What is frugal AI?

The expression "frugal AI" comes in France from a specific document: the Référentiel général pour l'IA frugale, published in June 2024 as AFNOR Spec 2314. It was led by the Ecolab of the French Ministry of Ecological Transition and drafted from January to June 2024 by a working group of companies, start-ups, public bodies and laboratories, with the support of Hub France IA, EcoInfo, ADEME and Arcep. Its 100 pages contain definitions, a measurement method, 31 good practices and rules for communication.

An "AFNOR Spec" is not a standard. The document states that the AFNOR Spec "must not be presented as a French standard or as equivalent to a French standard" and that its content "binds only its authors" (our translation). It is voluntary and aims to prepare future European and international standards. In addition, a draft European standard on assessing the environmental impact of AI systems, prEN 18287, has been in preparation at the joint technical committee of CEN and CENELEC since October 2025; we found no document linking it to the framework.

The framework defines the frugality of an AI service as the overall reduction of its resource needs and impacts, obtained by redefining uses or performance requirements. An AI service is frugal when 3 conditions are met.

1. It has been demonstrated that an AI system is needed, rather than another, less resource-intensive solution, for the same objective.
2. Good practices are adopted by the producer, the provider and the customer of the service to reduce its impacts.
3. Uses and needs have been questioned beforehand and aim to stay within planetary boundaries.

The document distinguishes frugality from efficiency. An efficient system applies known optimisations to consume less at each use, the need remaining fixed. A frugal service accepts a result judged sufficient and reasons about its total consumption: to take an example of our own, a model that consumes 2 times less per query, but is served 10 times more often, would be efficient without being frugal. The first condition may lead to not using AI at all; the framework states that "the injunction to develop an AI service must be open to challenge" (our translation).

The 31 good practices come mostly from testimonies and interviews with the organisations in the working group. They are sorted into 7 themes (relevance of AI, governance, skills, data, model performance, equipment, measurement) and placed on the stages of an AI system's life cycle, from launch to decommissioning. A single practice often applies to several stages: 19 of the 31 concern design and development.

Design and development concentrate the most practices, and the framework concludes that design and governance weigh on the whole life cycle of a project. Participants also rated each practice on its frugality gain and its implementation effort. 5 practices rank among both the 10 strongest in gain and the 10 easiest: analyse the need (BP01), plan for end of life (BP11), optimise the use of existing equipment (BP20), reuse already trained models (BP29) and favour more frugal models (BP30). The framework recommends starting with them.

To assess an AI service, the framework relies on life cycle assessment and on the ADEME framework for digital services. The unit of measurement is the service delivered over one year: the year's inferences, training and other fixed costs, and each new training run, including those that did not succeed. The system boundary covers users' devices, networks and data centres.

4 indicators have priority: climate change, depletion of mineral and metal resources, primary energy and final energy. Water comes at medium priority, with acidification, fine particles and ionising radiation. Measured data are preferred to set values, such as the maximum thermal power announced for a chip (TDP).

For electricity, the framework favours the grid emission factors where the service and its components operate, rather than market-based ones that take purchase contracts into account. To claim that a service has a positive balance on an impact category, it excludes offsetting, the purchase of low-carbon or renewable electricity and the recovery of heat from servers. This is the same rule the hub applies when it recalculates providers' figures depending on the grid.

Mistral AI states that it followed the framework's method for the life cycle assessment of its Mistral Large 2 model, published in July 2025 with Carbone 4 and ADEME and reviewed by Resilio and Hubblo. It is the only public application of the framework by a model provider that we have found.

The text gives no target value: no electricity consumption, emissions or water volume above which a service would be frugal. It states that the assessment of an AI service "can only be relative (between two systems) and not absolute" (our translation). A service is frugal by the approach it follows; to claim it, it must publish its life cycle impacts, with no level required.

The ranking of good practices rests on opinions. The framework itself cites its limits: "absence of indicators and measurements", an approach "based on expert opinion", questionable representativeness of the working group and the "failure to take rebound effects into account" (our translation). The gain from a practice is therefore not quantified; the document suggests quantifying it in updates to the compendium, and we found none.

No verification is required. The framework provides for no certification or label. To communicate on the frugal nature of a service, it asks for a life cycle assessment to be published with the 4 priority indicators, the detail of the good practices applied and the expected indirect effects; it also asks to state whether the assessment underwent a "critical review by an external third party or not" (our translation), and recommends renewing it at least once a year. It excludes social and ethical questions from its scope and takes no position on the usefulness of services by sector, such as health or security.