Where AI usage sits in the GHG Protocol's draft Scope 3 data types, and what to report today

The 2011 Scope 3 Standard is the final text and remains the one to report against. Everything about data types and a 95% boundary is proposed working draft. The GHG Protocol's March 2026 Phase 1 Progress Update is headed "all content is draft and subject to change", says its proposals have not been approved by the Independent Standards Board, and asks for comments to be held until public consultation. The Standard Development Plan of 29 July 2026 folds the Corporate Standard, Scope 2 Guidance and Scope 3 Standard into one consolidated standard, expected to be co-branded with ISO, with a consultation draft estimated for Q2 2027 and publication estimated for Q4 2028. The plan itself says that timeline may change.

Category 1, purchased goods and services, because an AI API or chat seat is a service your company buys from a third party. The GHG Protocol's Scope 3 Technical Guidance sets the category boundary at cradle-to-gate, which is why a factor covering serving electricity alone falls short of it. The revision does not change this placement so far. The Progress Update does not mention AI or cloud services, and its proposed Category 16 is aimed at other value chain activities such as facilitated emissions and licensing, where the company earns income from an activity it never buys, sells or owns. Purchased inference is a plain purchase. Required boundaries per category are still under review in the working group's second phase, so this reading should be checked again when the consultation draft appears.

ESRS E1 adds one sub-line. The revised ESRS E1, adopted on 3 July 2026 and published as Delegated Regulation (EU) 2026/1563, says at AR 24(g) that the undertaking shall "if applicable, disclose GHG emissions from purchased cloud computing and data centre services as a subset of the overarching scope 3, category 1", and its AR 26 table has a row for it. The 2023 text had the same subset at AR 51, conditioned on materiality. Neither text names AI. Treating metered AI inference as purchased cloud computing is our reading, a sensible one, and worth stating in the method note rather than assuming.

Two ESRS details matter for a consultant. AR 24(a) tells the undertaking to consider the Scope 3 Standard "Version 2011", so a draft GHG Protocol tier has no standing in an ESRS statement yet. And the delegated act applies to financial years beginning on or after 1 January 2027, with optional use for 2026 years, to a population that Omnibus I narrowed to companies with more than 1,000 employees and more than €450 million in turnover. A client below those thresholds that still wants an AI line is doing so by choice, for customers or lenders for instance. Gross emissions sit at disclosure requirement E1-8 in the revised numbering.

Not one to one, and the top of the draft ladder is out of reach for most AI lines today. SOMA's tiers sort services by the data a preparer can get. The draft sorts emissions by how specific both the activity data and the emission factor are. The two questions overlap without being the same, so Tier 2a, the exact billed token count, lands in the middle of the draft rather than at the top. Annex A of the Progress Update holds two competing options for the classification, and neither has been chosen. The table shows where each SOMA tier would fall under each option, alongside the method names in force now. This is SOMA's own reading of a working draft. The GHG Protocol has published no mapping for AI.

Both options also have an Unclassified tier for a company unable or unwilling to split its emissions, although under Option 2 that tier is still subject to further consideration. Figures are expressed in mass units, not percentages. Tier 1 is the one clean match. The development plan lists "phasing out, limiting, or removing the spend-based method" among the things to consider, while the working draft keeps it as a visible tier, so a spend-based AI line would be labelled rather than banned on current evidence.

Because the factor is not specific to the supplier. Option 1 asks for both specific activity data and specific emission factors, and summarises the test as physical or functional units, representativeness, and specificity to the sites where the emissions occurred. Billed tokens pass the activity side. A SAEF factor is a class mean from public benchmarks of open-weight models combined with a grid average for a region, so it fails the site test by design. Moving up requires something that does not exist yet, a provider's own allocated figure for your account. Under Option 2 the answer depends on one bracketed word, "[representative]", which the draft has not defined. If the final text accepts a documented class and region factor as representative, Tier 2a would count as primary. That is a question to raise at consultation, not an answer to assert today.

None of this says Tier 2a is poor data. The Progress Update states that "data type is not by itself an indicator of data quality or data accuracy", and the Technical Guidance makes the same point in Box 1.1, warning that supplier data can be less accurate than industry averages once allocation is involved. A proposed restriction sharpens that. Revision A8 would permit corporate-level allocation for homogeneous suppliers and bar it for diversified ones, so a large provider's company-wide footprint divided by revenue would not earn a specific label under the draft.

Evidence that lets someone else rebuild the number and check the tier you claimed. The list below is SOMA's recommendation, built from what the 2011 Standard requires you to describe for each category, namely data types and sources, data quality, methodologies, allocation and assumptions, plus what the draft would add.

