What is the carbon footprint of Microsoft 365 Copilot, ChatGPT Enterprise and other AI seats?

We found no provider that reports emissions per seat, so a company has to estimate. For 250 active seats sending the dataset default of 30 messages per seat per month for 12 months, the SOMA AI (somaai.earth) Tier 2b estimate is 5.02 kg CO₂e, with a band of 2.5 to 12 kg. That assumes a class B model (mid-size) served from US East - N. Virginia. Those two assumptions are declared by the company doing the estimate. This page does not know which model or region Microsoft uses for Copilot, and neither does the licence.

Not as a separate figure, on the evidence we found. A Microsoft Q&A answer states that the Microsoft 365 Emissions Impact Dashboard does not include Copilot for Microsoft 365, Copilot agents or other integrated AI services, and that the power is part of the aggregate footprint but not yet partitioned or exposed as a line item. The source is a Microsoft Q&A page, not product documentation, so check the current documentation before relying on it. The Azure dashboards cover an Azure customer's cloud consumption, which is a different thing from employees using Copilot. The tools comparison lists what each tool covers.

• Microsoft 365 Copilot and GitHub Copilot. No token export. The admin centre gives assigned seats and active users.
• ChatGPT Enterprise or Team. The admin dashboard gives seats and message counts, as the Scope 3 guide sets out.
• Claude for Work. The admin view reports seats and activity, as the Scope 3 guide sets out.
• Gemini for Workspace. A seat-licensed product of the same kind. We have not checked what its admin console reports, so ask your administrator.

Seats and active users, or message counts where they exist, are Tier 2b data in the Scope 3 guide. Keep the export itself, with its date, as the evidence behind the figure. Check your vendor's current admin documentation for what it exports.

Seats times messages per active seat per month times months gives messages. Each message is converted to an assumed input and output token count, the output tokens carry the reasoning uplift of 2.20, and the tokens are multiplied by the class and region factors. The dataset's default is 30 messages per active seat per month, with a low of 20 and a high of 65. The dataset names one source for it, the UK Department for Business and Trade pilot of 1,000 Microsoft 365 Copilot licences. See the published evaluation. A heavy-user setting of 325 messages comes from OpenAI's definition of power users in its workspace analytics, and its low end of 115 from a UK government coding assistant trial of GitHub Copilot. Count active users, not licences, since licensed seats that are never used send no messages. The reasoning uplift is explained on the reasoning page.

The band is the dataset's uncertainty band on tokens. It does not cover the choice of class or region, which is why the three rows differ more than the band does. Pick the class by the nearest published model and say so.

• The messages per seat are a benchmark from other organisations, not your own usage. Replace them with the count from your admin centre where it gives one.
• Copilot actions are not necessarily chat messages. The conversion to tokens is a general business-use assumption.
• The class and region are declared, not known, for a vendor that does not disclose them.
• The reasoning uplift is an author assumption, since no public source gives the multiplier.
• The method paper is under peer review.

Report the seat line as a Tier 2b estimate. State the number of active seats, the messages per seat and where each came from, the model class and region assumed, and the factor version. Give the band beside the central value. If a vendor later publishes real usage or an emissions figure, switch to it and note the change. The free calculator runs this estimate, and the Scope 3 guide places it in an inventory next to API usage. The methodology page has the equations.

Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

• What is the carbon footprint of one ChatGPT prompt? Per-prompt figures and why the region decides
• AI carbon footprint tools compared. What each tool covers and what it leaves out
• Why AI footprint estimates differ: published figures, dated. Published figures with their boundaries