What problem does it solve?
Supreme and constitutional courts sit at the top of a system that can send them enormous numbers of filings that repeat the same legal question. Brazil's Federal Supreme Court decides extraordinary appeals only when they raise a question with "general repercussion", a question of economic, political, social or legal relevance, often shared by thousands of similar cases in other courts; every appeal received is analysed by the court's own Secretaria de Gestão de Precedentes and, on classification, decided by the court president. The court describes an AI classifier's suggested theme as an indication that the justices always validate or confirm when they assess the case, not a classification that stands on its own. Colombia's Constitutional Court receives far more tutela actions, Colombia's fast track constitutional complaint for a violation of fundamental rights, than it can individually review, and has to select which ones raise an issue important enough for the court itself to decide.
Sorting filings at this scale by hand is slow and adds staff time with every filing received. It is also legally sensitive: which theme a filing is placed under can decide whether it is resolved on its own merits or bundled with thousands of others under a single leading case, so an error in classification is not a minor administrative slip.
How does it work?
- Digitise and structure the filing. Scanned appeals and petitions are converted to text and structured so a classifier can read them; Brazil's Victor system includes an image to text conversion step for this reason.
- Classify against the court's own themes. A model trained on years of the court's own decided cases suggests which recognised legal theme, or which selection criteria, the new filing matches.
- Surface the suggestion for staff to check. The suggested theme or selection flag appears to court staff, alongside the filing, for them to confirm, correct or escalate; the model output is a recommendation, not a ruling.
- Route by outcome. Confirmed filings that match an existing theme are grouped for combined resolution; filings flagged as raising a new or important issue go to the selection or admission process a justice ultimately decides.
- Retrain as the law moves. New leading cases and legal themes are added to the training set so the classifier keeps pace with what the court has decided since, rather than freezing its understanding of the law at deployment time.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Emerging
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | 85% | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time.
Indicative value
A supreme or constitutional court that receives 600,000 filings a year
EUR 666,667 to EUR 15.5 million
Staff triage time released per year
How this is calculated
Formula: filings * shareAutomatable * (minutesSavedPerFiling / 60) * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Filings received per year filings, filings per year | 200,000 | 620,000 | The high end, rounded down from 620,242, is Colombia's Constitutional Court's own figure for the tutela filings it receives a year on average, given by court president Alberto Rojas Ríos at the tool's launch. The low end, 200,000, is an editorial assumption for a smaller court, replace with your own. Source |
| Staff minutes saved per filing on theme classification minutesSavedPerFiling, minutes per filing | 40 | 60 | Brazil's Federal Supreme Court president was quoted in 2018, while Victor's classifier was still in testing rather than production, saying the classification work that would cost a court employee 40 minutes to an hour by hand is done by Victor in five seconds; the five seconds is not subtracted here because it is small next to the range. This overstates the time actually released where the court still has justices validate or confirm every suggestion: treat the low end of this range, not the high end, as the more realistic saving until you measure your own review time. Source |
| Share of filings routine enough for automated theme classification shareAutomatable, fraction of filings | 0.2 | 0.5 | Editorial assumption, replace with your own. Not every filing fits an existing theme; a meaningful share will always need a person's full read from the start. |
| Fully loaded cost of a court staff member's hour hourlyCost, EUR per hour | 25 | 50 | Editorial assumption. Replace with your own staff cost. |
What it leaves out: Time released, not cash saved, unless the court changes staffing. It leaves out the cost of building and validating the classifier against years of decided cases, the ongoing review every suggestion still needs, the cost of a wrong classification that reaches a person late, and the very different filing volumes and manual triage times across courts and jurisdictions.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Supremo Tribunal Federal
Brazil · Government and public sector · 2018
Victor is an artificial intelligence system Brazil's Federal Supreme Court (STF) began building with the University of Brasília in late 2017. Of its four planned steps (converting scanned filings to text, splitting a filing into its component documents, classifying those documents, and suggesting the theme of general repercussion a filing belongs to), only the text conversion step has been in production, since the end of December 2020: as of May 2021 it had processed over 10 million pages the same day they arrived. The theme classifier's own first lab results came in 2018, covering 27 themes of general repercussion. As of mid 2021, the court's own account says the document splitter and the document classifier ("classificador de peças") were still being readied for production with no date set, and that Victor as a whole is "mesmo ainda não definitivamente em funcionamento" (still not definitively in operation). Every appeal is analysed by the court's Secretaria de Gestão de Precedentes and decided by the court president; a suggested theme is an indication that the justices always validate or confirm when they actually assess the case.