• Tier 3. The supplier's statement, its boundary, its assurance opinion, and how your share was allocated.
• Tier 2a. The dated usage export, tokens split into uncached input, cached input and output, the model identifier, the serving region and how you know it, and the factor version soma-ai-ef-2026.10-dataset-v3.0.0.
• Tier 2b. The admin report with seats or messages, each default used in place of missing data, the use case assumed, and whether reasoning modes are in use.
• Tier 1. The invoice, the factor and its licence, your dated request to the vendor for usage data and the reply, and a note that the line covers the whole subscription.
• Every tier. Who checked that the list of AI services is complete, and the boundary with its named exclusions.

Check the factor itself against the proposed Revision A5, which recommends factors with no more than 5% cut-off, an uncertainty assessment, and information on completeness, validation and verification level. SAEF publishes its boundary, stated exclusions and uncertainty, with a class range of ±39% for class B at the world-average grid. It does not certify a cut-off percentage, has no third-party verification, and its method paper is under peer review. Say so in the note and let the assurance provider weigh it.

A hypothetical company has three AI services in one reporting year, each at a different tier. Line A is an API on a cloud platform with the region pinned to EU West - Ireland, 240 million billed tokens on a mid-size model at the dataset's default input and output mix. Line B is 60 active enterprise chat seats with no token export and an EU-only processing commitment, so the EU-27 average row applies, at the published default of 30 messages per active seat per month and 1,210 tokens per exchange. Line C is a SaaS tool with embedded AI, where the vendor holds the relationship and the invoice is all you have, €12,000 for the year.

Line B carries an activity band of -50% to +130% on its estimated tokens, which gives 0.49 to 2.24 kg CO₂e. Line C carries ±300% and is an upper bound on the AI part, since it prices the whole subscription. Under the draft's disaggregation the annex table would read 0.0069 t in Other and 1.4172 t in Spend-based under Option 1. The picture this reveals is the point of the proposal. About 99.5% of the AI footprint in this inventory sits in the line with the least information behind it, and the upgrade path for next year is obvious. Write the vendor a usage request now.

No, and it is worth correcting because the claim circulates. The 2011 Standard already says companies "shall account for all scope 3 emissions and disclose and justify any exclusions", so silent omission is already non-conforming. The proposed Revision B1 gives a numeric allowance, letting a company exclude a category or activity as long as total exclusions stay within 5% of total required Scope 3. Revision B5 adds de minimis emissions inside that cap, and its own example is Category 1 emissions from paper clips and staples. What the draft does add is a duty to know the size. Revision B2 would require quantifying all required Scope 3 every year to show the exclusions fit, by any method, and Revision B4 keeps the duty to disclose and justify them.

Apply that to the example. In a company with a 5,000 t CO₂e Scope 3 total, a hypothetical figure, the three lines above come to 0.028% of the inventory. Under the draft that company could exclude AI and still conform, provided it had a number showing the exclusion fits and said so. The practical conclusion is that some quantification is needed either way, and a rough Tier 1 or 2b screen is enough to support an exclusion. Reporting the line rather than excluding it is a choice made for other reasons, such as a customer questionnaire or a target that covers digital services.

Report under the final text and prepare cheaply for the draft. Six steps cover it, none of which depends on how the classification is settled.

1. Place the AI lines in Category 1 and, under ESRS, in the cloud computing and data centre sub-line if the service fits your reading of it.
2. Assign a SOMA tier per service, following the steps in the Scope 3 accounting guide.
3. Tag each line with its method name under the 2011 guidance and, in a working paper, with both draft options. The tag costs one column now and saves a reclassification later.
4. Keep totals in mass units per type, so a disaggregated annex table can be produced on request.
5. Record the factor version from the SOMA AI Emission Factors table, so the figure can be rebuilt after the factors change.
6. Set a data specificity goal for the AI lines, for example moving spend-based lines to token exports, which is the kind of goal proposed Revisions A6 and A7 recommend.

Avoid two statements. Do not claim conformance with the revised standard, which does not exist yet, and do not describe a draft tier as a GHG Protocol classification in a public report. "Prepared under the Scope 3 Standard (2011), with data types tracked against the working draft of March 2026" is accurate and will date gracefully.

SOMA turns providers' usage exports into a Scope 3 Category 1 entry with the tier, factor, version, sources and uncertainty attached, ready for an assurance provider to read. If two published numbers disagree in front of a client, start with why AI footprint estimates differ. Write to lili@somaai.earth.

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.

• How to account for AI usage in your Scope 3 inventory under CSRD. The full method, step by step
• Methodology. Tier definitions, boundary and uncertainty
• SOMA AI Emission Factors table and download. The factors used in the example