaning that these principles are satisfied if the expert has reviewed the automatedly generated decision and amended the default settings of the software by adapting them to the established facts and to his or her professional knowledge and experience?
7. Are the principle of explainability and the principle of traceability under Article 86(1) of Regulation No 2024/1689 in relation to the black box effect to be interpreted as requiring the expert, when producing an expert’s report assisted by AI or an automatedly generated outcome, to provide a full account of the algorithm for all the calculations and assumptions made, including elementary calculations which are used as substitutes for formulas and which do not provide any concrete information on the subject matter of the dispute?
8. Is Article 86(1) of Regulation No 2024/1689 to be interpreted as meaning that the right to an explanation is satisfied if the expert has verified the outcome generated by AI or an algorithm for automated outcome generation on the basis of his or her professional skills and experience?
9. Must Article 267 of the Treaty on the Functioning of the European Union, the principle of legal certainty and the principle of the obligation not to act be interpreted as meaning that a national court may make a reference for a preliminary ruling regarding the interpretation of an EU regulation which has entered into force but the provisions of which are not yet applicable, where the legal effects of that regulation are relevant to the resolution of a case pending before that court?
10. Must Article 113 of Regulation (EU) 2024/1689 … and the principle of sincere cooperation be interpreted as meaning that that regulation applies ratione temporis (by reason of time) to legal relationships that arise before the date on which it becomes applicable but that continue to produce legal effects after that date?
Sources
OJ notice C/2026/276 (EUR‑Lex) · Case file on CURIA · Regulation (EU) 2024/1689 (Artificial Intelligence Act)
Comment
Strip away the repetition and the ten questions are three. Does the Act reach software that helps an expert rather than a judge? How much explanation does Article 86(1) actually require? And can any of this be asked yet?
The classification question is narrower than it looks. Point 8 of Annex III captures AI systems “intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts”. A court-appointed expert is not a judicial authority, and his software was bought to do actuarial arithmetic, not to interpret law. On a literal reading the answer is no. But recital 61 explains the classification as addressing “the risks of potential biases, errors and opacity”, and it is hard to see why those risks change according to which participant in the proceedings runs the software, when the output lands in the same judgment either way. The referring court’s phrasing — “on the basis of which the court will deliver its final judicial decision” — is aimed squarely at that gap.
Here the Court’s data-protection case law offers a usable model. Case C‑634/21, SCHUFA (ECLI:EU:C:2023:957) concerned a credit agency that produced only a probability value; the actual decision to lend was taken by a bank. The Court refused to let that formal division of labour decide the matter, holding that the automated establishment of the probability value is itself “automated individual decision-making” where the third party receiving it “draws strongly on” it. Transpose that reasoning and the question is not whether the expert or the court is the formal author, but whether the court draws strongly on an output neither of them generated. On that test a human signature at the bottom of the report proves very little.
Which is exactly what questions 6, 7 and 8 circle. They offer the Court a menu of increasingly thin versions of human oversight: the expert verified the result against his experience; the expert adjusted the software’s default settings; the expert can explain the outcome but not the algorithm. Question 7 states the difficulty without disguise — must the expert give “a full account of the algorithm for all the calculations and assumptions made”? A yes makes proprietary software effectively unusable in court, since no vendor will disclose that. A no means explainability extends exactly as far as the licence agreement permits. The Court will have to find something between the two, and whatever it finds becomes the working standard for expert evidence across the Union.
Questions 9 and 10 look procedural and are anything but. The AI Act entered into force in August 2024 with obligations phased in over years; the referring court wants to know whether it may seek an interpretation of provisions not yet applicable, and whether the Regulation reaches relationships formed earlier that still produce effects. An affirmative answer would let national courts pull EU instruments forward into disputes the legislature had not yet reached — a significant thing for any regulation with a long transition, of which the AI Act is only the most prominent. Bulgaria’s courts, having already produced the Union’s first AI Act reference, are proving unusually quick to test what the new Regulation will actually mean.