# Green AI for business

> A practical guide for businesses on cutting the energy, cost and carbon of AI: model choice, suppliers, measurement, procurement and honest claims.

Source: https://greenai.guide/guides/business/  
Updated: 2026-11-19

## In short

Green AI for a business means getting the results you need from AI with less energy, money and carbon. Most of the saving comes from a few decisions: using AI only where it adds value, picking the smallest model that does the job well, choosing suppliers who publish their energy data, and measuring before you make any claim.

## Key points

- In most businesses the cheapest, fastest and greenest AI is usually the same AI: smaller models, fewer tokens, less rework.
- The International Energy Agency estimates data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's total, and projects about 945 TWh by 2030.
- Task-specific models can be orders of magnitude cheaper to run than general-purpose generative models for the same task.
- Ask suppliers for energy per request, the data centres they use and their measurement method, and write the answers into contracts.
- Make no green claim you cannot back with measured or clearly sourced figures.

## Why should a business care about green AI?

Because the things that make AI greener are mostly the things that make it cheaper, faster and more reliable. Less
computation means lower bills, quicker answers and less waiting, and the same habits that waste energy, such as
long prompts, repeated attempts and oversized models, also waste money and staff time.

A 2026 UK green AI sandpit of mathematicians, computer scientists and industry reached a similar view: buyers choose
on cost, speed and quality, so green AI has to come with those, not compete with them. That is good news for a
business. You rarely have to trade results for a smaller footprint.

There are four other reasons to pay attention:

- **Scale.** The International Energy Agency (IEA) estimates that data centres used about 415 TWh of electricity in
  2024, around 1.5% of global electricity, and projects that this will more than double to about 945 TWh by 2030, with
  AI the most important driver of growth. Your AI use is part of that demand.
- **Reporting.** If you report on energy and emissions, AI is now part of the number, whether you count it or not.
- **Regulation.** Rules are starting to require energy data from AI and data-centre providers (see below), which
  makes better supplier information available to you.
- **Trust.** Customers, staff and investors increasingly ask how you use AI. Clear, honest answers help.

## Where does the energy in AI go?

Most of a typical business's AI energy goes on **inference**, which means running a model to answer requests, not on
training it. You almost never train a large model yourself, but every prompt, every automated workflow and every AI
feature in your software runs inference on someone's hardware.

The main places energy is used:

| Stage | What happens | Who controls it |
|---|---|---|
| Training | A model learns from data, once or occasionally | Mostly the model provider |
| Fine-tuning | An existing model is adapted to your task | You or your supplier |
| Inference | The model answers each request | Shared: you choose how much and which model |
| Data centre | Cooling, power conversion, networking | The hosting provider |
| Hardware | Making the chips and servers (embodied carbon) | The provider, through how long kit is kept |

For a business, inference is the part you can change most. The amount depends on:

- **the model's size and type.** A study by Luccioni, Jernite and Strubell found that general-purpose generative
  models were orders of magnitude more expensive to run than task-specific systems for the same task. In their tests,
  generating images used on average about 2.9 kWh per 1,000 requests, while classifying text used about 0.002 kWh;
- **the number of tokens** (the chunks of text a model reads and writes). Long prompts, pasted documents and
  wordy answers all add up;
- **how many calls** a workflow makes. Agents that plan, check and retry can make dozens of calls for one task; and
- **where and when** it runs, because the carbon in a kilowatt-hour depends on the local electricity grid.

For a sense of scale per request, Google reports that a median text prompt to its Gemini apps used 0.24 Wh of
energy, and that this figure fell 33-fold over a year. Providers measure in different ways, so treat such numbers as
indicative, and see [measuring AI energy](/practice/measure-ai-energy/) for why the details matter.

## How do we choose the right model?

Use the smallest model that does the job reliably, and test that on your own work. For many business tasks, such as
sorting emails, extracting fields from invoices, summarising or translating, a small or specialised model is enough.

UNESCO and University College London found that using smaller, task-specific models could cut energy use by up to 90%
without losing performance on suitable tasks, and that shorter prompts and responses could cut it by more than half.

