--- title: "Measuring and managing your company's AI:\nwho can help?" description: "How to measure your company's AI footprint: free tools, platforms, specialist services and consultancies compared, and what 3 usage rules save." canonical: https://projetcelsius.com/en/blog/measure-ai-footprint-in-your-company/ author: "Guillaume Pakula" datePublished: 2026-11-06 dateModified: 2026-11-06 language: en translationOf: https://projetcelsius.com/blog/mesurer-piloter-ia-entreprise/ --- # Measuring and managing your company's AI: who can help? **Author**: Guillaume Pakula **Published**: 2026-11-06 **Section**: Practical guide **Canonical URL**: https://projetcelsius.com/en/blog/measure-ai-footprint-in-your-company/ **French original**: https://projetcelsius.com/blog/mesurer-piloter-ia-entreprise/ **Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille. > An organisation can calculate its AI footprint in-house, as an order of magnitude, with **free tools**. To track it and set **usage rules**, AI governance platforms, specialist services and consultancies, Projet Celsius among them, offer support **on quotation**. The item is still small today. Model size and where the computation runs change the footprint **by a factor of 10 to 18**, and keeping large models for the requests that need them already cuts it by **40%**. ## Key takeaways - An **order of magnitude** can be calculated in-house, with EcoLogits or Projet Celsius's Base IA. - The Copilot console counts requests and shows **no emissions** at all for that usage. - Offering **large models on request only** cuts an organisation's AI footprint by about 40%. - If 5% of requests become **videos**, the AI footprint is multiplied by 4 to 12. In 2025, **18% of French companies** with 10 or more employees used at least one artificial intelligence (AI) technology, 3 times more than in 2023, according to [Insee](https://www.insee.fr/fr/statistiques/9025878) (the French national statistics office, in French). On measurement, the only published figure covers large groups. In a [Capgemini survey](https://www.capgemini.com/insights/research-library/sustainable-gen-ai/) published in January 2025 of **2,000 executives** at organisations with a turnover above $1 billion, **only 12%** said they measured the footprint of their generative AI. > **Who wrote this article?** Projet Celsius, which wrote this article, is one of the providers compared. Each is described from its public pages as at 6 November 2026, and none has paid to appear here. For an office-based company, the item is still small, **a few kilos of CO2e a year**, as [our AI Bilan Carbone® calculator](https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/) shows (Bilan Carbone® being the French carbon accounting method), far behind the purchases and business travel of a [first Bilan Carbone®](https://projetcelsius.com/en/blog/bilan-carbone-cost-france/). **Measuring** it is mainly a way to track its growth. The **usage rules** that keep it in check take little effort. ## What do you need to count to measure AI? Measuring an organisation's AI means **counting its actual usage**, then converting it into electricity, CO2e (CO2 equivalent) and cost. A conventional responsible digital approach, known as Green IT, does not see that usage. It starts from an inventory of devices and data centres, which account for **50% and 46%** of the digital footprint in France according to the [ADEME study](https://ecoresponsable.numerique.gouv.fr/docs/2024/etude-ademe-impacts-environnementaux-numerique.pdf) (ADEME being the French Agency for Ecological Transition) of January 2025, and that study does not count generative AI. 3 reasons argue for starting in 2026: **usage is multiplying**, its nature is changing, and buyers and funders are starting to ask the question. ### Usage is multiplying Companies using AI that analyses written language, the family of ChatGPT, Copilot and Mistral, employed **48% of employees** in France in 2025, against **27% a year earlier**, again according to [Insee](https://www.insee.fr/fr/statistiques/9025878). Usage is also changing in nature: an **agent**, a program that chains dozens of steps on its own, consumes **far more than a question**, as our article on [AI electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) explains. ### Legal obligations Since 2 August 2025, the [EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689) has required **providers of general-purpose AI models** to document the energy consumption of their models (Annex XI). It **imposes no such requirement** on the companies that use them. For those companies, AI goes into the emissions report like