CSRD, ESRS E1, and why LLM emissions belong in your climate disclosure

The Corporate Sustainability Reporting Directive (CSRD) and the European Sustainability Reporting Standards (ESRS) push companies toward verifiable, granular climate data—not a single slide that says “we care about green AI.” For product and platform teams, the implication is clear: generative AI and LLM usage becomes a data problem, not only a communications problem.

From CSRD to ESRS E1

CSRD requires in-scope undertakings to publish a sustainability statement built on ESRS. Among environmental topics, ESRS E1 — Climate change is where most finance and sustainability teams first encounter hard requirements: not intentions, but greenhouse gas (GHG) emissions, targets, transition plans, and risk disclosures—presented in a way that is comparable over time.

Methodologically, ESRS E1 aligns with what carbon practitioners already use: the GHG Protocol framing of Scope 1 (direct), Scope 2 (purchased energy), and Scope 3 (other indirect emissions across the value chain). The reporting challenge is not the physics—it is traceability: scope boundaries, emission factors, and evidence that an auditor can follow.

Where LLM inference sits: usually Scope 3

For most organizations consuming third-party models or cloud APIs, inference workloads fall under Scope 3—purchased services and upstream value chain emissions—not on-site fuel combustion. That does not make them immaterial: for software vendors and AI-heavy products, token volume grows fast, models differ wildly in intensity, and customers increasingly ask for per-tenant or per-product footprints.

Regulatory expectations and stakeholder pressure are converging: ballpark estimates (“our AI is roughly one short flight a month”) age poorly when usage is recurring, multi-model, and multi-tenant.

Double materiality: two roles for LLMs

CSRD is built around double materiality: impacts on people and planet, and sustainability-related risks and opportunities affecting enterprise value. That lens separates two very different “AI stories”:

• AI as a sustainability topic: the climate footprint of your product’s model calls, governance, reduction roadmap.
• AI as a reporting tool: LLMs helping draft ESRS-aligned narrative disclosures—often a large share of the workload because much sustainability content is qualitative.

carbon-llm focuses on the first: instrumenting measurement with metadata-only API calls so prompts never leave your infrastructure, while you still produce auditable activity data for footprinting.

From “average AI” to activity data

ESRS E1 does not prescribe a single formula for “grams of CO₂ per LLM call,” but it rewards consistent methodologies, documented factors, and disclosures that tie numbers to management systems. In practice, teams move from:

• one-off spreadsheets updated quarterly, to
• continuous metering: tokens, model id, environment (test vs production), and time range—fed into coefficients with stated confidence and sources.

That is the same pattern as classic Scope 3 category work—only the activity driver is inference instead of kilograms shipped.

What ISVs should capture

If you expose LLMs to customers, you typically need attribution at least at tenant and model granularity, with a clear audit trail:

• Activity data: prompt and completion token counts from the provider’s usage payload.
• Model identity: stable model keys so coefficients map cleanly.
• Emission factors: LCAs or vetted estimates—versioned and cited.
• Privacy: metadata-only pipelines so customer prompts are not copied to a third party.

A REST /v1/track style integration keeps your application code almost unchanged while feeding the inventory your CSRD process will ask for.

Practical roadmap

1. Map every production path that calls an LLM (direct APIs, gateways, batch jobs).
2. Record tokens and model identifiers automatically—test vs live separated.
3. Apply documented coefficients; keep references and revision history.
4. Allocate to customers or business lines where materiality requires it.
5. Feed the totals into your climate narrative, targets, and transition planning—not as a siloed “green IT” line.

Bottom line

ESRS E1 does not create a magical “LLM CO₂” line item in law—it demands a serious GHG disclosure for climate-related impacts. For AI-heavy products, the winning pattern is simple: what is metered and traceable can be defended in assurance; what stays a hand-wavy average cannot.

If you already ship LLM features, the productive move is not another sustainability slogan—it is wiring measurement into the same code path as billing and analytics.