ESG Reporting for AI Usage
On this page
ESG reporting AI usage is quickly moving from a niche concern to a standard line item. As teams route more work through chat models, image generators, and reasoning systems, the energy and carbon behind those requests start to matter for disclosure. The good news is that AI fits neatly into frameworks you may already use.
This guide explains where AI belongs in your environmental accounting, what data you actually need to collect, and how to report it without overstating precision. The aim is a disclosure that a reviewer can trust and that your own team can defend.
Where AI usage fits in ESG reporting
When your company buys AI from a vendor, the emissions from running those models sit in your value chain rather than on your own premises. That makes them a scope 3 category, specifically purchased goods and services. You are not operating the data center, but you are responsible for the demand you create.
Treating AI as scope 3 keeps it consistent with how you already handle cloud hosting, software subscriptions, and other digital services. It also sets a realistic expectation: these figures are estimates built from activity data and emission factors, not meter readings from a device on your desk.
What data you need to collect
Credible reporting starts with usage. For text models, the core unit is tokens processed, since tokens drive energy which drives carbon. For images, the unit is generations. If you can capture these per request, you can build everything else on top.
From usage you apply a chain of estimates. A rough per-token energy figure varies by model class, then a data center overhead factor and a grid emission factor convert energy into carbon. Water follows from energy too. You do not need perfect numbers to start, but you do need to record your assumptions so results are reproducible.
- Tokens processed or images generated, ideally per model
- An energy estimate per unit, varying by model class
- A data center overhead multiplier (PUE, often around 1.56)
- A grid emission factor for the relevant region
- A water intensity factor if you report water use
Turning usage into carbon estimates
Per-token energy differs widely by model. Small models land near 0.0008 Wh per token, mid-sized models near 0.0015 Wh, frontier models near 0.0038 Wh, and reasoning models near 0.0040 Wh. Multiply tokens by the right figure, apply overhead of roughly 1.56, then multiply by a grid factor such as 0.395 kg CO2 per kWh for the US average.
For a worked example of this chain, our carbon footprint calculator and the guide on how to measure AI carbon emissions walk through each step. The key is consistency: use the same method every reporting period so trends are meaningful.
How to report it credibly
Disclose your method, not just your total. State which models you used, the emission factors you applied, and the region assumptions behind them. A number without context invites doubt; a number with a clear trail invites trust.
Be honest about uncertainty. AI footprint figures are approximate and vary by model and grid, so present ranges where appropriate and avoid false precision. If you offset or buy carbon-negative AI, report the gross impact and the offset separately so readers can see both. Platforms that measure each request, like Ecoia, make this separation straightforward.
Common pitfalls to avoid
The most frequent mistake is guessing at usage instead of measuring it. Estimating a monthly total from memory produces figures no auditor will accept. Capture real activity data from your provider or a carbon-tracked API instead.
A second pitfall is conflating offsets with reductions. Buying offsets does not lower the energy a model uses; it compensates for it. Keep the two ideas distinct in your report, and pair offsetting with genuine efficiency work. The guide on ESG for software carbon accounting covers this balance in more depth.
The headline: Treat AI usage as a scope 3 emission, base it on real usage data and stated assumptions, and disclose gross impact and offsets separately.
FAQ
Is AI usage really a scope 3 emission?
Yes, when you buy AI from a vendor rather than run the hardware yourself. The emissions occur in your value chain under purchased goods and services. This is the same logic used for cloud services and other digital purchases.
Do I need exact figures to report AI emissions?
No. AI footprint reporting relies on activity data and published emission factors, which are estimates by nature. What matters is that you record your assumptions, apply them consistently, and present uncertainty honestly rather than claiming false precision.
How do offsets appear in an ESG report?
Report your gross measured impact first, then show offsets or carbon-negative purchases as a separate line. This keeps the underlying energy use visible and prevents offsets from masking a rising footprint.
What is the single most useful data point to capture?
Tokens processed per model, or image generations for image work. Usage is the foundation of every downstream estimate, so measuring it accurately does more for credibility than refining any other factor.