- Accuracy: 85%, in tests, not yet in production
"Já temos feito testes no Projeto Victor, de inteligência artificial, que identifica os casos de recursos extraordinários ou de agravo em recursos extraordinários com acuidade de 85%"
Claimed by: organization
Corte Constitucional de Colombia
Colombia · Government and public sector · 2020
PretorIA is an AI system the Colombian Constitutional Court implemented in 2020, inspired by Argentina's Prometea, to help manage the tutela action (Colombia's fast track constitutional complaint for a violation of fundamental rights), of which the court receives over 600,000 a year. It searches, categorises and produces statistics on incoming tutela rulings so staff can identify the cases that raise an important or recurring issue and should be selected for the court's own review; categorisation and statistics were planned to be available initially only for health related rulings. Independent research describes it as functioning like a trained search engine: it does not decide which cases are selected, and every case the tool surfaces is still reviewed and chosen by a person. The court built PretorIA through a public private alliance with Universidad de Buenos Aires and Universidad El Rosario, alongside other supporting institutions.
- Interactions handled: about 4500, cases processed per day
"PretorIA aids this procedure by processing approximately 4,500 cases daily, helping the Court select which decisions should be reviewed."
Claimed by: independent
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A large set of the court's own historical filings, each labelled with the theme or selection decision a person made
- An owned, current taxonomy of the court's recognised legal themes or selection criteria
- Reliable text extraction for scanned and handwritten filings
Systems to integrate
- The court's own case management or digital docket system
- Document ingestion and optical character recognition for scanned filings
- The court's repository of leading cases and precedent, so the taxonomy stays current
Complexity: High
The classifier itself is a mature text classification problem, but building one a court can trust is slow: Brazil's Federal Supreme Court began its multi year partnership with the University of Brasília in late 2017. Its theme classifier showed its first lab results in 2018, covering 27 themes, and was reported at 85% accuracy in tests; by the court's own account, the optical character recognition step of Victor has been in production only since December 2020, with the document splitter and the document classifier still being readied for production and no date set for either, and Victor as a whole was still not definitively in operation. The hard parts are a large, clean set of the court's own decided cases to train on, a legal theme taxonomy the court agrees to use, and a validation and revalidation process the judiciary accepts.
- 1
Start with a suggestion, not a decision
Build the tool to propose a theme or selection flag that a justice or the reviewing judge validates, the design both Victor and PretorIA use, rather than one that classifies filings without review.
- 2
Digitise and structure the intake first
Convert scanned and paper filings to clean, structured text before classification runs; Victor's own pipeline treats this as a distinct step because a classifier is only as good as the text it reads.
- 3
Train on the court's own decided cases
Use years of the court's historical filings, each already labelled by a person's classification or selection decision, as the training set, so the model reflects how this court actually classifies, not a generic legal taxonomy.
- 4
Pilot on one filing type, then widen
Run the classifier alongside the manual process on one appeal type or a limited volume first, as Colombia's Constitutional Court planned to make PretorIA's categorisation and statistics available initially only for health related rulings, and compare its suggestions against what staff decided before relying on it more broadly.
- 5
Keep the taxonomy and the model current
Assign an owner to add new legal themes and leading cases as the court decides them, and retrain or revalidate the classifier on a schedule, not only when accuracy visibly drops.
Guardrails
- Every classification is a suggestion; a court official or justice confirms the theme or selection before it affects how a filing is routed or resolved
- No new precedent, deadline or substantive right is decided by the model; it sorts and prioritises only
- A record links each suggested classification to the human decision that followed, for audit and for measuring accuracy over time
- The theme and selection taxonomy is owned by a legal team, not inferred by the model on its own
KPIs to instrument
- Accuracy of suggested classifications against a sample staff reviewed independently
- Staff minutes per filing on classification, before and after
- Share of suggestions accepted unchanged versus corrected or escalated
- Backlog of filings awaiting classification or selection
Human in the loop
A justice or panel makes the actual admissibility, selection or merits decision in both deployments. PretorIA's own design keeps the reviewing judge as the sole decision maker. At Brazil's STF, every appeal received is analysed by the Secretaria de Gestão de Precedentes, and Victor's suggested theme is an indication that the justices always validate or confirm when they actually assess the case.
Common failure modes
- A confidently wrong classification on a genuinely new question
- A filing that raises a real new legal question gets sorted into an existing theme because it resembles one on the surface, and its novelty is missed. Sample newly classified filings for signs of a new issue and give staff an easy way to flag "does not fit any current theme."
- Bad text in, bad classification out
- A poorly scanned or handwritten filing produces garbled text that the classifier reads confidently but wrongly. Check extraction quality as its own step, the way Victor's image to text conversion is separated from theme classification, before trusting the classification.
