Order No. 59-P of the Chairman of the Supreme Court of the Russian Federation, dated 3 September 2026, approves a Concept that plans AI-generated draft judicial acts, access for all courts to court AI services by the second quarter of 2027 and a 2030 target of more than 95 percent of judges using those services, while stating that the final decision always remains with a human

What the order commands, and who is on the receiving end

Five numbered paragraphs, two pages. Paragraph 1 approves the attached Concept. Paragraph 2 directs the Supreme Court's Department for the Systematisation of Legislation and Analysis of Judicial Practice, jointly with the Department of Informatisation and Communications and in interaction with the Judicial Department at the Supreme Court, to organise the measures the Concept provides for and to keep the Concept, its calendar plan, its targets and its risks up to date. Paragraph 3 directs the V. M. Lebedev Russian State University of Justice to take the Concept's provisions into account in its educational activity, including professional development for judges and court civil servants.

Paragraph 4 orders publication in the Bulletin of the Supreme Court, posting on the court's official website and communication to lower courts. Paragraph 5 keeps control of execution with the Chairman. The signature block reads I. V. Krasnov, Chairman.

That is the whole of the operative order, and its addressees are all inside the court system. It is not a resolution of the Plenum, it does not touch the civil, arbitrazh, criminal or administrative procedure codes, and it says nothing to a party or a representative. Where the document reaches litigants at all, it does so through the Concept's plans and its risk matrix, and those are programme text. We return to that distinction in every section below because it is the one a reader is most likely to lose.

A word on timing, since this is dated back-coverage. The order carries the date 3 September 2026 on its first page. We read it on the Supreme Court's website in the week of 15 to 18 September after a ConsultantPlus notice that carries a site publication date of 15 September 2026. That date belongs to the ConsultantPlus notice. We have no evidence of when the court first posted the file, and we do not claim the order was unavailable before 15 September.

The principle the Concept keeps returning to: assistant, not replacement

Section II paragraph 4 states the purpose: to raise the capacity of the judicial system and the efficiency of the courts with modern technological solutions, drawing on best world practice with regard to domestic specifics, and, in our translation of the closing words, with the unconditional preservation of the key role of the judge in decision-making.

Paragraph 5 then lists eight principles. The second is the one in the pull quote. Our translation of that passage, from section II paragraph 5(2) on physical page 4: AI is an instrument of assistance to the judge and not a replacement for the judge; the final decision always remains with a human. The PDF wraps the passage across three lines; we joined them with single spaces and changed nothing else.

The other seven of the Concept's programme principles, in our paraphrase, are protection of human rights and freedoms including non-discrimination in the design of solutions; legality; transparency, explainability and controllability, under which algorithms are to be understandable, verifiable and where necessary auditable and every interaction with an AI system is to be logged; quality and reliability, with source data drawn from official publication sources and outputs verified; public order and ethics; confidentiality and personal data protection; and technological independence, meaning preferential use of domestic AI technologies and solutions.

Paragraph 6 sets five tasks. Four are about access, transparency, judicial capacity and workload. The fifth is reducing the number of errors connected with the human factor. The text says human factor generally and does not single out judges; our reading, marked as interpretation, is that section II otherwise casts the human as the safeguard and here casts human error as something to reduce.

Seven work areas, and two of them reach the judgment itself

Section III paragraph 7 describes end-to-end application across the life cycle of a case, from filing to enforcement and review by higher instances. Paragraph 8 then sets out seven areas, each with a list of functionality and a list of expected results. These are plans. Nothing in the text says any of them is running anywhere, and section IV paragraph 10 makes every solution pass four stages before it is scaled: audit of the process, piloting, evaluation of pilot results with a decision on scaling, and scaling itself.

The intake area at 8.1 includes automatic registration, completeness and jurisdiction checks, automatic identification of grounds for leaving a filing without movement, returning it or refusing it, a state-fee calculator, and a preliminary risk assessment for the applicant. The text qualifies that last item in its own parenthesis: a probable outcome for the typical category, without predicting the decision in the specific case, with an explanation of the reasons for the assessment. Area 8.2 covers classification by category and complexity and the distribution of cases among judges taking account of workload and specialisation. Area 8.3 covers hearing support: automatic transcription from speech recognition of all participants with speaker identification, real-time translation, and real-time provision to the judge of relevant norms, practice and case materials. Every item in this paragraph is functionality the Concept plans for under a heading the document itself labels functionality and expected results; the text reports none of it as deployed, and section IV paragraph 10 routes each solution through a pilot before any scaling.