Practical steps:

1. **List your AI tasks** and what "good enough" means for each, for example 98% of invoice fields correct.
2. **Try models from small to large** on a sample of real cases. Stop at the first that meets the standard.
3. **Route requests.** Send routine requests to a small model and only hard ones to a large model. This is often
   called model routing.
4. **Watch for false economies.** A model that is too small and fails, forcing a rerun on a big model, can use more
   energy than going straight to the big one. Measure success first time, not cost per call.
5. **Reuse, don't regenerate.** Where the same answer, template or piece of code is needed again, store and reuse it
   rather than asking the model each time. GreenCode, an ITEA research project, describes using retrieval, routing to
   the cheapest capable component and scheduling to keep its own AI use small.

Independent ratings can help you compare open models. The AI Energy Score, launched by Hugging Face and partners,
rates models from one to five stars for energy efficiency on common tasks. There is more on this in
[choosing the right model](/practice/choose-the-right-model/) and
[orchestration over generation](/articles/orchestration-over-generation/).

## What should we ask AI suppliers?

Ask for numbers, methods and locations, and put the answers in the contract. Many suppliers cannot yet give full
answers, but asking signals demand and shows you who takes this seriously.

Questions to include in tenders and renewals:

- **Energy per request.** What is the energy (Wh) per typical request or per 1,000 tokens for the models we will
  use? Is it measured or estimated, and how?
- **System boundary.** Does the figure include idle capacity, cooling and networking, or only the chips?
- **Where it runs.** Which data centres and regions host our workload? What is their power usage effectiveness
  (PUE, the ratio of total facility energy to computing energy) and water use?
- **Electricity.** What share is matched hour by hour with clean power, rather than covered by annual certificates?
- **Model options.** Can we choose smaller models, or route requests between models? Can we run on our own
  hardware where data must stay local?
- **Reporting.** Will you give us monthly energy and emissions data for our account, in a form we can use in our own
  reporting?
- **Hardware life.** How long do you keep servers in service, and what happens to them afterwards?
- **Change notice.** Will you tell us if you switch the model behind our service?

Two pieces of EU law are worth knowing, even outside the EU, because large suppliers apply them widely:

- The **EU AI Act** requires providers of general-purpose AI models to document the model's known or estimated
  energy consumption (Annex XI).
- The **Energy Efficiency Directive** requires operators of data centres with at least 500 kW of IT power demand to
  report energy performance, including energy and water use, to a European database (Delegated Regulation 2024/1364).

Our [sustainable AI procurement](/practice/sustainable-ai-procurement/) guide has a fuller question set, and
[hosting and data centres](/practice/hosting-and-data-centres/) explains what the answers mean. You can find
suppliers in the [products directory](/products/) and independent help in the
[consultants directory](/consultants/).

## How do we measure AI energy and emissions?

Measure what you can, estimate the rest openly, and track the things you control. Perfect figures are rarely
available, but a consistent method lets you see whether you are improving.

- **Your own hardware or cloud machines.** Open-source tools such as CodeCarbon estimate emissions from the power
  drawn by processors and memory and the carbon intensity of the local grid.
- **Third-party AI services.** Record tokens, calls and models used. Multiply by the supplier's energy-per-request
  figure if they give one. If not, use a published estimate and say so.
- **Units.** Report energy in kWh and emissions in kg CO₂e (carbon dioxide equivalent), and say which grid factor you
  used.
- **Per outcome.** Track energy per useful result, for example per invoice processed, not only in total. This shows
  the effect of rework and retries.

See [how to measure AI energy](/practice/measure-ai-energy/) and
[why measurement is the bottleneck](/articles/measurement-is-the-bottleneck/).

## How does AI fit into CSRD and ESG reporting?

AI electricity is part of your energy use, so it belongs in your existing emissions reporting rather than in a
separate report. In most reporting frameworks it sits in purchased electricity (Scope 2) where you run the hardware,
and in purchased services (Scope 3) where you use cloud or AI services.

In the EU, the Corporate Sustainability Reporting Directive (CSRD) was narrowed in 2026. The Council of the EU's
final approval in February 2026 limits it to companies with more than 1,000 employees and more than €450 million in
net annual turnover. If you are below that, you may still be asked for data by larger customers who are in scope, or
choose to report voluntarily. Outside the EU, check your national rules and any voluntary frameworks you follow.

Whatever your obligation, treat AI like any other material energy use: name the method, the boundary and the
uncertainty.

## How do we build good habits among staff?

Give people a few simple rules and the right tools by default, rather than asking them to think about energy every
time. Habits matter most where many people use AI many times a day.

