SOMA

    How should AI and cloud services be reported under ESRS E1 for CSRD?

    For sustainability and finance teams of companies inside the CSRD scope. Regulatory position checked on 12 November 2026, not legal advice. 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 should AI and cloud services be reported under ESRS E1 for CSRD?

    As part of Scope 3 gross emissions, in the category where the purchase sits, which is Category 1 (purchased goods and services) for AI APIs, enterprise AI seats and cloud services. The revised ESRS E1 asks for Scope 3 emissions from each significant category in paragraph 30, and application requirement AR 24(g) asks, where applicable, for emissions from purchased cloud computing and data centre services as a subset of Category 1. The text does not name AI, so placing AI services in that subset is an interpretation, and a company should say so. Disclose the method behind the line, the factor version and the uncertainty.

    Which rules apply, and to whom?

    The legal chain as of 12 November 2026. Each row links to the official text.
    InstrumentWhat it says that matters hereApplies from
    Directive (EU) 2026/470 (Omnibus I)CSRD reporting only for undertakings with more than 1,000 employees on average and more than EUR 450 million net turnover. Both conditions apply.Financial years starting on or after 1 January 2027
    Delegated Regulation (EU) 2026/1563, adopted 3 July 2026The simplified ESRS, including E1 with paragraph 30 and AR 24. In force 10 November 2026.Financial years starting on or after 1 January 2027, with an option for 2026 if the version is stated
    Commission announcement, 3 July 2026Adoption of the revised ESRS and of a voluntary standard for companies outside the mandatory scope.Not applicable

    A company below the thresholds has no ESRS duty to report AI, and an AI line in its inventory is voluntary or a response to a customer or lender. The same method serves both cases. The full timeline, the Spanish transposition position and the voluntary standard are on the Scope 3 and CSRD guide, so this page does not repeat them.

    What should the disclosure contain?

    ESRS E1 does not prescribe an AI method, so the disclosure has to carry enough for a reviewer to rebuild the number. Four items do most of the work.

    1. Tier and method. Say which of four data tiers produced each line. Tier 3 is a provider figure for your account, Tier 2a is billed tokens, Tier 2b is messages or seats and Tier 1 is spend. Several tiers in one table is normal.
    2. Factor version and source. Name the dataset and pin one version for the reporting year. SOMA AI uses the open SAEF dataset, version soma-ai-ef-2026.11-dataset-v3.1.0, and treats a change of version as a methodology change.
    3. Uncertainty. A token-based line carries a class range of ±48% to ±67% at the world-average grid by model class. A spend-based line carries ±300%. Print the band in the same row as the figure.
    4. Boundary and exclusions. Inference, the share of server hardware and amortised training are in. Provider offices, network transport and your own devices are out. Image, speech and video models are outside the current factor version.
    How AI and cloud sit in a Scope 3 table. The cloud line is the AR 24(g) subset of Category 1.
    PurchaseCategoryTypical dataTier
    Cloud compute and storageCategory 1, cloud and data centre subsetProvider carbon report for your accountProvider data, or Tier 1
    AI APIs with usage billingCategory 1, same subset by our readingTokens by model and region from the console2a
    Seat-licensed AI assistantsCategory 1, same subset by our readingActive seats and messages from the admin export2b
    SaaS with embedded AICategory 1Invoice only, AI share not separable1

    Why avoid double counting between cloud and AI?

    A company that runs an AI API through its own cloud account may already have those emissions in the provider's customer carbon report. Count each purchase once. If a token-based line covers a service, remove its spend from any spend-based line, and if a provider report already includes the AI workload, do not add a token estimate on top. Record which source was used in the inventory row.

    What is still open?

    • Whether an assurance provider accepts an estimate-based AI line as significant or immaterial is a judgement for your materiality assessment.
    • The ESRS text names cloud computing and data centre services, not AI, so the AI reading is ours.
    • The GHG Protocol data tiers are under consolidation and may change.

    The method paper behind the SAEF factors was accepted for publication in Sustainable Production and Consumption in November 2026, and the equations are on the methodology page. SOMA AI publishes both the dataset and this page.

    Keep reading