Area 8.4 is where the text reaches the merits, and, like every other area in paragraph 8, it is written as planned functionality with expected results in a programmatic Concept, not as a system in operation. Sub-area 8.4.1 lists systematic analysis of the parties' positions, help in analysing evidence, identification of contradictions in the evidentiary base, answers to questions about the case file with the source document indicated, and a pre-hearing brief for the judge. Sub-area 8.4.3 compares a draft judicial act with the legal positions of the Supreme Court and with decisions in analogous cases and warns the judge of possible errors. Sub-area 8.4.4 is intelligent assistance in sentencing: analysis of the circumstances and of mitigating and aggravating factors, statistics on penalties in analogous cases, a check of the qualification with regard to repeat offending, and reference recommendations based on the legislation and on the Supreme Court's clarifications.

Area 8.5 is the generation of draft judicial acts. Its functionality is the automatic formation of the structure and text of a draft decision, the text giving a court order and a ruling as examples, from the case materials, integration of the relevant norms, improvement of the text, and generation of individual blocks of a decision from the materials including audio recordings of the hearing. The expected result is stated as shorter document preparation time and the possibility for the judge to focus on the substance of the case. This is a contemplated function, not an operative one: it is gated by the four pilot stages in section IV paragraph 10, and nothing in the document says any court is generating draft judicial acts today. Area 8.6 is analytics on appellate and cassation practice feeding Plenum resolutions and practice reviews; area 8.7 is statistics and dashboards for court management.

The calendar runs to mid-2027; the numbers are for 2030

Annex 1 is a calendar with a quarter, a milestone and result, and a responsible organisation for each row; every row is a planned checkpoint, and the document reports none of them as reached. For the third quarter of 2026 it lists the formation of the Centre for Judicial Competences in AI and its start on implementation, a strategic session on applied use of AI technologies, and the summing up of a pilot use of AI products for official tasks with parameters defined for software, infrastructure and equipment. For the fourth quarter of 2026 it lists a decision on priority technological solutions, naming as the result the specific software, the procedure for placing AI services in the closed loop of the court system and the sources of funding, together with a decision on priority scenarios for official purposes.

For the first quarter of 2027 the calendar lists pilot use of the priority scenarios, with the procedure for scaling the most demanded scenarios and the process for training judges as the result. For the second quarter of 2027 it lists, again as a planned checkpoint, access for all courts to court AI services, with the result stated as AI services being used by judges in their work on a permanent basis, and a pilot of additional scenarios. The responsible-organisation column names the Supreme Court, the Judicial Department and the University. The PDF's text layer scrambles that column against the rows, so we do not assign a responsible body to any single milestone.

Annex 2 carries nine targets under a single column heading that reads, in our translation, target value (by 2030). In our paraphrase, outside quotation marks because the source is a two-column table: the share of judges using court AI services on a permanent basis, more than 95 percent; reduction in the time taken to hear simple, typical cases, by 20 to 30 percent; the share of hearing records prepared with AI, at least 90 percent; the share of draft procedural documents generated with AI, 30 to 50 percent; the share of drafts approved by judges without substantial edits, at least 80 percent; accuracy of evidence analysis and contradiction detection, at least 90 percent; time for a judge to find relevant judicial acts, no more than 3 seconds; reduction in complaints and in changes or reversals of judicial acts, by 15 to 20 percent; and the share of participants in proceedings using court AI services, at least 30 percent.

The table gives no baseline for any of the nine, no measurement method and no interim values. Section III paragraph 9 describes them as quantitative targets whose achievement is planned on completion of the Concept's measures. They are what the court intends to reach, and the document offers nothing that would let a reader check progress against them.

Where a litigator should read first: the risk matrix

Annex 3 on page 13 is a twelve-row table of risk, probability, impact and mitigation measures. Section VI paragraph 15 requires the Centre for Judicial Competences to update it as implementation proceeds and at least once per half-year. Section V paragraph 12 provides for the creation of that centre inside the Supreme Court as an expert group whose regulation is to be approved by the Chairman, and annex 1 schedules its formation as a third-quarter 2026 milestone; it is to draw on staff of the court, the Judicial Department and the University, judges of lower courts and, where necessary, representatives of state bodies, research organisations and technology partners. The probability and impact ratings below were read from a layout-preserving extraction of the PDF, because the plain text layer scrambles those two columns.

