Superior Court of Quebec
Quebec Superior Court pilots nine AI agents with 22 judges; most want to keep them
Between December 2025 and March 2026 the Superior Court of Quebec gave 22 volunteer judges nine specialised AI agents for preparatory work such as editing, translation and finding legal articles. The court's own report says most judges want to keep the tools, with limits that include the need to check every answer.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
The problem
Judges in Quebec's trial court handle thousands of civil, family, commercial and criminal cases a year, and much of the work before a decision is reading, editing and looking things up. Some judges were already using AI tools on their own, with no institutional framework. The court chose to run a supervised test instead of leaving that use unmanaged.
What they built
Nine separate AI agents inside Microsoft Teams, built with Microsoft Copilot Studio. One edits and rewords text, one translates between French and English, four search the Civil Code, the Code of Civil Procedure, the Criminal Code and bankruptcy law, one checks legal citations, one gives IT support and one covers an internal family-law reference. The agents were not allowed to write judgments or interpret the law.
How they did it
Twenty-two volunteer judges, picked to reflect the court's mix of regions, case types and comfort with technology, took part from 8 December 2025 to 16 March 2026. They were split into three groups with staggered access. The agents ran in the Quebec Ministry of Justice's Microsoft 365 environment. Judges fed each agent excerpts by hand, conversations were private between users, and the agents could not browse the web. Judges met weekly to share uses and learn about AI risks. The court measured results with three surveys, platform usage data and 90 synthetic test questions.
Results
The agents logged 893 sessions, 479 of them on the writing assistant, which every responding judge used. Per the court, the drafting and translation agents scored 9.3 to 9.8 out of 10 on its test questions, and legal-research agents scored between 7.8 and 9.0. In the final survey 89% wanted to keep using the tools and 84% said AI fit their role. The court reports no judge said AI influenced a decision. The court says it will keep the most used agents, build in-house AI expertise and study a wider rollout.
What went wrong
The court's test found one invented author first name in the citation agent, which scored 0 out of 10 on that question. A Microsoft update to the underlying model delayed the launch by several weeks and forced the team to retest all agents. Judges who had access withdrawn partway reported a clearly worse experience than those who kept it. The court also could not guarantee that data stayed in Canada, and told participants so.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Volunteer judges in the pilot, and their share of the court's bench | 22 judges (11.17%)2 | |
| Total usage sessions across the nine agents | 8931 | |
| Sessions on the writing assistant, the most used agent | 4791 | |
| Judges in the final survey (19 of 22 answered) who want to keep using these tools | 89%1 | |
| Final-survey respondents who say AI is compatible with their judicial role | 84%1 | |
| Final-survey respondents who recommend a wider rollout | 63%1 | |
| Court's quality scores for the drafting and translation agents, from 90 synthetic test questions | 9.3 to 9.8 out of 101 |
Lessons for your project
Build several narrow agents instead of one general tool. The court gave each agent one task, such as translation or one code of law, so it could control what each could and could not do.
Keep staff data out of the tool by design. Judges pasted in excerpts by hand and the agents could not browse the web or see court files on their own.
Run weekly sessions where users share tips and learn about AI risks. The court says the most useful new uses, like transcribing scanned handwritten notes, surfaced there.
Plan for upkeep. A platform update to the underlying model forced the team to retest every agent, and the court calls for dedicated in-house AI skills.
Keep the check-every-answer rule. The court's own test found one invented author name in the citation agent, so verification stays with the user.
Caveats
- All figures come from the court's own report and summary (April 2026). The court wrote them about its own pilot, and no independent party audited them. Trade outlet Droit-inc repeats the court's figures.
- The survey results rest on 19 respondents out of 22 volunteers, who chose to take part. The court itself notes this limits how far the results can be generalised.
- The court's report names the model underneath Microsoft Copilot as GPT-4.0, updated to GPT-4.1 early in the pilot. It does not say exactly which model each agent used.
- The full report separately says 68% (13 of 19) recommend continuing the project without reservation, a different question from the 89% who want to keep using the tools. The report also says 90 days in one limits section, against 98 days elsewhere.
- The pilot did not allow drafting judgments or interpreting the law, so no result applies to those uses.
- The court site blocked automated downloads. Both court documents were read from web.archive.org copies, which carry the same April 2026 date.
Sources
- Projet pilote d'intelligence artificielle: Sommaire des résultats
Cour supérieure du Québec · 2026-04 · Filing · coursuperieureduquebec.ca
- Rapport du projet pilote IA
Cour supérieure du Québec · 2026-04 · Filing · coursuperieureduquebec.ca
- Les juges de la Cour supérieure conquis par l'IA
Droit-inc · 2026-05-26 · News · droit-inc.com
Summarized in our own words from the sources above. Spotted an error? See corrections.