How Companies Track Scope 3 Emissions From AI Usage
On this page
As AI becomes a standard line in every workflow, it quietly becomes a line in your carbon accounting too. For most organizations that line sits in Scope 3, the hardest scope to measure and the one regulators are paying the most attention to. This guide explains where AI usage fits, why it is tricky to track, and how to turn it into auditable data.
Scope 1, 2 and 3 in plain terms
The Greenhouse Gas Protocol splits emissions into three scopes:
- Scope 1 is direct emissions from sources you own or control, such as company vehicles or on-site combustion.
- Scope 2 is indirect emissions from the energy you purchase and consume, mainly electricity for your own operations.
- Scope 3 is everything else in your value chain, both upstream and downstream, from purchased goods and services to the cloud and SaaS tools you use.
Scope 3 is usually the largest and the least visible, because it lives in other companies' operations.
Why AI usage is Scope 3
When your application calls a hosted AI model, the electricity is burned and the cooling water evaporated in a third party's data center. You did not buy that power directly, so it is not Scope 2; it is a purchased service, which puts it in Scope 3. The footprint is real, though: AI workloads are estimated to draw on the order of 200 TWh per year, with data-center cooling consuming millions of liters of water daily, and projections of roughly 300 million tons of AI-related CO2 by 2030. As your AI usage grows, so does this slice of your Scope 3 total. The difference between offsetting it and going further is covered in carbon neutral vs carbon negative AI.
The measurement challenges
The core problem is data. Most AI providers do not expose the footprint of an individual request, so companies are forced to estimate from token counts and public averages, which vary widely by model and region. Even good estimates require knowing the model class, since energy per token ranges from around 0.0008 Wh for a small model to roughly 0.0038 to 0.0040 Wh for frontier and reasoning models, multiplied by data-center overhead (PUE near 1.56) and converted with grid intensity (around 0.395 kg CO2 per kWh on a US average). The full method is on our How it Works page, and the per-model detail in AI energy usage by model.
The shortcut: instead of reconstructing AI emissions after the fact, use a provider that measures every request at the source and lets you export it. That removes the biggest source of error in Scope 3 AI accounting.
Per-seat and per-team attribution
Reporting a single company-wide AI number is a start, but most organizations need to attribute it, to a team, a product, a customer or an individual seat, the same way they attribute software spend. That requires per-request data tagged to a user or API key. With that in place, you can answer questions like which team's AI usage is growing fastest, or which feature carries the heaviest footprint, and act accordingly. Ecoia rolls per-request carbon, water and energy up per seat and per team, as part of green AI for business.
CSRD, SB-253 and exportable data
Regulation is making this concrete. The EU's CSRD drives large companies toward detailed value-chain disclosure, including Scope 3. California's SB-253 requires large companies doing business in the state to report Scope 1 and 2, with Scope 3 to follow. To satisfy auditors you need data that is granular, traceable and exportable, not a one-line annual estimate. A measured AI platform produces exactly that: per-request figures you can export into your carbon accounting and offset certificates you can attach to audits and investor decks.
FAQ
Why is AI usage counted as Scope 3 and not Scope 2?
Scope 2 covers electricity your own operations purchase. When you call a third-party AI API, the electricity is consumed in someone else’s data center, so it falls under Scope 3, the indirect emissions in your value chain. The exception is if you run models on infrastructure you control, which can shift parts into Scope 1 or 2.
What makes AI emissions hard to track?
Most AI providers do not report the footprint of individual requests, so the data simply is not there. Energy per request also varies by model, region and load. Without a provider that measures and exports per-request figures, companies are left estimating from token counts and public averages.
Do regulations actually require this?
Increasingly, yes. The EU’s CSRD pushes large companies toward detailed value-chain (Scope 3) disclosure, and California’s SB-253 requires large companies doing business in the state to report Scope 1, 2 and eventually 3 emissions. AI usage is a growing part of that Scope 3 picture.
How can we attribute AI emissions to teams or seats?
You need per-request data tagged to a user, key or team. Ecoia tracks carbon, water and energy per request and can roll it up per seat and per team, so you can attribute the footprint the same way you attribute software cost, and export it for reporting.