Scope 3 and LLM usage: what to measure first for CSRD-ready data

As EU sustainability reporting and customer due-diligence questions mature, generative AI spend and usage are showing up next to travel and cloud in Scope 3 conversations. The goal is not to chase every chat — it is to build defensible activity data where AI is material to your business.

Start from materiality, not from tools

Ask whether LLM calls are customer-facing, embedded in products, or internal productivity. The same API usage might map to different categories (e.g. purchased services vs use of sold products) depending on how you define financial control and boundaries — a theme we cover in Category 1 vs 11 for AI.

Activity data you can actually collect

The most audit-friendly path is usually provider usage records: prompt and completion tokens per model, per project or tenant, with timestamps. That aligns with coefficient-based CO₂e methods (see tokens → CO₂e). Spreadsheets of “we think people use ChatGPT a lot” do not survive assurance.

Connect to climate disclosures

ESRS E1 and GHG Protocol alignment still expect clear scopes and methodologies. LLM inference often lands in Scope 3; the important part is consistency with your other emissions data and a narrative that matches your transition story. Our CSRD & ESRS E1 primer walks through the framing.