any **purchase of a digital service**, under indirect emissions, as our guide to [including AI in your Bilan Carbone®](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/) explains. French local authorities have one more reason. The Territoires d'IA funding window of the [Banque des Territoires](https://www.banquedesterritoires.fr/sites/default/files/2026-07/Juillet2026_Banque_des_Territoires_Cahier%20des%20charges%20_guichet%20cofinancement_IA.pdf) (in French), opened on 22 May 2026, co-funds **30% to 50% of their AI licences** for 1 year. It adds **10% of funding** to applications that commit to a frugal AI approach, as defined by the AFNOR framework (AFNOR being the French standards body). Since 1 January 2025, municipalities and groupings of municipalities with more than 50,000 inhabitants must also maintain a [responsible digital strategy](https://www.legifrance.gouv.fr/jorf/article_jo/JORFARTI000044327312) (in French) that is subject to an **annual review**. > **What is frugal AI?** The AFNOR [general framework for frugal AI](https://www.ecologie.gouv.fr/presse/publication-du-referentiel-general-lia-frugale-sattaquer-limpact-environnemental-lia) (AFNOR Spec 2314, in French), published in June 2024 after 6 months of work by 150 contributors, is voluntary and free to consult. It proposes a method for assessing the environmental impact of an AI service and 31 good practices. ### The 4 pieces of information to count 4 pieces of information make up an organisation's footprint, and they are the same ones used afterwards to set usage rules: - Who uses AI, and how often: occasional, regular and heavy users do not weigh the same, and **the number of licences** says nothing about actual usage. - What people use it for: a quick question, a summarised document, a [generated image](https://projetcelsius.com/en/blog/ai-image-carbon-footprint/), a video or an hour of a coding agent, with a gap of **1 to more than 1,000** in electricity. - Which model: according to the Base IA, Projet Celsius's database of AI impact figures, a large model consumes **about 4 times more** than a medium model for the same question, and **24 times more** than a small one. - Where the computation runs: served in France, the same question emits **about 11 times less** CO2e than with world average electricity, 41 grams per kWh against 458 according to [Ember data](https://ember-energy.org/data/yearly-electricity-data/) used by the Base IA. ## Can you estimate your AI footprint without help? An order of magnitude can be calculated **in-house**, with free tools and the volumes from the consoles. Setting usage rules also requires knowing which model each team uses, which the consoles rarely say. ### An order of magnitude with a public calculator The [EcoLogits calculator](https://calculator.ecologits.ai/), an open tool, estimates the footprint of a volume of queries model by model, with **a mode for a whole organisation**. Projet Celsius's Base IA publishes **its factors by task**, from a quick question to an hour of agent work, and [our calculator](https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/) uses them to derive a company's footprint from its headcount and usage. For a first question, "does AI weigh much for us?", these tools are enough. Our article on [the carbon footprint of AI](https://projetcelsius.com/en/blog/chatgpt-carbon-footprint/) gives the orders of magnitude for a query. ### Vendor consoles The [Microsoft 365 Copilot usage report](https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage) counts **active users** and the number of requests sent. The documentation for the [Microsoft 365 emissions dashboard](https://learn.microsoft.com/en-us/power-bi/connect-data/service-connect-emissions-impact-dashboard-microsoft-365), for its part, states that **Copilot is not included**. At Google, [Google Cloud carbon reporting](https://docs.cloud.google.com/carbon-footprint/docs/covered-services) covers **the Gemini API and Vertex AI**; Gemini in Google Workspace does not appear in the list of covered services. The consoles therefore give **volumes, rarely emissions**. ### The 4 most common obstacles Moving from an order of magnitude to usage rules almost always runs into 4 problems: - The consoles count users and requests, rarely the model used, **never where the computation runs**, and AI used outside the company's tools does not show up. - Several tools add up, an office assistant, a conversational assistant, the AI built into business software, each with its own export, units and period. - Agents are billed **in monthly credits**, with no published equivalence in hours of work or in electricity. - A