greenai.guide

Home › Good practice › How to write a green AI policy

Good practice

How to write a green AI policy

Updated · 5 min read

In short

A green AI policy sets out when your organisation uses AI, how it chooses and runs it efficiently, how it measures and reports the impact, and what it will and won't claim. Keep it short, practical and tied to existing processes such as procurement, data protection and sustainability reporting, and review it yearly.

Key points

  • A good policy covers both efficiency (using AI well) and sufficiency (using it only where it adds value).
  • Make the defaults do the work: smallest adequate model, AI features off unless needed, output limits on.
  • Tie the policy to processes that already exist: procurement, data protection, sustainability reporting, training.
  • Measure with a stated method and report per unit of useful work, so growth in use is visible.
  • Give the policy an owner, a review date and a short list of measurable commitments.

What is a green AI policy for?

It turns good intentions about AI's environmental impact into defaults, decisions and checks that people actually follow. Without one, each team chooses models, tools and suppliers on capability and price alone, and nobody can say what the organisation's AI use costs in energy, carbon or water.

A good policy does two things. It makes AI use efficient: the right model, run well, on clean electricity. And it makes AI use sufficient: AI where it adds value, not by reflex. At a 2026 UK green AI sandpit of mathematicians, computer scientists and industry, most participants in a poll wanted green AI to focus on both equally, rather than on either alone. The sandpit also warned against policies that read as "don't use AI": they tend to be ignored. Aim for using AI responsibly and well.

Who should be involved?

Bring together the people who already own the processes the policy touches. A short working group is enough:

  • someone who owns the policy and is accountable for it (often sustainability, digital or operations)
  • IT or engineering, who choose and run models and tools
  • procurement, who buy AI services
  • data protection and legal, because data minimisation and green claims overlap with their work
  • communications, who make public statements
  • a few everyday users, who know where AI is really used

How do I write it, step by step?

Start from what you use now, set a small number of defaults and commitments, and build in measurement and review.

  1. Map current AI use. List tools, features, models and suppliers, including AI bundled into software you already have. Note rough volumes.
  2. Agree your principles. Three to five plain statements, for example "we use the smallest tool that does the job well".
  3. Set defaults. Defaults change behaviour more than guidance does. Decide which AI features are on or off by default, the default model size, and output and history limits in tools you build.
  4. Define when heavier use needs a reason. For example: a large frontier model, image or video generation at scale, training or fine-tuning, or always-on agents.
  5. Choose how you will measure. Agree a method, a boundary and a functional unit; see measure AI energy. Accept estimates at first, labelled as such.
  6. Write buying rules. Add sustainability questions and scoring to AI procurement; see sustainable AI procurement.
  7. Set rules for claims. Nothing is described as green without evidence; see honest green claims.
  8. Plan training. A short, practical session for all staff, plus deeper material for people who build or buy AI.
  9. Pick a few measurable commitments with dates.
  10. Name an owner and a review date, and publish the policy internally.

What should the policy say? A template outline

Use this outline as a starting point and cut what doesn't apply.

1. Purpose and scope

  • Why the organisation has this policy, in two or three sentences.
  • What it covers: AI tools used by staff, AI built into products and services, AI bought from suppliers, and any training or fine-tuning.
  • What "AI" means here (generative AI only, or all machine learning).

2. Principles

  • Use AI where it adds clear value; consider simpler methods first.
  • Use the smallest model or tool that meets the quality bar.
  • Avoid waste: repeated regeneration, oversized inputs and outputs, idle agents.
  • Prefer low-carbon hosting and cleaner hours for flexible work.
  • Measure honestly and claim only what we can show.

3. Everyday use

  • Expected habits for staff (see use AI efficiently).
  • AI features off by default unless a team needs them.
  • Guidance on image and video generation.

4. Building and running AI

5. Buying AI

  • Standard sustainability questions and tender scoring.
  • Standard contract clauses for reporting, location and opt-outs.

6. Measurement and reporting

  • Method, boundary and functional unit, for example the Software Carbon Intensity approach in ISO/IEC 21031:2024.
  • What is reported, how often and to whom; whether figures are measured or estimated.
  • How AI fits into existing sustainability or emissions reporting.

7. Claims and communication

  • Who approves public environmental claims about AI, and what evidence is needed.

8. Training and awareness

  • Who gets what training, and how often.

9. Roles and governance

  • Policy owner, working group, and how the policy links to AI governance (for example an ISO/IEC 42001 AI management system, if you have one).

10. Commitments and review

  • Three to five measurable commitments with dates.
  • Annual review date, and what triggers an earlier one (a new major AI tool, a new regulation).

How do I make it stick?

Build the policy into the moments where decisions are already made.

  • Add sustainability questions to the existing approval form for new tools.
  • Put energy per request next to cost and latency in technical reviews.
  • Add AI to the sustainability report you already produce.
  • Make training short, practical and part of induction. The sandpit heard of a university that made AI literacy, including AI's environmental consequences, compulsory for all staff, treated like data-protection training.
  • Watch for the rebound effect: if efficiency gains are swallowed by more use, the totals will show it. Report totals as well as per-unit figures.

Checklist

  • [ ] Current AI use mapped, including bundled features and suppliers
  • [ ] Principles cover both efficiency and sufficiency
  • [ ] Defaults set for features, model size and output limits
  • [ ] Heavier uses need a stated reason
  • [ ] Measurement method agreed and annexed
  • [ ] Procurement questions and contract clauses included
  • [ ] Rules for environmental claims included
  • [ ] Training planned for all staff and for builders and buyers
  • [ ] Owner, commitments and review date named

What to do next

  1. Map your organisation's AI use this month, including AI switched on inside existing software.
  2. Draft principles and defaults using the template above, and test them with a few users.
  3. Agree a measurement method and collect a first, estimated baseline.
  4. Add the sustainability questions to your next AI purchase.
  5. Set three measurable commitments and a review date, and publish the policy internally.
  6. See the business guide, the charities guide or, for schools, the teachers' page; find help in the consultants directory and training in courses.

Frequently asked questions

Do we need a separate green AI policy?

Not necessarily. Many organisations add a green AI section to an existing AI-use, IT or sustainability policy. What matters is that the commitments are written down, owned and checked.

How long should a green AI policy be?

Two to four pages is usually enough. Put detailed methods, such as how you measure energy, in a separate annex that can be updated without re-approving the policy.

Should a green AI policy ban some uses of AI?

Outright bans are hard to enforce and one person's frivolous use is another's creative work. It is usually better to set defaults, require a justification for heavy uses and make the costs visible.

Which standards can a policy refer to?

Common reference points include the Software Carbon Intensity specification (ISO/IEC 21031:2024) for measurement, ISO/IEC 42001 for AI management systems, and the W3C Web Sustainability Guidelines for digital services.

Sources

  1. ISO/IEC 42001:2023 Information technology – Artificial intelligence – Management system, ISO (2023)
  2. Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
  3. Web Sustainability Guidelines, W3C (2025)
  4. AI large language models: new report shows small changes can reduce energy use by 90%, UNESCO (2025)
  5. Thinking about using AI? Here's what you can and (probably) can't change about its environmental impact, Green Web Foundation (2024)
  6. Green Claims Code: making environmental claims, Competition and Markets Authority (2021)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy