AI in the civil service: one in five employees exposed, and a transformation still without a pilot

A third of public jobs exposed: where the figure comes from

The figure has been circulating for a year: 36 % of public employment would be exposed to significant changes due to generative AI. It comes from a study by the consulting firm Roland Berger, The Public Sector in the Age of Gen AI, published on 19 September 2025. Its authors analyzed 450 occupations representing 351 million public jobs worldwide, task by task, using the method of the International Labour Organization (ILO). Result: approximately 125 million full-time equivalents exposed, of which 22 % are jobs likely to be "augmented" (AI assists the agent without replacing them) and 7,5 % potentially automatable. Conversely, 64 % of public employment would not be or would be little affected.

Two clarifications are necessary. First, this is a global figure and a measure of exposure, not a forecast of job cuts. Second, it concerns jobs and not "transformed occupations." Applied to France by the iFRAP Foundation based on Roland Berger's work, exposure would reach approximately 38 % of the 5,7 million public agents. The occupations most affected by automation are office functions: secretarial work, reception, call centers, data entry.

The inspectors' diagnosis: rapid but disorderly diffusion

For France, the reference is now the joint report of the General Inspectorate of Finance (IGF), the General Inspectorate of Social Affairs (IGAS) and the General Inspectorate of Administration (IGA), commissioned by Matignon in September 2025, dated April 2026 and made public on 2 July 2026. Transposing the ILO method to French administrations, it results in a more cautious framing than private sector estimates:

• one public agent in five is, in theory, significantly exposed to AI;
• one in seven is significantly exposed, especially in administrative and support functions;
• one in four is weakly exposed (social workers, nurses, etc.);
• one in two is very little exposed, in missions in direct contact with the public (teachers in front of their students, law enforcement in the field).

The inspections describe an adoption that is "recent and multifaceted." DINUM lists more than 140 AI products within the state, with very strong disparities: from zero to 74 projects depending on the ministry. In local authorities, adoption follows size: 90 % of metropolitan areas are engaged, while 73 % of municipalities with fewer than 3 500 inhabitants have no AI project.

"Artificial intelligence (AI) constitutes a major lever, but neither automatic nor exclusive, for transforming public action." (IGF-IGAS-IGA report, summary, April 2026)

"Shadow AI," symptom of absent oversight

The most revealing phenomenon is the use of consumer tools without employer validation. According to a study by the National Union of General Directorates of Local Authorities (SNDGCT), cited by the inspections, 40 % of local authority agents would use AI tools in an unregulated manner. The inspections see this as proof of a real appetite, but also of major risks: data security, inequalities between agents and between territories, reliability of decisions.

The same SNDGCT noted, in its reflection paper published in December 2025, that the fear of replacement could make AI perceived as a "civil servant killer." The sovereignty issue here is very concrete: when an agent copies a file into a commercial conversational assistant, the public data leaves the administration's perimeter without any rule having authorized it.

To address this, DINUM, together with DITP and DGAFP, published on 26 June 2026 the first interministerial guide on the use of AI for state agents. It sets out five principles: the agent remains responsible for their decisions; sensitive data can only be processed with authorized tools; transparency is required when AI plays a substantial role; usage must be sober, ethical, and respectful of copyright; agents must undergo continuous training.

The sovereign tool, between ambition and trial and error

The State attempted to propose its own tools. The most publicized experiment, "Albert France services," assisted advisors in France services offices in their responses to users. It will not be generalized "in its current form," announced DINUM on 9 January 2026. It has refocused its efforts on the "AI Assistant," a conversational agent integrating Mistral models and tested by 10 000 ministry agents, and on Albert API, a shared platform hosted in a SecNumCloud-qualified environment.

The resources remain modest. According to the inspections, between 2023 and 2025, DINUM committed 8,7 million euros on AI, the ministries 26,2 million, and France Travail 18,7 million. The report deems these fundings weak and fragmented, and recommends making secure general-purpose AI tools available to all agents, pooling infrastructure (graphics processors, hosting), and imposing reversibility, auditability, and data ownership clauses in public contracts to limit technological dependence.

Real gains, but conditional

The report documents concrete gains. The State's procurement directorate anticipates more than 60 % in time savings for supplier identification; the legal affairs directorate of the economic and financial ministries has halved the time spent on document research. As part of the "Public Services +" program, document-based AI has reduced the average response time to users from 19 to 3 days. In Plaisir (Yvelines), a municipal conversational agent reduced the rate of missed calls from 55 % to 8 %.

But the inspections insist: these gains only materialize after a learning phase, and involve investment and operating costs often higher than forecast. The theoretical exposure of a profession says nothing about actual productivity gains, let alone job cuts. To avoid a "graveyard of proof of concept," the report calls for starting from the real work of agents, co-building uses with them and with users, and developing a "technological social dialogue."

Who controls the transformation?

The political question is shifting. It is no longer about knowing whether AI will enter the administration, since it is already there, often through the back door, but who sets the rules. Three risks are emerging. The first is a two-speed administration, between equipped metropolises and small municipalities without resources. The second is a suffered dependence on commercial tools adopted by agents for lack of alternatives. The third is a transformation of professions driven by digital directorates rather than by general directorates, without forward-looking management of jobs and skills. The thirteen proposals from the inspections, from three-level training to the pooling of "commons," outline a response. The question remains whether it will be funded.