Sustainable AI: 10 Best Practices
On this page
Sustainable AI is less about grand gestures and more about a handful of habits repeated consistently. Whether you are one person using a chat tool or a team shipping AI features, the same sustainable AI best practices apply, and most of them save money as well as carbon.
This is a practical list of ten. Some are for individuals, some for teams, and all of them are things you can start this week. Together they cover the full loop: use less, measure what remains, and remove more than you emit.
1 to 3: Measure, right-size, prompt well
The foundation is measurement. Before you optimize anything, know your footprint using the tokens to energy to carbon method, where a typical text answer lands under a gram to a few grams of CO2. Start with how to measure AI carbon emissions.
Next, right-size the model, since model choice is roughly a five times lever from a small model at about 0.0008 Wh per token to a reasoning model near 0.0040 Wh. Then prompt efficiently by keeping instructions concise and specific, which reduces tokens on both sides of the exchange.
- Measure your footprint before optimizing
- Match the model size to the task
- Write concise, specific prompts
4 to 6: Trim context, batch, stop regenerating
Waste hides in the input. Paste only the portion of a document you actually need rather than the whole thing, and summarize long history instead of resending it every turn. The model reads every token you give it.
Batch related questions into one prompt to cut repeated context, and resist reflexive regeneration. When a reply is close, a short follow-up refines it for far less than a full redo. These tactics are expanded in energy-efficient prompting.
- Trim context to what matters
- Batch related questions together
- Refine instead of regenerating from scratch
7 and 8: Reserve heavy compute and images
Reasoning models and image generation are the expensive end of the menu. A reasoning model uses around 0.0040 Wh per token and an image runs roughly 2.0 to 3.5 Wh, so treat both as deliberate choices rather than defaults.
Use reasoning models when a problem genuinely needs step-by-step depth, and generate images when you actually need one rather than for idle experimentation. The cost of that discipline is small, but repeated across a team it adds up. See reasoning models energy cost.
- Use reasoning models only for hard problems
- Generate images intentionally, not idly
9 and 10: Offset verified and report honestly
The last two practices close the loop. Offset what remains with verified, retired credits, and prefer automatic per-request offsetting so nothing slips through. Ecoia measures each request and offsets past 200 percent of the measured impact, landing on the carbon negative side.
Finally, report honestly. Present numbers as estimates with ranges, note that they vary by model and region, and avoid fake precision. Teams accounting for Scope 3 AI emissions especially benefit from transparent, documented figures.
- Offset remaining impact with verified credits
- Prefer automatic per-request offsetting
- Report honestly with ranges, not fake precision
Make it a team habit
Individual effort helps, but sustainability sticks when it is built into how a team works. Set a default model for routine tasks, share prompt templates that are already tight, and put measurement on a dashboard everyone can see.
For organizations, a platform that tracks and offsets automatically removes friction. Explore green AI for business to see how these practices scale beyond a single user.
The headline: Sustainable AI comes from repeatable habits: measure, right-size models, prompt efficiently, reserve heavy compute, offset with verified credits, and report honestly.
FAQ
Which best practice has the biggest impact?
Right-sizing the model usually wins because model choice is roughly a five times lever. Moving routine work from a frontier or reasoning model to a smaller one cuts energy sharply with little loss of quality. Measurement is what lets you see that gain.
Are these practices only for large teams?
No. Individuals benefit from concise prompts, right-sized models, and fewer regenerations just as much as teams do. Teams gain extra leverage by standardizing defaults and dashboards, but the core habits work at any scale.
Do sustainable AI habits cost more time?
Most of them save time. Concise prompts and right-sized models often return usable answers faster, and skipping needless regeneration is quicker by definition. The main new habit is measurement, which a calculator or automatic tracking makes nearly effortless.
How does reporting fit into sustainable AI?
Honest reporting keeps the whole effort credible. Presenting figures as ranges, noting regional and model variation, and avoiding fake precision builds trust and supports Scope 3 accounting. It also prevents the overclaiming that undermines genuine green work.