measurement does not say what to change. You need **scenarios**, such as keeping large models for certain uses, and a body to decide between them. ## What does AI without rules cost? For an office-based organisation, AI currently weighs **a few kilos to a few hundred kilos** of CO2e a year, a very small share of its footprint. The risks lie in what the item becomes when usage changes, and in what it costs. ### An item that changes scale An **hour of a coding agent** emits as much as **nearly 1,500 questions** to a medium model, according to the Base IA, and a **10-second video clip** as much as about 1,100. In the example of a local authority with 1,000 users detailed below, turning 5% of requests into [generated videos](https://projetcelsius.com/en/blog/ai-video-environmental-impact/) **multiplies the footprint by 4 to 12**, depending on clip length. ### A cost that follows model size Large models also cost more. On [Anthropic's public price list](https://platform.claude.com/docs/en/about-claude/pricing) as at 6 November 2026, a million output tokens costs **$0.50** with Claude Haiku 5.5 and **$20** with Claude Opus 5.5, from the same generation, **40 times more**. Keeping large models for the requests that need them cuts **the bill as much as the footprint**. ### Questions from clients and elected officials A client asks its suppliers about AI in a climate questionnaire, an elected official wants to know what the tools deployed emit, an executive committee decides on a licence budget. Without a figure specific to the organisation, **the answer stays general**. > The future size of the AI item will depend above all on the usage rules organisations set now. ## Who offers to measure the AI footprint? **5 families of providers** help measure an organisation's AI and set its usage rules, from platforms designed for large groups to free open tools. A trained person can also do part of the work in-house. ### AI governance platforms They are aimed at large companies that deploy many AI systems and want to keep a register of them. [Fruggr](https://www.fruggr.io/fr) presents itself as an **AI governance platform** that maps usage, shadow AI included, and links each cost to a result. Its [English page](https://www.fruggr.io/en) also announces measurement of the carbon footprint of all its AI systems. Its consulting offer adds [8 training courses of 2 to 8 hours](https://www.fruggr.io/fr/expertise/formations-ia/), including how to build an AI usage charter. **No price is published**, and access is through a demo. It is software to deploy, designed for IT, finance and compliance. An AI platform can also show the footprint to each user. At Orange, the internal Live Intelligence platform shows each employee **their carbon footprint in grams** and steers them by default towards the lowest-impact models, according to [the page on its AI jobs](https://orange.jobs/fr/fr/tout-savoir-metiers-ia-orange/live-intelligence) (in French). The display only covers queries made through that platform. ### Carbon and responsible digital consultancies They serve organisations that want to train their teams or assess a specific digital service. [Carbone 4](https://carbone4.com/fr/nos-formations-au-catalogue/module-ia-environnement) offers a **15 to 20-minute online module** on AI and the environment, free for individuals and on quotation for a group. Carbone 4 worked with ADEME on the [**life cycle assessment** of Mistral Large 2](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai), which Mistral published in July 2025 and presents as the first comprehensive assessment of an AI model. [Hubblo](https://hubblo.org/fr/), which reviewed that study, carries out bespoke life cycle assessments of digital solutions, AI included. Their prices are on quotation. Neither **publishes an offer** to measure the usage of a whole organisation. ### Specialist AI measurement services They are aimed at organisations that already spend a lot on AI. [TokenClimate](https://tokenclimate.com/offre) promises to tell clients what their AI usage costs and what it emits, team by team, then to look for reductions with them, in a **monthly report**. It targets organisations that spend **at least €2,000 a month** on AI, and gives its price in writing after a 30-minute call. That threshold rules out most SMEs, and the offer does not cover the Bilan Carbone®. ### Projet Celsius We support SMEs, mid-cap companies (ETI in French) and local authorities, from a few dozen to a few thousand users, that want **a defensible figure** and usage rules. The engagement follows 4 steps: - The diagnostic of actual usage starts from console exports and a short