The French government's Direction interministérielle du numérique deployed an experiment of the Assistant IA on 10,000 licences between December 2025 and June 2026, and the Direction interministérielle de la transformation publique evaluated the pilot's usage, perceptions and impacts with the ministries of the economy and finance and of justice, together with researchers from LaborIA-Inria, the Laboratoire Fabrique de Pensée Critique at INSA Rennes and the Laboratoire Interdisciplinaire Sciences Innovations Sociétés at CNRS. The Assistant IA is a sovereign conversational agent for the everyday administrative work of civil servants: drafting emails, summarising documents and translating text, while keeping public data confidential.
The evaluation was deliberately mixed-method, combining four questionnaires to the economic and finance ministries, including 1,359 responses qualifying the tester panel in December 2025 and 465 end-of-experiment responses in June 2026, two questionnaires to the Ministry of Justice with 700 responses in January 2026, around twenty semi-directive interviews, and three in-person focus groups with 33 participants from the two partner administrations.
The focus groups used an experimental sequence in which volunteers performed the same task twice, once with and once without the assistant: writing a reply to a chain of solicitation emails and producing a summary of a six-page document. Measured on those timed tasks, drafting an email went from 21 minutes without AI to 19 minutes with it, a gain of 2.5 minutes, while article synthesis went from 33 minutes to 28 minutes, a gain of 5.4 minutes.
Quality scores from two independent blind graders out of 20 rose from 14.14 to 15.09 for email and from 14.15 to 15.43 for synthesis, gains of nearly one point and more than one point respectively. 75 percent of testers in the two ministries considered the assistant useful for their job, 65 percent would recommend it to a colleague and 56 percent said they now use non-sovereign solutions such as ChatGPT or Perplexity less often in their professional work.
Usage itself was production-oriented, with 70.9 percent of uses asking the assistant to produce a concrete result or execute a task, against 28.4 percent asking for information and 0.6 percent expressing an opinion, and 60.5 percent of testers judged they had progressed in command of AI tools during the evaluation period. Since the start of the experiment, 50 percent of respondents used the assistant at least once a week.
The evaluation is explicit about its limits, stating that the results cannot be generalised to all civil servants, that the sample was unrepresentative with category A and A+ agents making up 86 percent and 35 percent of respondents in management roles, and that the figures apply to the two partner ministries. It also documents thirteen deferred "debts" created by AI adoption, grouped into five families spanning skills, real work, oversight, technology and organisational load.