French State (DINUM and DITP)
French State tests its Mistral-based Assistant IA with 10,000 civil servants and measures faster, better summaries
France's digital directorate DINUM gave 10,000 civil servants a general-purpose AI assistant built on Mistral AI models and hosted on a sovereign cloud. The public-transformation directorate DITP evaluated it with surveys and timed focus-group tasks. 75% of surveyed users found it useful, and timed tasks and blind scoring showed gains.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
75%
Surveyed pilot testers in the finance and justice ministries who consider the Assistant useful for their job1
65%
Same testers who would recommend it to a colleague1
56%
Same testers who say they use non-sovereign AI tools less often at work thanks to the Assistant1
of the order of 12%
Time saved on drafting tasks in focus groups1
The problem
French civil servants were already using public chatbots for drafting and summarising, which raises data-leak risks. The state wanted a secure tool of its own for everyday tasks, and wanted evidence on whether it helps before offering it to all staff.
What they built
The Assistant IA is a general-purpose chat assistant run by DINUM, the state's digital directorate. It runs on language models from Mistral AI, a French company, on the state's Albert generative-AI platform, with data hosted in the SecNumCloud-certified cloud of Outscale. It supports tasks such as drafting notes, summarising reports, reformulating text, extracting information and translating.
How they did it
DINUM deployed 10,000 licences for civil servants. The launch article reported infrastructure costs of about 300,000 euros plus model licences, and training for users. DITP's behavioural-science team then evaluated the pilot with researchers from LaborIA-Inria, INSA Rennes and CNRS. It sent several surveys to testers in the economic and finance ministries and the justice ministry, held about 20 interviews and ran three in-person focus groups. In the focus groups, volunteers did two tasks, an email reply and a six-page document summary, once with and once without the assistant.
Results
DITP reports that 75% of surveyed testers found the Assistant useful for their job and 65% would recommend it. Also, 56% said they used non-sovereign tools less often. Focus groups showed time savings of about 12% on drafting and up to 16% on document synthesis, with better quality in blind scoring. For the synthesis task, average time fell from 33 to 28 minutes and graders' scores rose from 14.15 to 15.43 out of 20.
What went wrong
Not everything was positive. In the survey, 57% considered that other generative AI tools answer their needs better. The main brakes were lack of trust in answers and need for training, and one manager said that, in general, it does not save time because everything must be validated.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Surveyed pilot testers in the finance and justice ministries who consider the Assistant useful for their job | 75%1 | |
| Same testers who would recommend it to a colleague | 65%1 | |
| Same testers who say they use non-sovereign AI tools less often at work thanks to the Assistant | 56%1 | |
| Time saved on drafting tasks in focus groups | of the order of 12%1 | |
| Time saved on document synthesis tasks in focus groups | up to 16%1 | |
| Timed article-synthesis task in focus groups of about 30 volunteers: average time without AI and with AI | 33 min without AI, 28 min with AI (average gain 5.4 min)2 | |
| Blind score out of 20 for the synthesis task, without AI and with AI, from two independent graders | 14.15 without AI, 15.43 with AI2 |
Lessons for your project
Measure with timed tasks, not only opinions. DITP paired surveys with focus groups where people did the same task with and without the assistant, and independent graders scored the results.
Report the mixed signals. A majority of surveyed testers still thought other generative AI tools answered their needs better, and the savings differed by task.
Train and target by task. Gains were larger for summarising than for drafting emails, so start where the benefit is clearest.
Offer a sanctioned tool to reduce unsanctioned use. The state framed the assistant as a secure alternative to public chatbots for staff.
Caveats
- The figures come from the French government's own evaluation team (DITP), and the publication is a government summary, not a peer-reviewed study. Le Monde Informatique, an independent outlet, covers the launch but not the results, so none of the measured figures is independently confirmed.
- The surveys covered testers in the finance and justice ministries only. The DITP page speaks of six pilot ministries, and the PDF says the 10,000 licences were deployed by DINUM; the percentages apply to those two ministries.
- Respondents were not representative: category A and A+ civil servants were 86% of the sample and 35% were managers. Most testers already had some experience of generative AI.
- The focus-group figures come from about 30 volunteers in an experimental setting. The report warns they cannot be generalised to all civil servants.
- The sources disagree on duration. The DITP page says ten months, the PDF says the experiment ran from December 2025 to June 2026, and the launch article says eight months from October 2025. The deployed date uses the December 2025 start given in the evaluation PDF.
- Le Monde Informatique names Mistral AI models and the Albert platform, not a model version.
Sources
- « L'Assistant » IA : bilan de l'expérimentation menée auprès des agents
DITP (modernisation.gouv.fr) · 2026-06-16 · Press release · modernisation.gouv.fr
- Expérimentation de l'Assistant IA dans la fonction publique : évaluation pilote des usages, perceptions et impacts
DITP (modernisation.gouv.fr) · 2026-06 · Research paper · modernisation.gouv.fr
- L'État ouvre l'expérimentation de la GenAI à 10 000 agents
Le Monde Informatique · 2025-10-28 · News · lemondeinformatique.fr
Summarized in our own words from the sources above. Spotted an error? See corrections.