How are local authorities adopting AI?

The share of authorities involved doubled in 3 years, and its progress is slowing: 13 points from 2024 to 2025, 7 points from 2025 to 2026. 20% plan to launch one within the year; with them, the proportion has stood at around 3 authorities out of 4 since 2025. Adoption follows size: in 2026, all regions, 94% of metropolitan authorities and 70% of departments have a project, compared with 40% of departments in 2025.

Almost all of these authorities use generative AI, the kind that writes, summarises and answers: 86% of the authorities involved are testing or deploying it, against 52% in 2024. Older AI, which recognises images or forecasts consumption, also exists, in fewer projects. Political sponsorship remains a minority: 35% of authorities say that AI is driven by their elected officials, 67% of regions and 44% of metropolitan authorities.

Local authority AI first serves their own offices: 54% of the authorities involved use it for internal management, against 45% in 2024, and 31% for relations with users, against 20%. The other areas concern 4 to 12% each.

The survey does not break internal management down by department. The authorities that give testimony in the note describe office uses: drafting letters and minutes, legal analysis, translation and help with analysing tenders in Cannes, where more than 1,000 staff have been trained; drafting job offers and matching CVs to job descriptions in Saint-Cloud, with measured time savings for the first and "more mixed" results for the second; meeting summaries in the Loiret; an internal assistant, hosted in local data centres, at Rennes Métropole. Communication is the only department measured separately: in 96% of the 219 authorities surveyed by Cap'Com and the Observatoire in spring 2026, communication staff produce text or images with generative AI.

Uses aimed at the public are rarer: handling road reports, diagnosing the state of roads from images. These are 2 of the first 4 use cases of the Territoires d'IA programme of the Banque des Territoires, launched in May 2026, along with help with the financial analysis of medico-social establishments and assistance for rural town hall secretaries.

30% of authorities have drawn up a list of AI tools authorised or made available to their staff. Among them, 55% cite Copilot, Microsoft's assistant built into office software, and 52% Mistral, the French publisher; they are followed by Claude (23%), ChatGPT (21%) and Gemini (16%), and the French service DelibIA, which relies on Mistral (18%). The note sees Mistral's place as a choice of sovereignty. In the Ipsos poll published with it, 94% of French people consider it a priority that public service data be hosted in France or in Europe. Our analysis on French AI shows that, for the climate, the location of the servers matters and the nationality of the publisher does not enter the calculation.

In communication, these uses often remain free: in the Cap'Com survey, 58% of the 137 use cases described by communication departments rely on a free tool, and only 8% of organisations pay for licences. 41% of the authorities involved say they find it hard to choose between offers. The uses that staff make on their own initiative, which the survey does not measure, are judged important by the people interviewed in the note; our analysis on AI at work brings together what is known about them elsewhere.

Local authorities set rules before they measure. In 2025, 57% of the authorities concerned wanted to adopt a reference document; in 2026, 75% had set about it. 68% plan to train their staff, an obligation under the European AI regulation since February 2025 for organisations that deploy it. Social dialogue follows at a distance: 26% have put AI on the agenda of the bodies that represent staff. Our model charter compares 16 published charters, 4 of them from local authorities.

Residents expect these rules. In the Ipsos poll of July 2026 published with the survey, 92% of French people want a charter of authorised or prohibited uses in the public service, 93% want to be informed when an AI is used, and 81% prefer to deal with an official rather than with a messaging service run by an AI.

40% of the authorities involved do not yet have a tool to monitor the effects of their projects. Among all the authorities involved, 39% cite significant time savings, 24% faster procedures, 10% better service quality and 5% savings; 28% judge the effects very limited.

Environmental impact holds back 36% of the authorities involved, against 31% of those that have not started: it is one of the obstacles cited more often once the project has been launched, like access to quality data (44% against 28%) and fear of disappointing results (25% against 20%). The note adds "and its difficult measurement": none of the authorities that give testimony in the note provides the electricity or emissions of their uses, and the publication gives no figure for them. Our model charter offers an article that does this count every year. Rennes Métropole describes its assistant as "significantly less energy-hungry" than the large American models, with no measurement to support it. Residents are attentive to this: 79% of French people consider that the public service should limit the massive use of data and the use of AI because of their environmental impact, even if it reduces its efficiency; they were 75 to 77% from 2022 to 2025.

The order of magnitude can be calculated. The AI Impact Database estimates the use of AI at work by the 5.85 million French public employees at about 1,400 MWh per year, or about 240 Wh per employee, users or not, with the usual uses of an assistant. This average applies to the whole civil service; a local authority, without teachers or carers, has a different mix of occupations. For a local authority of 1,000 employees, this comes to about 240 kWh per year: about 10 kgCO2e if the servers run on French electricity, about 110 kgCO2e on the world average electricity. It is an estimate: it counts neither the services that the authority deploys for its users, nor the autonomous assistants that chain tasks, which bring a developer's use to about 400 kWh per year on their own, nor the manufacture of workstations.