An AI Sustainability Checklist for Teams
On this page
An AI sustainability checklist turns good intentions into repeatable actions. Most teams want to use AI responsibly but lack a concrete list of what to actually do. This article provides one, organised so you can adopt it step by step.
The checklist covers measuring your usage, choosing efficient models, cleaning up the power behind them, offsetting what remains, and reporting the result. Work through it once to set a baseline, then revisit it each quarter as your usage grows.
Step one: measure your usage
You cannot improve what you never quantify, so start by measuring. Capture tokens processed and images generated, ideally per model, since these are the units that drive energy and carbon. Even a rough baseline reveals where your footprint concentrates.
From usage you can estimate impact using the standard chain: energy per token by model class, overhead near 1.56, and a grid factor around 0.395 kg CO2 per kWh. The steps are laid out in how to measure AI carbon emissions, and an AI carbon footprint calculator can do the arithmetic for you.
- Log tokens and image generations per model
- Estimate energy, carbon, and water from usage
- Set a baseline you can compare against later
- Decide how often you will re-measure
Step two: choose efficient models
Model choice is the biggest lever most teams have. A small model near 0.0008 Wh per token can handle many routine tasks that people reflexively send to a frontier model near 0.0038 Wh. Reserve the heavyweight and reasoning models for work that genuinely needs them.
Trim waste while you are at it. Shorter prompts, tighter output limits, and caching repeated calls all cut tokens directly. The practices in how to reduce your AI carbon footprint show how much a few defaults can save across a busy team.
Step three: clean up the power
Where your compute runs matters. Providers on cleaner grids, or those pursuing round-the-clock carbon-free energy, emit less for the same work. Favour them when you can, and schedule flexible batch jobs for hours when the grid is cleaner.
You will not achieve perfectly clean power because grids are mixed and variable, as explained in renewable energy powered AI. The goal is to lower the carbon intensity of each request, not to wait for a flawless grid that does not yet exist.
Step four: offset what remains
Efficiency and clean power reduce impact but never erase it, so offset the remainder. Choose measured, verifiable offsets tied to your actual usage rather than vague annual estimates. The distinction between compensating and reducing is covered in what is carbon offsetting.
Some platforms build this in. Ecoia measures each request and offsets beyond 200% of the impact, making usage net negative without extra effort on your side. You can see the mechanics in how it works and decide whether to offset yourself or let the platform do it.
Step five: report and review
Close the loop by reporting. Record gross impact and offsets separately, state your assumptions, and fold the figures into any ESG disclosure. Treating AI as a scope 3 emission keeps it consistent with the rest of your value-chain accounting.
Then review on a schedule. Usage grows, models change, and grid factors shift, so a checklist is only useful if you run it again. A quarterly pass keeps your baseline current and catches drift before it becomes a surprise in an annual report.
The headline: A workable AI sustainability checklist is measure usage, choose efficient models, clean up power, offset the rest, and report it, then repeat each quarter.
FAQ
Where should a team start with AI sustainability?
Start by measuring usage in tokens and image generations, ideally per model. A baseline tells you where your footprint concentrates and gives you something to improve against. Everything else on the checklist builds on that first measurement.
How much does model choice really matter?
A great deal. Small models can use roughly a fifth of the energy per token of frontier models, so routing routine tasks to lighter models is often the single largest saving available. Reserve the heaviest models for work that truly needs them.
Do we still need to offset if we use efficient models?
Yes. Efficiency and clean power lower impact but never remove it entirely, so offsetting handles the remainder. Choose measured offsets tied to actual usage, and keep them separate from reductions in your reporting.
How often should we run the checklist?
Quarterly works for most teams. Usage grows, models change, and grid factors shift over time, so periodic review keeps your baseline accurate and prevents unpleasant surprises when annual reporting comes around.