- **Pick the right tool.** Search engines, templates, spreadsheets and existing software are often enough. Use AI
  where it saves real time or improves quality.
- **Be concise.** Ask clear questions, ask for short answers, and don't paste whole documents when a section will
  do.
- **Avoid "slot machine" use.** Regenerating an image or answer many times is costly. Refine the request instead.
- **Default to smaller models** in company tools, with larger ones available when needed.
- **Turn off what isn't used.** AI features switched on by default in software, and idle test machines, all draw
  energy.
- **Schedule heavy jobs.** Batch work that can wait, such as overnight document processing, can run when the grid is
  cleaner.

Put these in a short policy; our [green AI policy](/practice/green-ai-policy/) guide includes a template, and
[using AI efficiently](/practice/use-ai-efficiently/) has more everyday tips.

## How do we talk about it honestly?

Make claims only about what you have measured, against a stated baseline, and include the limits. Vague claims like
"sustainable AI" or "carbon-neutral AI" invite scrutiny from regulators and customers.

Good practice:

- State the **baseline** ("compared with our previous model") and the **measurement method**.
- Separate **efficiency** (less energy per task) from **total use** (which may still rise). Cheaper AI often leads to
  more AI use, which is known as the [rebound effect](/glossary/#rebound-effect).
- Don't rely on offsets or annual renewable certificates alone to call a service green.
- Remember that a provider's self-reported figure may not be comparable with another's.

There is more in [honest green claims](/practice/honest-green-claims/) and
[efficiency is not sufficiency](/articles/efficiency-is-not-sufficiency/).

## What to do next

1. **Map your AI use.** List the tools, models and automated workflows in use, and roughly how often each runs.
2. **Pick one high-volume task** and test whether a smaller model meets your quality standard.
3. **Start tracking tokens and calls** for your main AI services, and energy where you run your own hardware.
4. **Add the supplier questions above** to your next tender or contract renewal.
5. **Publish a short staff policy** on when and how to use AI, using our [policy guide](/practice/green-ai-policy/).
6. **Review claims** in your marketing and reports against the [honest claims guide](/practice/honest-green-claims/).
7. **Get help if you need it** from the [consultants directory](/consultants/), and look for tools in the
   [products directory](/products/).

## Frequently asked questions

### Does green AI cost more?

Usually not. The main levers, such as smaller models, shorter prompts and fewer repeated runs, also cut API bills and response times. The extra cost is mostly staff time to measure and to set up routing between models.

### Do we have to report the energy use of our AI?

Not as a separate item in most places. But AI electricity is part of your energy use and emissions, so it falls inside wider reporting such as the EU's CSRD for companies in scope, or voluntary ESG reporting.

### Is a smaller model always greener?

No. A model that is too small can fail and force retries or a second run on a larger model. Test on your real task: the greenest model is the smallest one that gets it right first time.

### How can we measure the energy of a cloud AI service we don't run ourselves?

Ask the supplier for energy per request and the method behind it. Where they won't say, use published per-prompt estimates with clear caveats, and track the things you control, such as tokens and number of calls.

## Sources

- [Energy and AI](https://www.iea.org/reports/energy-and-ai), International Energy Agency (2025)
- [Power Hungry Processing: Watts Driving the Cost of AI Deployment?](https://arxiv.org/abs/2311.16863), Luccioni, Jernite and Strubell, ACM FAccT (2024)
- [Smarter, smaller, stronger: resource-efficient generative AI and the future of digital transformation](https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90), UNESCO and University College London (2025)
- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [Council signs off simplification of sustainability reporting and due diligence requirements](https://www.consilium.europa.eu/en/press/press-releases/2026/02/24/council-signs-off-simplification-of-sustainability-reporting-and-due-diligence-requirements-to-boost-eu-competitiveness/), Council of the European Union (2026)
- [AI Act, Annex XI: technical documentation for providers of general-purpose AI models](https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-11), European Commission AI Act Service Desk (2024)
- [Commission Delegated Regulation (EU) 2024/1364 on the rating scheme for data centres](https://eur-lex.europa.eu/eli/reg_del/2024/1364/oj), EUR-Lex (2024)
- [CodeCarbon](https://codecarbon.io/), CodeCarbon (2026)
- [Announcing AI Energy Score ratings](https://huggingface.co/blog/sasha/announcing-ai-energy-score), Hugging Face (2025)
- [Green AI and energy-aware computing](https://greencode.ai/benefit/green-ai-and-energy-aware-computing/), GreenCode (ITEA) (2026)
