--- title: "What is the carbon footprint of AI in your company?\nIs it a lot?" description: "A company's AI emits from a few kilos to several tonnes of CO2e a year, and its use has tripled in 2 years. A calculator, the method and sourced figures." canonical: https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/ author: "Guillaume Pakula" datePublished: 2026-11-03 dateModified: 2026-11-03 language: en translationOf: https://projetcelsius.com/blog/bilan-carbone-ia-entreprise/ --- # What is the carbon footprint of AI in your company? Is it a lot? **Author**: Guillaume Pakula **Published**: 2026-11-03 **Section**: Explained **Canonical URL**: https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/ **French original**: https://projetcelsius.com/blog/bilan-carbone-ia-entreprise/ **Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille. > A company's AI emits from **a few kilos of CO2e a year**, for an office-based SME, to **several tonnes** when agents work on their own, generated videos multiply or the company sells a product built on AI, according to [Projet Celsius's AI Database (in French)](https://projetcelsius.com/ia/base/). And the figure is growing fast: AI use has **tripled in 2 years** in French companies, and the electricity used by the world's data centres is expected to **more than double by 2030**. The calculation comes down to 4 questions, and you can work out an order of magnitude yourself. ## Key takeaways - A company's AI accounts for **a few kilos to several tonnes** of CO2e a year. - **10 developers working with coding agents** emit 1.2 tCO2e of AI a year. - **Use has tripled in 2 years**, and data centre electricity should double by 2030. - An office-based SME with 50 staff stays at **about 12 kgCO2e a year**. The world's data centres used **415 TWh** (terawatt-hours) of electricity in 2024, about 1.5% of global consumption. The [International Energy Agency (IEA)](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf) expects them to use **more than double that in 2030**, around 945 TWh, **slightly more than the whole of Japan today**, and names AI as **the main driver of this increase**. In French companies, use is on the same curve: **18%** of companies with 10 or more employees used AI in 2025, 3 times as many as in 2023, according to [Insee](https://www.insee.fr/fr/statistiques/9025878), the French national statistics office (in French). The calculator above places each company using the engine of the [AI Database (in French)](https://projetcelsius.com/ia/base/), Projet Celsius's open database on the impact of AI. The result depends less on size than on use: **kilos with office assistants, tonnes when agents work on their own**. Agents, images and videos are precisely the uses that are spreading fastest. ## Does AI weigh much in a company's carbon footprint? **It depends mainly on what you hand over to it.** For an office-based company, the footprint is counted **in kilos of CO2e** (CO2 equivalent) a year, while the Bilan Carbone® (the French carbon accounting method) of an SME is counted in hundreds or thousands of tonnes. It becomes **heavy in 3 cases**, which have one thing in common: the AI works far longer than a person asking it a question. ### An office-based company: a few kilos a year An **office employee** who uses AI daily, for emails, meeting notes and questions, consumes **about 1.3 kWh a year** in data centres. With the world's average electricity, that comes to **0.59 kgCO2e a year**, the equivalent of 4 km by car. A services SME with 50 staff, a quarter of whom use AI, reaches **12 kgCO2e a year**. A company with 500 staff reaches **145 kg**. These figures are tiny next to a full carbon footprint assessment. SMEs supported by Décarbon'Action (a French carbon footprint programme) that have published their assessment on the platform of ADEME (the French Agency for Ecological Transition) report a median of **4,834 tCO2e a year** for companies with 50 to 99 employees, according to [our analysis of their assessments](https://projetcelsius.com/en/blog/industrial-sme-carbon-footprint-six-months/). Office AI accounts for **less than one hundred-thousandth** of that, far behind [purchases, often half of the footprint](https://projetcelsius.com/en/blog/scope-3-emissions-guide/). ### The 3 cases where AI weighs heavily The footprint changes scale as soon as **AI works on its own for long periods**, or **produces moving images**. The following 3 cases come from the same database, with the same conventions: - A team of 10 developers working with **coding agents**, programs that write, test and fix part of the code on their own: 1.2 tCO2e a year, more than 200 times the AI of an office employee, per developer. - An agency that produces its videos with AI: **8.4 kgCO2e for a single 3-minute video**, made of 18 generated shots, as much as 14 years of an office employee's AI. - A software publisher serving an AI feature to 100,000 users: **3.7 tCO2e a year**, because every query from every customer runs servers. **The third case is the one most often forgotten.