How should AI be counted in a local authority's responsible digital report?

56% of local authorities with more than 3,500 inhabitants started an AI project in 2026, and 36% of those see its environmental impact, "and the difficulty of measuring it", as an obstacle (our analysis of AI adoption by local authorities). The largest already have a framework for counting: the annual report of their responsible digital strategy. This guide sets out what that framework requires and proposes the AI line it lacks.

Article 35 of the law of 15 November 2021 aimed at reducing the environmental footprint of digital technology, known as the REEN law, requires municipalities with more than 50,000 inhabitants and inter-municipal authorities with their own tax powers and more than 50,000 inhabitants to define, by 1 January 2025 at the latest, a responsible digital strategy, with their objectives for reducing the environmental footprint of digital technology and the measures to achieve them. This strategy is reported on annually in the report on the sustainable development situation (article L. 2311-1-1 of the Code général des collectivités territoriales). Departments and regions are not covered.

The decree of 29 July 2022 specifies that the preliminary work programme includes an assessment of the environmental impact of digital technology and its uses in the area, and that the strategy sets objectives, associated monitoring indicators and measures. It gives no indicator, no scope and no calculation method.

The indicators cited by the 2023 guide relate to hardware and staff: number of staff trained, average lifespan of the equipment, consumption of server rooms. Among the published strategies we have read, those of Angers Loire Métropole, Montpellier Méditerranée Métropole and Métropole Rouen Normandie mention AI as a subject to govern, without listing it among the items measured.

The AI line is built use by use, like the AI item of a Bilan Carbone®, with one difference: a local authority also serves the public, and its services open to the public are included in the count. For each use, you need an annual volume and the country where the servers run.

The country of the servers matters more than any other setting: for the same electricity consumed, emissions are about 9 times higher with the world average electricity, 458 gCO2e per kWh, than with French electricity at the ADEME factor, 51.9 g, the one a regulatory carbon footprint requires. Administration consoles give the number of users and requests; the model and the country of the servers have to be requested from the supplier, and the State's purchasing sheet offers a clause asking the supplier for this information, which the local authority can reuse at each contract renewal. A model hosted on the local authority's own servers is already in the consumption of its server room: the AI line flags it so as not to count it 2 times.

The following example completes the line for a typical inter-municipal authority with 1,500 staff, which has made an assistant available to its staff, transcribed the meetings of 200 of them, opened an assistant on its internal documents and a reception assistant on its website, and translated its website. Each value comes from an entry in the AI Impact Database, with its default settings, scaled to the volumes in the example.

The staff assistant accounts for more than half of the line. Its value starts from the civil service average: about 240 Wh per staff member per year, users or not, based on the Database entry on public sector staff, which breaks down the 5.85 million staff by occupation. An office worker who uses it every day consumes more, about 1.3 kWh a year according to our analysis by occupation ; a local authority that knows its number of active users can start again from this value with the hub's estimation tool. The transcription is the Database's for 200 people instead of 100; the document assistant takes the Database as it stands, 150 users and 20,000 pages indexed 2 times a year.

The order of magnitude changes in 2 cases. An model hosted in the server room, on 2 GPUs switched on all year, consumes about 10,400 kWh per year, 16 times the line in the example. Some coding agents, which write, test and fix code by chaining tasks, bring the use of one developer to about 400 kWh a year. These 2 cases are the ones to spot first in the inventory.

The decree also refers to uses in the area. If the local authority chooses to count its inhabitants' use of AI, the order of magnitude changes again: the Database estimates the personal use, studies and work of the inhabitants of a town of 100,000 inhabitants at about 41 MWh a year, more than 60 times the line in the example, including the work of the staff who live there. We recommend presenting them separately, since the local authority has no direct control over them.