SOMA

    How do I include SaaS AI tools like ChatGPT Enterprise in my Scope 3 inventory?

    For sustainability and procurement teams adding a seat-licensed AI tool to a Scope 3 inventory. Sources checked on 12 November 2026. SOMA figures are estimates. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Page updated

    How do I include SaaS AI tools like ChatGPT Enterprise in my Scope 3 inventory?

    Book them in Scope 3 Category 1, purchased goods and services, as one inventory row per tool. Estimate the row from seats, then messages, then tokens. If you can only get the invoice, use spend as a labelled screening figure. For 200 active seats on a mid-size model served from US East - N. Virginia, the SOMA AI estimate is 4.02 kg CO₂e for 12 months, with a band of 2 to 9.2 kg.

    Which category does a SaaS AI tool belong in?

    Category 1 of the GHG Protocol Scope 3 Standard, because your company buys a service that a provider delivers. It is not Scope 2, since the provider buys the electricity. The ESRS E1 position, including why the cloud subset of Category 1 is relevant, is on the ESRS E1 page.

    What data do I take from the tool, and how do I turn it into tokens?

    Admin data by tool, as the publishers describe it. Check your own console, since exports change.
    ToolWhat the admin side givesTier it supports
    ChatGPT Enterprise and EduWorkspace analytics with a per-user CSV export that includes whether a user was active and the total messages sent. No tokens.2b, from messages
    Microsoft 365 CopilotThe Copilot usage report gives enabled and active users and, per user, prompts submitted, with a CSV export. No tokens.2b, from prompts
    Other seat toolsAsk the vendor what its admin console exports. Keep the dated export as evidence.2b or 1

    The conversion from messages to tokens is an assumption you declare. The dataset takes 928 input and 282 output tokens per exchange, so 1210 in total, and multiplies output tokens by 2.20 for reasoning. Where an export has no message count, the dataset falls back to 30 messages per active enterprise seat per month, with a range of 20 to 65, and 325 for heavy use. Count active users, since unused licences send nothing. The seat benchmark and its sources are explained on the seat footprint page.

    What does a worked example for 200 seats look like?

    Tier 2b estimate for 200 active seats over 12 months, class B model. Kilograms of CO₂e. Estimates, computed from the open dataset.
    CaseRegionMessagesCentralBand lowBand high
    Default, 30 per seat per monthUS East - N. Virginia72,0004.0229.2
    Default, same usageEU North - Sweden72,0000.9980.52.3
    Heavy, 325 per seat per monthUS East - N. Virginia780,00043.522100

    Steps. 200 seats times 30 messages times 12 months gives 72,000 messages. Each message becomes 928 input tokens and 282 output tokens with the reasoning uplift on output, and the tokens are priced with the class B factor for the region. The provider does not publish which model or region serves your seats, so both are assumptions to record. The row is a Tier 2b estimate and the band does not cover the choice of class or region.

    What if I only have the invoice?

    Multiply spend excluding VAT by 0.09613 kg CO₂e per EUR. That is the EXIOBASE 3.8.2 — Computer and related services factor of 0.1181, deflated to 2019 prices, with an uncertainty of ±300%. We give no worked number here, because ChatGPT Enterprise has no public list price and an invented one would set the result. For a seat product the screening figure overstates AI use, since the subscription pays for software, storage, support and margin as well as inference. For API purchases, where the invoice tracks tokens more closely, the method paper behind these factors, accepted for publication in Sustainable Production and Consumption, finds spend-based figures 1.7 to 12.3 times a token-based estimate on deflated spend. Treat spend as a screening figure on a different boundary, never as an upper bound, and replace it with the seat estimate once you have the export. Do not relabel a whole subscription as AI unless a separate add-on price evidences the AI share, and remove any spend already covered by a token or seat line so the purchase is counted once.

    What goes in the inventory row?

    • Tool, vendor, contract period and annual spend.
    • Active seats and messages, with the date of the export they came from.
    • Tier, model class and region assumed, and why.
    • Factor version, central value, band and the word estimate.

    The full row format and the review questions are in the Scope 3 guide, and the equations are on the methodology page. The free calculator runs the seat estimate. SOMA AI publishes both the factors and this page.

    Keep reading