** AI sold within a product does not sit with employees' tools: it runs **each time a customer uses it**, and [it is counted with sold products](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/). For video, the breakdown by model is in [our article on the impact of generated videos](https://projetcelsius.com/en/blog/ai-video-environmental-impact/). ### What these figures leave out The AI Database counts the electricity used by queries, data centre included. It leaves aside the **manufacturing of servers and chips**, **the training of models** and employees' computers. These emissions exist, and equipment manufacturing weighs heavily in [a company's digital footprint](https://projetcelsius.com/en/blog/corporate-digital-carbon-footprint/) as in [the footprint of AI over its whole life cycle](https://projetcelsius.com/en/blog/why-ai-is-bad-for-the-environment/). The orders of magnitude above are **estimates built on assumptions**, which the [AI Database (in French)](https://projetcelsius.com/ia/base/) sets out in detail and lets you adjust. > 10 developers working with coding agents emit as much as 100 office-based SMEs with 50 staff each. ## Why will the AI carbon footprint grow? The 2026 figures describe a use that is still young. **3 forces are pushing this figure upwards**, and they add up. ### More companies, more employees In 2025, 18% of French companies with 10 or more employees used at least one AI technology, **compared with 6% in 2023**. The increase covers every size of company, and **58% of companies with 250 or more employees** were already using it. Among companies using AI, 43% use an AI that generates text or speech, and **41% an AI that generates images**, videos or sound. ### Agents that work on their own A quick question to an assistant uses about [0.3 Wh (watt-hours)](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/). One hour of a **coding agent**, which chains 120 steps on its own while rereading everything it has already done each time, uses **440 Wh**, as much as **nearly 1,500 questions**. In the calculator, moving from an assistant to agents **multiplies by 13** the footprint of a services SME. Use is growing in the number of people, and it is also changing in nature. > The number of employees matters less than the kind of tasks handed to AI: one hour of an agent uses as much energy as 1,500 questions. ### Data centres doubling Worldwide, the IEA expects data centre consumption to **rise from 415 TWh in 2024 to about 945 TWh in 2030**. **The United States accounts for the largest share** of this increase, ahead of China. Every company that adds AI uses takes its share of this electricity, whether it counts it or not, and [AI's electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) is becoming a grid issue as much as a climate one. ## How do you calculate the carbon footprint of AI? The calculation comes down to **4 questions**, in this order: how many people use AI, how many tasks each of them hands over to it, how much electricity each task uses, and how much carbon that electricity emits. ### How many people, how many tasks Not everyone in a company uses AI, and not at the same pace. For regular use, the AI Database assumes that **a quarter of employees** use it, with **about 6 requests per working day** in an office job. It counts 2 requests for occasional use and 21 for intensive use. An accountant, a developer and a graphic designer do not have the same tasks: the database counts them **job by job**, based on the breakdown of jobs in each sector published by Eurostat. ### What a task costs Each task has its own electricity use, measured or estimated, and the gap ranges from **1 to more than 1,000** between a quick question and one hour of an agent. The chart below places them on a scale where each step multiplies by 10. A generated image, for example, uses **about 10 times as much as a question**, as explained in [our article on the footprint of an image](https://projetcelsius.com/en/blog/ai-image-carbon-footprint/). What makes up a company's footprint, then, is less the number of its employees than **the kind of tasks** they hand over to AI. One hour of an agent a day **weighs more than 1,000 questions**. ### From electricity to carbon The last step converts electricity into CO2e, using the carbon intensity of a kilowatt-hour in the country where the servers run. By default, the AI Database uses **the world's average electricity, 458 g of CO2e per kWh** in 2025 according to [Ember](https://ourworldindata.org/grapher/carbon-intensity-electricity), because **you rarely know where a query runs**. With servers in France, at 41 g per kWh, the same AI emits **11 times less**. ## Can you calculate your company's AI footprint yourself? **Yes, for an order of magnitude.** The figure becomes harder to defend when it has to go into **a verified carbon footprint assessment**, or when AI is part of what the company sells. ### An order of magnitude in a few minutes The calculator at the top of this article is enough to place your company. To go further, Projet Celsius's [AI carbon footprint tool](https://projetcelsius.com/en/tools/ai-carbon-footprint-calculator/) **quantifies a use from its volumes**. The [AI Database (in French)](https://projetcelsius.com/ia/base/) gives, for each task, each job and each type of organisation, its value, its range, its sources and its assumptions, **openly available**, under a CC BY 4.0 licence. The [hub on the impact of AI (in French)](https://projetcelsius.com/ia/) explains where these figures come from. ### What the Copilot and GitHub consoles show The admin accounts of these tools already provide part of the volumes. The [Microsoft Copilot usage report](https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage) shows **active users by app** and the **number of prompts sent per user**, which can be exported to a spreadsheet. [GitHub Copilot metrics](https://docs.github.com/en/copilot/concepts/copilot-metrics) track active users, chat requests and **agent activity**. None of these dashboards gives **either the electricity or the emissions**. They provide the hardest data to estimate, **the real volume**, which still has to be multiplied by the electricity of each task. ### Where it gets harder An in-house calculation reaches its limits in **4 situations**: - AI is hidden in software: the email client, the customer relationship management software or the translation tool use it without any invoice mentioning it. - Agents dominate use: their consumption depends on how long the tasks last, and licences count in monthly credits with no published equivalence in agent hours. - AI is in the product sold: every use by a customer counts, and the volume is read in the application's logs, not in an admin console. - The footprint has to be [verified by a third party](https://projetcelsius.com/en/blog/in-house-carbon-footprint-third-party-verification/): you then need dated assumptions, a source for every figure and a stated uncertainty. These limits are not specific to AI: [doing your own Bilan Carbone®](https://projetcelsius.com/en/blog/do-your-own-bilan-carbone-cost/) always runs into **data collection** more than into the calculation itself. ## And in the company's Bilan Carbone®? In a Bilan Carbone®, AI used by employees sits with **purchases of digital services**, such as hosting or software, among the indirect emissions of the value chain, which the methods call Scope 3. AI built into a product sold is counted with **the use of sold products**. The detail of this classification is in [our guide to including AI in your Bilan Carbone®](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/). For most companies, this line **will not change the result of the assessment**. Measuring it now nonetheless gives a **baseline year**, to track a use that has tripled in 2 years and to spot the moment when agents or a product take it to the tonne. A [first SME Bilan Carbone®](https://projetcelsius.com/en/blog/industrial-sme-carbon-footprint-six-months/) can include it by collecting usage volumes along with the other purchases. For us at Projet Celsius, the useful question concerns the uses that will change the scale of this line: coding agents, video and AI sold to customers. For an office-based company, the line currently comes to **a few kilos** and does not deserve days of calculation. We recommend **identifying these 3 uses now**, **asking suppliers for their volumes**, and measuring the line once to have a point of comparison. ## Key takeaways - A company's AI accounts for a few kilos of CO2e a year, about 12 kg for an office-based SME with 50 staff, up to several tonnes with coding agents or AI sold to customers. - Coding agents, generated video and AI built into a product sold change the order of magnitude, and these are the uses spreading fastest. - Use has tripled in 2 years in French companies, and the IEA expects the world's data centres to use more than double in 2030. - The calculation comes down to 4 questions: the people, their tasks, the electricity of each task and the carbon of the kilowatt-hour. - An order of magnitude can be calculated on your own with the AI Database. A figure that has to be verified in a Bilan Carbone® requires real volumes and a documented method. To see what lies behind these figures, read [what is the carbon footprint of a data centre](https://projetcelsius.com/en/blog/data-center-carbon-footprint/), and to choose the most frugal tool, [which AI has the smallest environmental impact](https://projetcelsius.com/en/blog/chatgpt-vs-claude-vs-gemini-environmental-impact/). ## Frequently asked questions ## Frequently asked questions ### What is the carbon footprint of AI in a company? It ranges from a few kilos to several tonnes of CO2e a year, depending on use. A services SME with 50 staff using an assistant daily emits about 12 kgCO2e a year, and a company with 500 staff about 145 kg, according to Projet Celsius's AI Database, with the world's average electricity. A team of 10 developers working with coding agents emits 1.2 tCO2e a year, and an AI feature served to 100,000 users 3.7 tCO2e. ### How much CO2 does an employee using AI at work emit? An office employee who uses an assistant such as ChatGPT daily emits about 0.59 kgCO2e a year for their work use, data centre included, the equivalent of 4 km by car. A Microsoft 365 Copilot licence used all year accounts for about 1.3 kgCO2e, and a GitHub Copilot licence about 19 kg, because a developer hands far more tasks over to AI. ### How do you measure AI energy use in a company? By multiplying 4 pieces of data: the number of people who use it, the number of tasks they hand over to it, the electricity of each task and the carbon intensity of the electricity. The admin consoles of Microsoft Copilot or GitHub Copilot give usage volumes. Projet Celsius's AI Database gives the electricity of each task, openly available. ### Should AI be included in a company's Bilan Carbone®? Yes, like any purchase of a digital service, among the indirect emissions of the value chain (Scope 3). For an office-based company, the line stays very small next to purchases, energy and travel. It becomes significant for a software publisher selling AI within its product, or in a team working with coding agents. ### Does an AI query use a lot of electricity? A quick question to an assistant uses about 0.3 Wh, data centre included, according to a Google measurement and consistent estimates used in Projet Celsius's AI Database. A generated image uses about 10 times as much, and one hour of work by a coding agent nearly 1,500 times as much. ### Is the AI carbon footprint lower with servers in France? Yes. French electricity emits about 41 gCO2e per kWh in 2025, compared with 458 g on average worldwide, according to Ember. The same use of AI therefore emits about 11 times less CO2e with servers in France. You rarely know where a query runs, however, which is why the world average is used by default. ## Sources - International Energy Agency, [Energy and AI](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf), April 2025, executive summary, p. 13. - Insee, [Les technologies de l'information et de la communication dans les entreprises en 2025](https://www.insee.fr/fr/statistiques/9025878) (in French), Insee Première No. 2120, Clément Lefebvre, 21 July 2026. - Projet Celsius, [AI Database, environmental database of artificial intelligence (in French)](https://projetcelsius.com/ia/base/), version 0.3 of 2 November 2026, CC BY 4.0 licence. - Ember and Our World in Data, [Carbon intensity of electricity generation](https://ourworldindata.org/grapher/carbon-intensity-electricity), 2025 data, accessed 3 November 2026. - ADEME, [Impact CO2, transport comparator](https://impactco2.fr/outils/transport) (in French), accessed 3 November 2026. - United States Environmental Protection Agency (EPA), [Greenhouse Gas Equivalencies Calculator, calculations and references](https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator-calculations-and-references). - ADEME, [GHG assessment dataset](https://data.ademe.fr/datasets/bilan-ges) (in French), Licence Ouverte 2.0, assessments of companies supported by Décarbon'Action, data of 2 November 2026. - Microsoft, [Microsoft Copilot usage report](https://learn.microsoft.com/en-us/microsoft-365/admin/activity-reports/microsoft-365-copilot-usage), Microsoft Learn, 14 July 2026. - GitHub, [Copilot usage metrics](https://docs.github.com/en/copilot/concepts/copilot-metrics), GitHub Docs, accessed 3 November 2026.