questionnaire to teams, to distinguish **usage patterns** and the models used. - The footprint is calculated **by team and by use**, with the public, dated factors of our Base IA, which anyone will be able to check once they are published. - The usage rules, 3 or 4 of them, are quantified **as scenarios**, like those in the next section, then tracked in Pilotage IA, the online tool we launched in November 2026. It groups teams by usage pattern, sets the scenarios side by side and produces the action plan and the exports for management or elected officials. - The AI item goes into the organisation's **Bilan Carbone®**, with its other digital purchases, as described in our guide to [a company's digital footprint](https://projetcelsius.com/en/blog/corporate-digital-carbon-footprint/). The factors that drive our price are set out below. We do not give **legal advice** on the EU AI Act, and we do not measure servers that an organisation runs itself. Nor do we have a platform that routes queries to a model. ### Open tools and frameworks They suit teams that want to estimate on their own. [EcoLogits](https://ecologits.ai/latest/), backed by the CodeCarbon association, is a **suite of open tools** that estimate the footprint of generative AI models in use. [CodeCarbon](https://github.com/mlco2/codecarbon) measures **the energy of code** running on your own machines or servers, which is useful for a team that trains or hosts its own models. On method, the AFNOR framework and the [RIA31 framework](https://institutnr.org/ria31-le-referentiel-ia-ethique-et-responsable) of the Institut du Numérique Responsable (the French institute for responsible digital practices) bring together good practices. For images and video, the [GenAI footprint](https://www.labanquepostale.com/content/dam/lbp/documents/communiques-de-presse/2026/cp-genai-juin2026.pdf) alliance, launched on 18 June 2026 by 9 large European companies, including Publicis, AXA, Engie and La Poste, feeds its measurements into EcoLogits and into e-footprint, an open tool hosted by the Boavizta association. All these tools are **free**. You have to **enter the volumes yourself**, and none of them says which rules to adopt or how to get them accepted. ### A trained person, in-house In-house work suits organisations that have a CSR (corporate social responsibility) or digital manager available. **A trained person** reads the consoles, runs a calculator and proposes rules; the [Institut du Numérique Responsable MOOC](https://institutnr.org/) or the Carbone 4 module provide the basics. The cost is limited to **time spent**. What is missing is the **outside perspective** of a third party. ## How much do a few usage rules save? A few usage rules cut the footprint **by 40% to more than 90%**, and complement the [individual habits](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/) that our article quantifies one by one. The 2 calculations below come from the Base IA engine, with world average electricity unless stated otherwise. ### For an office employee Take an accountant who hands **865 tasks a year** to AI: quick questions, emails, spreadsheets, PDFs, summaries. With a large model and extended reasoning for everything, their year emits **6.0 kgCO2e**. With the default setting of a medium model, **378 g**. Moving simple tasks to a small model gives 294 g, **20 times less** than the heaviest mode and **22% less** than the default setting. ### For a local authority with 1,000 users Take a fictitious local authority with a few thousand staff, 1,000 of whom use an AI assistant. Over a year, each hands to AI 300 quick questions, 40 web searches, 20 PDFs to query and 10 summaries of long documents, a total of **370,000 requests**. With 30% large models and an unknown computation location, the year emits **252 kgCO2e**, about **250 g per user**. Agents or images would push this total up, as [our calculator](https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/) shows. This total is also the figure that elected officials and the annual review of the responsible digital strategy expect. The first rule concerns **model size**. Keeping large models for the requests that need them cuts the footprint by **40%**, doing without them by **52%**, and a small model for everything by **86%**. The ranking of [the most polluting AI](https://projetcelsius.com/en/blog/most-polluting-ai/) shows the same gaps model by model. ### Computation location, images and video The second rule concerns **where the model runs**. Served in France, the same models **lower the footprint by 91%**, because French electricity emits **about 11 times less** per kWh than the world