Scope 3 Category 1 vs Category 11 for AI / LLM services

A frequent question in CSRD / ESRS E1 workflows is: “Do our AI/LLM inference emissions go into Scope 3 category 1 or category 11?” The honest answer is: it depends on boundaries — who controls the activity, and what you consider part of the value chain for the reporting entity.

Scope 3 categories are boundary decisions

Emission factors and token-to-CO₂e math can be consistent. What changes between category 1 and category 11 is the accounting frame: the “who pays” and “what is sold / purchased” perspective your auditor expects you to apply.

Category 1 (purchased goods and services): where it often fits

Category 1 is commonly used when LLM inference is treated as a purchased serviceconsumed by your organization (or for your organization’s internal operations). Practically, if you:

• buy API calls from third parties as an input to your services, and
• report the resulting footprint within your reporting entity’s climate disclosure,
• use token-based activity data mapped to supplier factors (or documented estimates),

then a Scope 3 category 1 narrative is often the simplest starting point for LLM carbon measurement.

Category 11 (use of sold products): when AI may shift the conversation

Category 11 becomes relevant when the emissions occur in the use phase of your sold products. For AI-heavy vendors, that can mean your customers run inference as part of using your product/service — and you need to decide whether those inference emissions are “in the product’s use” boundary your reporting frame covers.

Note: this isn’t an automatic switch. Some organizations still treat inference as a purchased service (category 1), even if the output is customer-facing, because the reporting entity boundary and consolidation logic matter more than the marketing story.

A practical decision framework

When you need to be consistent (and defensible in assurance), document answers to these questions:

• Financial control: who controls the inference activity and its contractual responsibilities?
• Reporting perimeter: are you reporting a single legal entity or a group consolidation?
• Materiality: is inference footprint material enough to require auditor-level consistency?
• Operational attribution: do you attribute tokens per customer/tenant using a stable mapping?
• Assumptions transparency: what do you assume when providers do not provide direct supplier data?

What helps your auditor (even before the “final” category)

Auditors usually care less about your first guess and more about whether you can show a coherent boundary logic and traceable activity measurement. If your token-to-CO₂e computation is repeatable and your coefficient provenance is documented, you can iterate on category decisions without rewriting everything.

Further reading (selected)

For deeper framing, use the GHG Protocol Scope 3 guidance as the base reference and align your interpretation with your own consolidation policy. This post is educational; discuss the final classification with qualified advisors and your auditor. See also: auditor evidence checklist for LLM emissions.