evidoria

← Back to browse

Good practice Imported

Albert API — France's sovereign generative-AI platform for civil servants

France · Paris · See the France profile · See the Paris profile

Evidence: Descriptive / self-reported Top 66% 53/100 · Ask Evidence Copilot about this practice

DINUM's Albert API (open-source Llama/Mistral, SecNumCloud) provides sovereign generative AI to French civil servants. A France Services pilot found 71% ease-of-use; 10,000 agents from 8 ministries tested it from Oct 2025; full roll-out paused pending a 2026 assessment.

71 %
Advisors who found the tool easy to use (pilot)
58 %
Advisors who would recommend the tool to colleagues (pilot)
10,000
Civil servants enrolled in expanded eight-month test (from October 2025)
43,000 Q&A sheets
Training corpus size
2.5 million
Target user base (French civil servants)
Albert API — France's sovereign generative-AI platform for civil servants Albert API — France's sovereign generative-AI platform for civil servants

Details

Promoter
Direction interministérielle du numérique (DINUM)
Period
2024–present
Keywords
digital government, AI infrastructure, civil service, productivity

Context

Albert is the French state's sovereign generative AI platform, launched by DINUM (Direction interministerielle du numerique) in April 2024 for use by French civil servants in Paris and across the administration. Built on open-source foundations — Meta's Llama 3.1 and Mistral AI's models — its codebase is published on GitHub and Hugging Face, and its inference endpoint runs on French SecNumCloud-certified infrastructure so that sensitive administrative data never leaves French jurisdiction.

Objectives

Albert aims to give France's roughly 2.5 million civil servants a sovereign generative AI tool for administrative tasks, keeping sensitive data within French jurisdiction via SecNumCloud-certified hosting and open-source models.

Activities

DINUM trained Albert on a corpus of 43,000 Q&A sheets from service-public.fr, piloted it with advisors, and then expanded testing to 10,000 agents from eight ministries in an eight-month test beginning October 2025, targeting applications such as tax administration (an estimated 16 million annual citizen requests) and environmental project review (about 4,000 dossiers a year).

Results

In the initial pilot, 71% of advisors considered the tool easy to use and 58% said they would recommend it to colleagues. During the expanded eight-month test with 10,000 agents from eight ministries, significant technical malfunctions and erroneous responses emerged.

Conclusions

Following the technical issues identified in expanded testing, DINUM paused full generalization of Albert in its current form; a comprehensive assessment is expected in summer 2026 before any broader rollout decision.

Implementation
Indicative cost
High (€500k–€5M) — No public budget figure is disclosed; classified as high cost given sovereign SecNumCloud-certified infrastructure, custom LLM training/hosting, and a target user base of 2.5 million civil servants.
Time to results
Medium (1–3 years) — Launched April 2024, with an expanded eight-month test running from October 2025 into 2026 and a comprehensive assessment expected summer 2026; classified as medium given roughly two years of development and testing without yet reaching a stable, fully scaled state.
Staffing & skills
DINUM (Direction interministerielle du numerique) — lead public agency, Mistral AI — model co-development / technical partner

Conditions for success

  • Open-source model foundations (Llama 3.1, Mistral) allowing code transparency on GitHub/Hugging Face
  • SecNumCloud-certified sovereign hosting to satisfy data-jurisdiction requirements
  • Domain-specific training corpus (43,000 Q&A sheets from service-public.fr)
  • Phased rollout from small pilot to a supervised 10,000-agent, eight-ministry test before any full generalization

Common failure modes

  • Significant technical malfunctions and erroneous responses emerged during the expanded eight-month, 10,000-agent test, prompting DINUM to pause full generalization of the tool in its current form pending a comprehensive assessment expected summer 2026.

Where it fits

Governance type
centralized interministerial digital agency (DINUM) under national government
Scale
national, targeting ~2.5 million civil servants; currently paused after a 10,000-agent expanded test
Income level
high-income (France)

Commonly funded by

National / regional programmes Digital Europe Programme Recovery and Resilience Facility (national plans)

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful