How do you account for AI in a Bilan Carbone®?

Assistant subscriptions, Copilot licences and API calls are recorded in the accounts alongside other software. A Bilan Carbone® that counts its purchases of services in euros, as the method tolerates for services, therefore already includes AI, at the monetary ratio for IT services.

Neither the Bilan Carbone® method of the Association pour la transition Bas Carbone (ABC) nor the GHG Protocol has a rule specific to AI, and ADEME's Base Carbone has no factor for AI, GPUs or the cloud. Their general rules are sufficient, taken in the order of a carbon footprint: the line item, the activity data, the factor, the uncertainty.

To give a sense of scale, we take as a reference a carbon footprint of 8.5 tCO2e per employee. This is the median of 1,618 footprints of service companies published on the ADEME platform for the years 2022 onwards, with their indirect emissions, divided by the declared headcount or, failing that, the middle of its band (our calculation). Depending on how headcount is counted, this median ranges from 7.0 to 10.1 tCO2e, and half of the footprints lie between 3.2 and 40 tCO2e per employee.

Each use of AI is classified by a single question: who runs the servers? An organisation that buys an assistant, an API or cloud computing is buying a service. The electricity of the data centre belongs to its provider: it is included neither in the organisation's Energy item nor in its Scope 2.

Category 11 of the GHG Protocol covers the scope 1 and 2 emissions of end users, over the whole expected lifetime of the products sold in the year. An AI sold in a product and served from its provider's servers therefore remains in that provider's electricity or cloud purchases; only the energy of the customers' devices and servers goes into category 11, and into the Use item.

Employees' computers are already counted in the Energy and Fixed assets items. The method makes this point about digital services: an organisation's electricity consumption is not counted a second time under services. An in-house server room adds the refrigerant leaks from its air conditioning, in the Other direct emissions item and in Scope 1.

A Bilan Carbone® multiplies activity data by an emission factor. For AI, the easiest data to obtain is the furthest from what the servers consume: the amount on the invoices. The rest is held by the tools' administrators.

• For an assistant under licence: the number of licences and the level of use. The Microsoft 365 Copilot usage report gives active users by application and requests to Copilot Chat, but not the requests made in Word, Teams or Outlook, meeting summaries among them, which make up more than a third of the usage profile in the database.
• For coding tools: active users and agent activity (GitHub Copilot metrics), or the number of developers who work with agents.
• For an API: tokens read and written, by model, in the provider's console.
• For rented computing: the GPU hours on the invoice, or the provider's carbon tool, using its location-based value.
• For an AI served in a product: the requests, in the application logs.
• For in-house servers: the measured kWh, or the GPUs, their operating hours and their load.

The GHG Protocol describes 4 methods for purchases, from the most supplier-specific to the most general: supplier data, a hybrid method that starts from the supplier's activity data, average factors per unit purchased, and spend. For AI, the first exists only for part of the cloud. The Google Cloud carbon tool covers Vertex AI and the Gemini API; the Microsoft 365 emissions dashboard covers 8 applications, and Copilot is not one of them. The GHG Protocol guidance warns that the most supplier-specific data is not always the most accurate.

With no factor published by assistant providers, the hub's AI Impact Database calculates the electricity of each use, data centre included, from published measurements and usage assumptions that it sets out in detail. The table gives the factors needed for a carbon footprint, using world average electricity.

Carbon then depends on where the servers are: 458 gCO2e per kWh for the world average electricity, 210 in the European Union, 384 in the United States, 41 in France, in 2025 and on a life cycle basis (the hub map shows every country). The country where a request is processed is rarely published, but some providers publish a zone: Microsoft commits to processing requests to Microsoft 365 Copilot from its European customers within the European Union, excluding Anthropic models. For these customers, the licence is counted with European Union electricity, or 0.60 kgCO2e per year. The world average applies only when no zone is documented or fixed by contract. For an organisation's own servers in France, the reference factor for a carbon footprint is ADEME's, 0.0519 kgCO2e per kWh for 2024, a value smoothed over 2021 to 2024.

Google measures 0.24 Wh for the median text request to its Gemini assistant, and derives from it a carbon figure that takes account of its purchases of decarbonised electricity. The Bilan Carbone® method accepts only grid electricity for an organisation's energy; for consistency, we recalculate the providers' figures using the grid in the servers' zone.

Take a French services company with 200 employees and a median footprint of 1,700 tCO2e, which equips all its employees with Microsoft 365 Copilot, used at a typical level, with servers in the European Union. The 200 licences emit 0.12 tCO2e per year, or 0.007% of the footprint.

Coding agents change the scale of the line item, because one agent hour uses as much electricity as more than 2,200 quick questions. A developer who runs agents for 594 hours in the year emits 185 kgCO2e with world average electricity, 140 times a licence counted the same way.

The product feature weighs even more, because every request from every customer runs a server: 20,000 active users at 3 requests a day generate 22 million requests a year. If the company serves it from purchased cloud capacity, the line goes into its purchases of services; if it serves it from its own servers, into its electricity. In this example, the line item stays below half a per cent of the footprint. With a footprint at the 1st quartile of services companies, 3.2 tCO2e per employee, the licences would make up 0.02% and the full example 0.9%.

