CSRD AI Emissions Reporting: A Practical Step-by-Step Guide for Sustainability Teams

If you are a sustainability or ESG professional preparing your organisation's CSRD disclosure, you have almost certainly discovered that AI emissions are required under ESRS E1 — and that your AI vendors provide none of the data you need to calculate them.

This guide covers what the reporting obligation actually requires, why major AI vendors have no data to give you, three approaches to gathering emissions figures yourself, and what auditable disclosure looks like at each quality tier.

For an overview of how CSRD intersects with AI usage more broadly, including the ESRS E1 scope and which organisations are affected, see our guide to CSRD AI carbon reporting requirements.

What ESRS E1 Actually Requires You to Report

CSRD's climate standard — ESRS E1 — requires disclosure of Scope 1, 2, and 3 greenhouse gas emissions. Scope 3 covers indirect emissions in your organisation's value chain. Category 11 of the GHG Protocol Scope 3 standard covers "Use of sold products" — which the GHG Protocol technical guidance has extended to include the operational use of digital services and SaaS tools purchased by your organisation.

The AI-specific obligation follows directly: your employees use AI tools, those tools run in data centres, those data centres consume electricity, that electricity generates CO₂e. The resulting emissions are a Scope 3 Category 11 liability your organisation must measure and disclose.

The reporting obligation applies regardless of whether AI usage is centralised on a single enterprise platform or distributed across individual employee subscriptions. Shadow AI — personal ChatGPT, Copilot, or similar subscriptions used for work — is within scope. A complete disclosure requires knowing how much AI your organisation is actually using, which starts with a shadow AI audit. For a framework for discovering and governing shadow AI usage, see our analysis of the shadow AI governance gap.

Why Your AI Vendors Have No Data to Give You

The fundamental problem: AI vendors process millions of queries on shared hardware, in shared data centres, across hundreds of thousands of customers simultaneously. Attributing a specific fraction of data centre energy consumption to a specific customer's queries requires per-customer metering at the inference level — infrastructure none of the major vendors have built.

OpenAI, Microsoft Azure OpenAI, Google Gemini, and Anthropic Claude all provide API-level token counts. None translate those token counts into CO₂e figures in their reporting dashboards. As of Q1 2026, no major AI platform provides a carbon data export suitable for CSRD disclosure.

This gap is expected to close as regulatory pressure increases — the EU AI Act and CSRD together create strong incentives for vendors to build emissions reporting. In the interim, the gap is yours to bridge.

Three Approaches to AI Emissions Data — Ranked by Auditability

Sustainability teams currently have three options for obtaining AI emissions figures. They differ substantially in cost, time, accuracy, and audit defensibility:

Approach 1: Manual Estimation Using Academic Energy Models

The methodology for manual estimation:

1. Retrieve total token counts by model for the reporting period from your API usage dashboards
2. Apply published energy-per-token figures — Patterson et al. (2022, Google) and Luccioni et al. (2023, Hugging Face) are the standard references for frontier models
3. Identify the data centre regions used by each AI platform — this determines grid carbon intensity (g CO₂e per kWh), sourced from national grid emissions factors
4. Multiply energy consumption by grid carbon intensity to produce CO₂e
5. Apply ±30–50% uncertainty bounds, since data centre PUE (Power Usage Effectiveness) varies significantly between providers and facilities

This process takes 2–4 weeks and costs €5,000–€20,000 in sustainability consultant fees. The resulting disclosure will carry wide confidence intervals. Auditors reviewing CSRD reports increasingly flag estimated Scope 3 figures without measurement methodology documentation — your disclosure must include a detailed methodology note explaining every assumption.

Approach 2: Platform-Level Native Tracking

The most defensible approach: deploy an AI platform that tracks carbon emissions natively, per token, per model, per team, from every query. Govarix does this automatically. Every interaction generates a logged token count, applied against the model's published energy profile, producing a CO₂e figure stored in the carbon ledger at the moment of inference.

This produces a measured disclosure rather than an estimate. The audit trail is the query log itself — every query, model, token count, timestamp, and resulting CO₂e entry is available for auditor review. Uncertainty is under ±5%, reflecting only variation in grid carbon intensity rather than estimation of energy consumption.