average. A medium model served in France, which combines size and location, cuts it by 94%. This rule depends on what vendors offer, detailed in our article on [so-called eco-friendly AI](https://projetcelsius.com/en/blog/most-eco-friendly-ai/). The third rule concerns **images and video**, quantified in the "With images and videos" tab of the figure above: 3% images add 6% to the footprint, while 5% videos multiply it **by 4 to 12**. Keeping generated video for the teams that need it avoids this change of scale. ## How much does an AI diagnostic cost? **No public fee schedule** exists for measuring an organisation's AI and setting its usage rules. The open tools and the AFNOR framework are free. Governance platforms, specialist services and consultancies work **on quotation**, Projet Celsius included. ### What drives the price 4 factors make up most of a diagnostic quote: - **The number of users** and teams to distinguish, from a few dozen in an SME to several thousand in a local authority or a group. - **The number of tools** and access to their consoles: a clean export can be read in an hour, scattered usage calls for a questionnaire to teams. - **The expected deliverable**, ranging from a figure for a report to scenarios for an executive committee or a note for elected officials. - **Tracking over time**, from a one-off measurement to a tool updated every year, with the item included in the Bilan Carbone® that is [often entrusted to the same consultancy](https://projetcelsius.com/en/blog/carbon-footprint-consultancy-how-to-choose/). ### Where to start? For an SME, a mid-cap company or a local authority, we recommend a consultancy that starts from actual usage for the diagnostic and the usage rules, and that is the work we do: the footprint by team, 3 or 4 rules quantified as scenarios, then their tracking in Pilotage IA. The most useful rules are a lighter model by default, large models on request, [images and video](https://projetcelsius.com/en/blog/ai-video-environmental-impact/) kept for the teams that need them, and computation served in France where the offer exists. Other providers suit other profiles better: a large group that wants to keep a register of dozens of AI systems and their compliance will turn to **a governance platform**, and an organisation that already spends more than €2,000 a month on AI will find a monthly report at **a specialist service**. A team that trains its own models will measure its servers **with CodeCarbon**. A legal question on the EU AI Act is a matter for **a lawyer**. A small organisation that only wants an order of magnitude can stop at the EcoLogits calculator or ours. ## Key takeaways - Measuring an organisation's AI means counting who uses it, for what, with which model and where the computation runs. - No law requires a company to measure its AI footprint: the EU AI Act targets model providers. - Generated video and agents can change the scale of an item that is still small in an office-based organisation. - Keeping large models for the requests that need them cuts the footprint by about 40%, and a small model or computation served in France by nearly 90%. - Open tools and frameworks are free. Platforms, specialist services and consultancies, Projet Celsius among them, work on quotation. To gauge the weight of AI in your own footprint before talking to a provider, start with [our AI Bilan Carbone® calculator](https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/). ## Frequently asked questions ## Frequently asked questions ### Are companies required to measure the carbon footprint of their AI? No. Since 2 August 2025, the EU AI Act has required providers of general-purpose AI models to document their energy consumption, not the companies that use them. For those companies, AI goes into the emissions report as a purchase of a digital service. In France, however, municipalities and groupings of municipalities with more than 50,000 inhabitants must maintain a responsible digital strategy. ### How much CO2 does a company's AI use emit? A few kilos of CO2e a year for an office-based SME, about 12 kg for 50 people according to Projet Celsius's Base IA, and several tonnes with coding agents or generated video. For a fictitious organisation of 1,000 users that hands 370,000 requests a year to AI (questions, searches, documents), the calculation gives about 250 kgCO2e with world average electricity. ### Does Microsoft Copilot show its CO2 emissions? No. The Microsoft 365 Copilot usage report gives the number of active users and requests sent, and the documentation for the Microsoft 365 emissions dashboard states that Copilot is not included. These