Row 1 is model hallucination, described as non-existent norms, acts and citations, rated high probability and very high impact. Its mitigations, which like every measure in this paragraph are listed as contemplated mitigations in a programme annex and not as rules in force, are a mandatory retrieval-augmented architecture with links to the source, verification of the result by the judge, a ban on the use of public AI services, and stress-testing of models. Row 3, leakage of personal data and restricted information, lists an isolated network, user training and a ban on transferring data to public clouds. Row 6, profiling of judges by their decisions, lists a full ban on personalised analysis of judges and anonymisation of judge data in analytics. Row 7, technological dependence on foreign models, lists domestic solutions and mandatory placement in a closed loop.

Row 5 is the one to set beside the pull quote; its measures, like every row in annex 3, are listed as contemplated mitigations in a programme annex, not as rules in force. The risk is a reduction in the independence and initiative of the judge, rated medium probability and high impact. The mitigations are training, a ban on default selection, positioning of AI systems strictly as an assistant, a mandatory statement of reasons for rejecting an AI recommendation, and interaction logs for later analysis. The Concept's principle is that the final decision remains with the human. The matrix's mitigation is that a human who rejects the recommendation records why. Both sentences are in the document, and a reader who quotes only the first has left out the second, which is the mechanism that qualifies it.

Row 10 is the risk of participants filing AI-generated materials with unreliable references, rated medium probability and high impact. The listed mitigations are mandatory declaration of the use of AI when filing documents, automatic checking of references to judicial acts and normative legal acts, and procedural sanctions. Row 12 is violation of the parties' right to a fair trial, rated low probability and very high impact, with the mitigations listed as the party's right to know that AI was applied, the right to challenge AI conclusions, and the mandatory final role of the judge. Every measure named in this paragraph is a contemplated mitigation listed in an internal risk matrix approved as part of a programme that binds no one outside the court system; none is a procedural rule in force, and the order sets no date for any of them.

What rows 10 and 12 are, and what they are not, decides how a practitioner should file this article. They are entries in a matrix of contemplated measures approved as part of a programme. They are not a procedural rule, not a practice direction and not a sanction schedule. In our reading, a declaration duty for representatives, or a sanction for a fabricated citation, would need a procedural instrument; this order contains none and promises none by any date. We do not say any of these measures is in force, and we do not say when or whether one will be.

What we did not verify

What we opened: the 13-page PDF at the Supreme Court URL given above, read end to end, including order paragraphs 1 to 5, the approval block, Concept sections I to VII, annex 1 calendar, annex 2 targets and annex 3 risk matrix. We extracted the text with pdftotext, ran a second layout-preserving extraction to read the risk table's probability and impact columns, and confirmed the file's SHA-256 against the copy retained by the gate reviewer. We also read the saved ConsultantPlus notice page, which states a site publication date of 15 September 2026 for the notice.

What we did not open: Federal Law No. 243-FZ of 26 July 2026 and Presidential Decree No. 490 of 10 October 2019, which the Concept cites; the Bulletin of the Supreme Court issue in which paragraph 4 orders publication; any regulation on the Centre for Judicial Competences; any Russian procedural code; and the chronology in this desk's earlier article on Russia's federal AI law, which we deliberately did not import because its identifiers have not been reconciled with the citation in Concept section I paragraph 3. We did not verify the signature cryptographically and we obtained no timestamp of the court's first posting.

What we refuse to claim: we do not say the order was first published on 15 September or was unavailable before then; the only 15 September date we hold belongs to a ConsultantPlus notice. We do not say the order took effect on any date, because it contains no effective-date clause. We do not say any of the seven work areas is deployed, because the text describes functionality and expected results and routes every solution through a pilot. We do not treat the 2030 figures as outcomes or as commitments to litigants; the table gives no baseline and no method. We do not say the order creates a disclosure duty, a sanction or a procurement entitlement, because rows 10 and 12 of the risk matrix are contemplated mitigations. We do not assign responsible organisations to individual calendar rows, because the PDF text layer scrambles that column. We do not say which lower courts the Concept covers beyond the order's own words. Where the word must or mandatory appears above, it is our translation of the instrument's own text and not the desk's characterisation.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Russia's AI law
• Philippines Supreme Court AI judiciary framework
• AI court rules tracker
• Europe AI regulation tracker
• AI Regulation News hub