Step-by-Step: The Five-Step AI Emissions Reporting Process

Whether you are using manual estimation or platform-level tracking, the reporting process follows five steps:

Model Carbon Intensity Variance: An 8× Range

Model choice is a carbon decision. The variance between the most and least carbon-intensive commonly deployed models is approximately 8× per query — a material factor that sustainability teams should communicate directly to AI procurement and governance functions.

At 1 million AI queries per month, shifting 50% of routine tasks from a large frontier model to a small efficient model reduces monthly emissions by approximately 1.4 tonnes CO₂e. For a 250-employee organisation, that is a reduction of around 16.8 tonnes CO₂e annually — a material sustainability programme contribution achievable through governance policy alone, with no reduction in AI capability for complex tasks.

Govarix displays per-model carbon intensity in real time. Admins can set model routing rules that direct simple tasks to efficient models by default, with large models available on request. The carbon impact of routing decisions appears in the carbon ledger immediately. For the full framework for AI cost and model governance, see our guide to enterprise AI cost control and token budgets.

Real-World Equivalents for Stakeholder Communication

Raw CO₂e figures communicate poorly to board members, investors, and employees. Translating AI emissions into real-world equivalents makes the numbers visible and comparable to existing sustainability metrics your stakeholders already understand:

Govarix's ESRS E1 export includes these equivalents automatically, so a quarterly board sustainability update can present "our AI usage in Q1 was equivalent to 12,400 km of car driving" alongside the formal CO₂e figure — without additional manual calculation.

What Your ESRS E1 Export Should Contain

A complete AI emissions disclosure for CSRD purposes requires more than a single CO₂e figure. Each component has a specific purpose in the auditor review process:

• Total CO₂e for the reporting period — the headline figure, in metric tonnes with uncertainty range
• Breakdown by AI model — how much each model contributed, enabling year-over-year model mix analysis
• Breakdown by department or team — which business units drove the most AI carbon, enabling targeted reduction initiatives
• Total energy consumption in kWh — required separately from CO₂e under ESRS E1 energy disclosure
• Query volume by model — the underlying activity metric that auditors use to validate the CO₂e calculation
• Carbon intensity factors applied — the g CO₂e/kWh figure for each model, with source citation
• Real-world equivalents — for stakeholder communication materials
• Methodology note — a plain-language description of how each figure was calculated, suitable for direct inclusion in the CSRD disclosure document

For Govarix deployments, this entire output is produced by the ESRS E1 export function in the Admin → Costs → Carbon panel. The export is based on measured token counts — not academic estimation — making it the most defensible form of AI emissions reporting currently available to enterprise sustainability teams.

CSRD Non-Compliance: What Is at Stake

CSRD penalties are enforced by EU member states and vary by jurisdiction, but the exposure is substantial. Fines for non-reporting or materially inaccurate Scope 3 disclosure reach €1M in key jurisdictions including France and Germany. Non-EU companies with significant EU revenue face equivalent obligations under CSRD's third-country provisions.

Beyond financial penalties, the reputational and market consequences are significant. ESG ratings agencies — MSCI, Sustainalytics, ISS — now include AI emissions data in their scoring models. "No data available" is not treated as neutral: missing AI emissions data signals inadequate sustainability governance and depresses scores. Institutional investors managing over $50 trillion in assets have committed to TCFD-aligned portfolio analysis, making ESG data quality a direct factor in equity investment decisions.

The competitive dimension runs in both directions. Enterprises that produce auditable, model-level AI emissions data can use that transparency as a procurement differentiator — particularly when selling to other large enterprises with their own CSRD supply chain disclosure obligations. Your AI carbon data is a risk management asset, not only a reporting burden.

For guidance on how the EU AI Act's risk classification requirements interact with CSRD disclosure obligations, see our EU AI Act risk classification guide.

AI Emissions as a Governance Lever

Real-time carbon visibility makes AI emissions a manageable variable rather than a fixed cost. Model routing policies — directing routine tasks to efficient small models and reserving large frontier models for tasks that genuinely require them — are carbon reduction measures with no productivity trade-off.

Govarix shows the carbon impact of every model routing decision in real time. A sustainability manager can set a monthly carbon budget for AI usage alongside a financial budget, receive alerts when either approaches threshold, and review the model mix driving the most emissions — all within the same admin interface used for cost and usage governance.

This integration between AI governance and sustainability reporting is the model the CSRD was designed to incentivise. Measurable, reducible, auditable emissions data — from a system that is already running.

Frequently Asked Questions