The carbon footprint of generative AI
finally built into your footprint.

A line item regulation hasn’t caught up with yet

Writing assistants, conversational agents, image generation: the use of generative AI tools has become widespread in companies, across every sector. From a carbon standpoint, this usage falls under upstream Scope 3, within purchased digital services (GHG Protocol category 1).

But no official emission factor exists yet for large language model inference in ADEME’s Base Carbone. The database covers thousands of factors, updated every year, from electricity to transport to office supply purchases. AI inference simply isn’t in it yet. The result: a real line item, but one invisible in most carbon footprints on the market.

How Oakbon estimates this footprint, in decreasing order of precision

In the absence of an ADEME factor, Oakbon relies on available digital life-cycle assessment (LCA) literature, using a cascading method that always retains the most precise data available, without ever adding together two sources that would measure the same usage.

1. Actual queries, the most precise source. If you know your annual volume of text queries (LLM), Oakbon applies 0.005 kgCO₂e per query.

2. Failing that, AI licenses. If you don’t track your queries but know your number of licenses or active seats (ChatGPT Enterprise, Copilot, etc.), Oakbon applies 15 kgCO₂e per license per year, an intermediate estimate built on the assumption of around 3,000 annual queries per active license. The two sources are never added together: a license exists precisely to generate queries.

3. Generated images, in addition. Each AI-generated image is counted separately, at 0.030 kgCO₂e per image, about six times the footprint of a text query.

4. Monetary fallback, as a last resort. If none of the above data is available, Oakbon uses your annual spend on AI subscriptions and APIs, at a rate of 60 kgCO₂e per €1,000 excl. VAT.

An estimate shown separately, never added to Scope 3

Each of these factors carries high uncertainty, from 70 to 80% depending on the source, compared with much tighter margins for official ADEME factors. Silently adding them to the Scope 3 total would degrade the reliability of the whole footprint.

Oakbon therefore makes a deliberate methodological choice: the AI footprint is calculated, shown with its uncertainty range, and positioned alongside the Scope 3 total, never inside it. The report states the order of magnitude this represents, without ever adding it to the total. Wording used in the report: “Estimated factors (digital LCA literature, high uncertainty), in the absence of an ADEME factor dedicated to AI inference.”

Why measure it anyway

The value of this line item isn’t in its size, often modest compared with the rest of Scope 3, but in what it prepares you for.

Transparency, first: more and more carbon questionnaires and ESG audits are starting to ask about the use of digital and AI tools. Being able to answer with a figure, even an uncertain one, beats an empty box.

Anticipating a future ADEME factor, next: the day Base Carbone publishes an official factor for AI inference, a plausible development given how fast the database keeps expanding each year, you’ll already have your usage history ready to be recalculated with a more precise factor, without starting from scratch.

Internal steering, finally: knowing the order of magnitude lets you guide digital sobriety choices without waiting for a regulatory obligation.

Three levers for AI sobriety

Right-size the model to the task. A lightweight model is enough for most everyday uses; reserve the heaviest models for tasks that genuinely justify them.

Refine your prompts. Grouping requests and clarifying the instruction on the first try limits unnecessary back-and-forth.

Reserve image generation for genuine needs. A generated image weighs about six times as much as a text query.

In summary: your AI footprint checklist

1. Generative AI usage falls under upstream Scope 3, category 1, purchased digital services.
2. No official ADEME factor exists yet for AI inference: Oakbon relies on available digital LCA literature.
3. The method always retains the most precise data: actual queries, then licenses, then monetary fallback.
4. Queries and licenses are never added together: they measure the same usage.
5. Image generation weighs about six times as much as a text query.
6. This estimate stays separate from your “core” Scope 3, shown alongside it, never inside it.

Ready to measure your AI usage too?

The AI Footprint module is built into the Oakbon questionnaire, to switch on if your company uses generative AI tools.

From 69€/month · No commitment