volumes are a starting point: you need to apply factors by model and by computation location to them. ### Which free tool can estimate an organisation's AI footprint? The EcoLogits calculator, an open tool backed by the CodeCarbon association, estimates the footprint of a volume of queries by model and offers a mode for an organisation. Projet Celsius's Base IA publishes its factors by task and feeds a calculator for companies. CodeCarbon measures code running on your own servers. ### How can an organisation reduce its AI footprint? Through a few usage rules. Take an example of 1,000 users calculated with Projet Celsius's Base IA. Keeping large models for the requests that need them cuts the footprint by 40%, a small model for everything by 86%, and the same models served in France by 91%. Turning 5% of requests into videos would, by contrast, multiply the footprint by 4 to 12. ### How much does an AI footprint diagnostic cost? There is no public fee schedule: consultancies, specialist services and platforms work on quotation. The price depends on the number of users and teams, the tools and access to their consoles, the expected deliverable (a figure, scenarios, a note for elected officials) and tracking over time. Open tools are free. ## Sources - Insee, [L'usage de l'intelligence artificielle multiplié par trois entre 2023 et 2025](https://www.insee.fr/fr/statistiques/9025878) (Insee note on the tripling of AI use, in French), July 2026. - Capgemini Research Institute, [Developing sustainable Gen AI](https://www.capgemini.com/insights/research-library/sustainable-gen-ai/), 14 January 2025 (survey of 2,000 executives, August-September 2024). - European Parliament and Council, [Regulation (EU) 2024/1689 on artificial intelligence](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689), Official Journal of the European Union of 12 July 2024, Article 53 and Annex XI. - Banque des Territoires, [specifications of the Territoires d'IA co-funding window](https://www.banquedesterritoires.fr/sites/default/files/2026-07/Juillet2026_Banque_des_Territoires_Cahier%20des%20charges%20_guichet%20cofinancement_IA.pdf), version 1.1 of 10 July 2026 (PDF, in French). - [Law No. 2021-1485 of 15 November 2021 on reducing the environmental footprint of digital technology](https://www.legifrance.gouv.fr/jorf/article_jo/JORFARTI000044327312), Article 35, Légifrance. The Légifrance pages are in French. - Ministry of Ecological Transition, [publication of the general framework for frugal AI](https://www.ecologie.gouv.fr/presse/publication-du-referentiel-general-lia-frugale-sattaquer-limpact-environnemental-lia), 28 June 2024 (AFNOR Spec 2314, in French). - Microsoft, [Microsoft 365 Copilot usage report](https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage) and [Emissions Impact Dashboard for Microsoft 365](https://learn.microsoft.com/en-us/power-bi/connect-data/service-connect-emissions-impact-dashboard-microsoft-365), documentation consulted on 6 November 2026. - Google Cloud, [Carbon Footprint, covered services](https://docs.cloud.google.com/carbon-footprint/docs/covered-services), updated 26 October 2026. - EcoLogits, [documentation](https://ecologits.ai/latest/) and [calculator](https://calculator.ecologits.ai/); [CodeCarbon](https://github.com/mlco2/codecarbon), consulted on 6 November 2026. - Mistral AI, [Our contribution to a global environmental standard for AI](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai), 22 July 2025. - Anthropic, [API pricing](https://platform.claude.com/docs/en/about-claude/pricing), consulted on 6 November 2026. - TokenClimate, [offer](https://tokenclimate.com/offre), consulted on 6 November 2026; Fruggr, [AI governance platform](https://www.fruggr.io/fr), [AI roadmap](https://www.fruggr.io/fr/expertise/feuille-de-route-ia/) and [AI training](https://www.fruggr.io/fr/expertise/formations-ia/), consulted on 7 November 2026 (in French). - Ember, [Yearly electricity data](https://ember-energy.org/data/yearly-electricity-data/), carbon intensity of electricity in 2025, used by the Base IA; ADEME, [assessment of the environmental impact of digital technology in France](https://ecoresponsable.numerique.gouv.fr/docs/2024/etude-ademe-impacts-environnementaux-numerique.pdf), update for the year 2022, January 2025 (in French); La Banque Postale, [launch of the GenAI footprint alliance](https://www.labanquepostale.com/content/dam/lbp/documents/communiques-de-presse/2026/cp-genai-juin2026.pdf), press release of 18 June 2026 (PDF, in French). - Projet Celsius, Base IA, calculation engine of 6 November 2026 (electricity: 41 gCO2e per kWh in France, 458 worldwide).