Back to the 200 licences. At €20 excl. VAT a month, a price close to what Microsoft lists for Copilot Business, they cost €48,000 a year. At the Base Carbone ratio for programming, IT consulting and information services, 75 kgCO2e per thousand euros in 2023, they come to 3.6 tCO2e, or 0.21% of the footprint and nearly 30 times the electricity of their servers. The ratio for publishing, which covers software publishing, would give 4.6 tCO2e; we use the first, which includes data hosting.

The 2 figures do not have the same scope. The monetary ratio, a sector average with 80% uncertainty, covers the supplier's entire chain, from its servers and their manufacture to the rest of its business, while saying nothing about AI. Counting by use includes only the electricity of the servers and the data centre; it is a floor. Manufacturing the servers would add around 8% with world average electricity and 17% with European Union electricity, according to the database entry on manufacturing per GPU hour. Model training has no published factor per use.

The only life cycle assessment published by a provider, that of Mistral Large 2 by Mistral AI with Carbone 4 and ADEME, gives 1.14 gCO2e for a response of 400 tokens, server manufacture included, training excluded. The difference from the database's quick question, 0.14 gCO2e, is mainly due to the size of the model and the electricity grid, and cannot be transposed to a licence.

The Bilan Carbone® method strongly recommends against using monetary ratios, tolerates them for services and minor sources, and limits their share over successive footprints: at most 10% in non-specific ratios at the Standard level, none at the Advanced level. With a monetary ratio, it notes, the only reduction lever becomes spending less. For a licence at a fixed price, switching to a small model or to servers in France would not change the figure, and a price cut would lower it without the servers consuming any less.

The footprint then loses what electricity does not cover: the manufacture of servers and the training of models. We add manufacturing when the database provides it, and we report training as a limitation, with a low completeness score in the uncertainty matrix.

The database range, from 0.48 to 0.78 kgCO2e for a licence with European Union electricity, covers only the electricity of the tasks, under its assumptions. The settings move the figure further: the same licence emits 0.18 kgCO2e a year for light use and 1.8 for heavy use, from 0.29 to 1.7 kgCO2e depending on the size of the model, which Microsoft does not publish, 0.12 with servers in France and 1.3 with world average electricity.

The actual volume of use, which the consoles partly provide, and the servers' zone therefore reduce uncertainty more than a more precise factor per task would; the size of the model remains unknown for most assistants.

The Bilan Carbone® method requires an uncertainty for each line, derived from quality scores for the factor and the activity data, expressed as a 95% confidence interval. For AI, the number of licences, taken from contracts, is very high-quality data; the level of use and the factor call for low scores for reliability and completeness, and for geographical and technical representativeness when the location and the model are unknown.

An organisation that installs an open model on its own servers moves to a different line item. With 2 H100 GPUs switched on all year and running at 60% load, the server and its room consume about 10,400 kWh a year. The database counts the room as a recent data centre, which is optimistic for a small server room. In France, at the ADEME factor, this comes to 0.54 tCO2e, entirely in the Energy item of the Bilan Carbone®; in the GHG Protocol format, 0.36 t go to Scope 2 (combustion at the power plant) and 0.18 t to category 3 (upstream and grid losses). With world average electricity, on a life cycle basis, it would be 4.8 tCO2e.

Manufacturing goes into the Fixed assets item, spread over the lifetime of the hardware in the organisation, which the method lets you estimate. With 5 years, the lifetime used by the database entry, it comes to about 1 tCO2e per GPU, including its share of the server (our calculation), or 0.42 tCO2e a year for 2 GPUs. With French electricity, manufacturing thus amounts to 77% of the electricity emissions; with world average electricity, 9%.

The GHG Protocol, by contrast, requires all manufacturing to be recorded in the year of purchase, in category 2: about 2.1 tCO2e for the 2 GPUs, then nothing. Over 5 years the total is the same but the individual years differ widely, which a footprint published in the 2 formats should state.

1. List the uses of AI and, for each, who runs the servers: assistants, coding tools, APIs, cloud, product, own or rented servers.
2. Classify each use: purchases of services for what a provider serves, Energy and Fixed assets for the organisation's servers, Use for the devices and servers of customers.
3. Collect volumes: licences and level of use, requests from the consoles, tokens, GPU hours, queries, kWh.
4. Apply one factor per use, with the electricity of the servers' zone when the provider documents it or a contract fixes it, the world average otherwise, and the ADEME factor for the organisation's own servers in France.
5. Remove the amounts counted by use from the line in euros, add the manufacture of servers when the database provides it, and score the completeness of the line accordingly.
6. Write down the scope, the zone used and the uncertainty of each line.

Our reading: for an office-based organisation, this item comes down to a single line, and we count it by use, because in euros a licence shows neither the effect of a smaller model nor that of servers in France, nor the point at which agents or an AI sold to customers change the scale of the item.

Reduction levers, the order of magnitude by company size and service providers are covered in 3 blog articles: our guide to integrating AI into your Bilan Carbone®↗, the Bilan Carbone® of AI in business↗ and measuring AI in business↗.