- An "uncertain" queue nobody owns
- Filings the model is not confident about pile up without a service level target because no team is explicitly responsible for them. Give the uncertain queue an owner and a target time to clear, not just a lower confidence score.
What are the risks and rules?
EU AI Act
High risk
Annex III point 8(a) makes AI high risk when it is intended to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts. Classifying a filing by legal theme and flagging it for selection or combined resolution goes to which law applies and how, not only to organising the file, so the narrow procedural task exception in Article 6(3) is unlikely to apply here, unlike a tool that only summarises or indexes a file for a person to read.
Rules that apply
Guidance
- EU AI Act Annex III, high risk AI systems (European Union, Europe). Point 8(a) covers AI used by or for judicial authorities to research and interpret facts and law and apply the law to a concrete set of facts, which case theme classification and selection for review fall under.
- Article 6, classification rules for high risk AI systems (European Union, Europe). Sets out the narrow exception for AI that performs a narrow procedural task, improves the result of a previously completed human activity, detects decision making patterns without replacing the prior human assessment, or performs a preparatory task to an Annex III assessment, plus the Article 6(4) duty to document that assessment when a provider relies on it.
- EU AI Act Recital 61, purely ancillary administrative activities (European Union, Europe). Excludes AI for purely ancillary administrative activities that do not affect the actual administration of justice in individual cases, such as anonymisation or communication between personnel; theme classification and case selection affect the individual case, so this exclusion does not cover them.
Controls to put in place
- Validation of the classifier's suggestions against an independently reviewed sample before launch and on a recurring schedule after
- A named legal owner for the theme or selection taxonomy, who approves every addition or change
- Full record of every suggested classification and the human decision that followed
- A working route for staff to flag a filing that does not fit any current theme, reviewed regularly for signs the taxonomy needs to grow
Frequently asked questions
- Does AI decide which cases a supreme or constitutional court hears?
- Not in either deployment on this page. Independent research describes Colombia's PretorIA as functioning like a trained search engine that helps staff identify cases, with the reviewing judge kept as the sole decision maker. Brazil's Victor suggests the legal theme of an appeal that court staff at the Secretaria de Gestão de Precedentes analyse, and the justices always validate or confirm that suggestion when they actually assess the case.
- How accurate is this kind of classification?
- The CNJ news service, republished on Brazil's Federal Supreme Court's own news site, quoted then court president Dias Toffoli reporting Victor identified extraordinary appeal cases with 85% accuracy in tests, a figure from the court's own account rather than an independent audit, and from testing rather than production use. Accuracy will vary by court, filing type and how well the taxonomy fits the caseload, so measure it on your own filings before relying on it.
- How is this different from AI that summarises case files for a judge?
- Case file summarisation condenses a filing, evidence or earlier decisions into a structured summary for a person to read faster. This use case classifies a filing against the court's own legal themes or selection criteria to decide how it should be routed, a narrower and more legally consequential task that this page's own EU AI Act analysis treats as high risk, because the narrow procedural task exception in Article 6(3) is unlikely to apply to it.
- Is this high risk under the EU AI Act?
- Yes, as designed here. Annex III point 8(a) covers AI that assists a judicial authority in applying the law to a concrete set of facts, which legal theme classification and case selection for review both do, so the requirements for high risk systems, including risk management, logging and human oversight, apply.
How to cite this page
Blits.ai AI Use Case Library, "AI triage for case admissibility and selection at supreme and constitutional courts", last verified 30 September 2026, https://www.blits.ai/ai-use-cases/case-admissibility-and-selection-triage. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 30 September 2026: Published after review by an automated review workflow (independent skeptic review).
- 30 September 2026: Fixed adversarial review blockers: removed the unsupported claim that PretorIA's categorisation widened beyond health rulings (El Tiempo only says it was planned to launch there); dropped the implied full production replacement of Victor and stated instead that the STF always validates or confirms the suggested theme when a case is assessed; corrected general repercussion from a 'recurring' issue to a question of relevance shared by many similar cases, matching the STF's own wording; removed the contested 4,500 cases a day figure from metaDescription and replaced it with the court's own tutela filing volume. Also restored 'always validates' in implementation steps, humanInTheLoop, FAQ 1 and the Victor evidence summary; changed the classification step to say a justice or the reviewing judge validates, not staff; dropped 'about' from the five seconds quote; attributed the 85% article to the CNJ news service (republished by STF); tightened the Article 6(3) guidance wording away from 'purely preparatory or procedural'; noted the 40 to 60 minute saving comes from a 2018 test stage remark; and softened the unsourced 'does not scale' claim.
- 30 September 2026: Unpublished by an automated review workflow (independent skeptic review).
- 29 September 2026: First published