AI and Scope 3 Emissions

by Sustainability Tracker 17/11/2026

Guide, Thought Starters

This article was contributed by Sustainability Tracker.

How Australian Small Businesses Should Measure and Report AI Use

Last updated: 17 November 2026

Contributors: Anne Kehlhofer – Director, ESG, Xero; Sustainability Tracker; Danielle King – Director, Green Moves Australia; James Tilbury – Partner, Linden Sustainability; Sydney Straver – Director, &BLOOM Sustainability & ESG

Quick answer: AI subscriptions and cloud AI services belong in your Scope 3 inventory, usually under purchased goods and services. Most AI providers give customers little or no emissions data, so you will usually need to base emissions on spend basis or estimate by activity data. The method you choose can change the result by a factor of more than 200, so documenting your method matters as much as the number. 

Where to start: If your organisation is new to this, start with a conversation. Find out who is using which AI tools and how often, and whether AI is covered by your sustainability and procurement policies. That gives you the usage data any estimate needs. If you’re a smaller business outside mandatory reporting, the prompt is more likely to be a larger customer asking for your emissions data for its own Scope 3 reporting. The same steps apply, and a simple estimate with your method written down gives you something credible to send.

Do AI tools count towards your Scope 3 emissions?

Yes. Under the GHG Protocol, Category 1 (purchased goods and services) covers upstream emissions from producing goods and services that a company buys in the reporting year, including intangible services. AI subscriptions, AI features bundled into software, and AI accessed through a cloud platform all fall within that.

Cloud providers treat it the same way. Google Cloud’s reporting methodology allocates Google Cloud’s own Scope 1, 2 and 3 emissions to customers based on usage, and customers then include those figures in their reports as Scope 3.

Google’s emissions reports are for business customers, through the Google Cloud and Google Workspace admin consoles. Free personal accounts, such as a free Gemini app account, don’t come with one.

When does AI reporting matter under Australia’s mandatory climate reporting?

Australia’s mandatory climate reporting regime sits under the Corporations Act. Entities prepare their reports using the Australian Sustainability Reporting Standards (ASRS), with AASB S2 as the climate standard. Reporting is phased in by entity size. According to ASIC’s Regulatory Guide 280, Group 1 covers entities meeting at least two of three tests (consolidated revenue of $500 million or more, gross assets of $1 billion or more, 500 or more employees) for periods starting from 1 January 2025. Group 2 and Group 3 follow for financial years beginning on or after 1 July 2026 and 1 July 2027.

GroupFirst reporting periods beginningSize test (at least two of three)
Group 11 January 2025Revenue $500m+, gross assets $1bn+, 500+ employees
Group 21 July 2026Revenue $200m+, gross assets $500m+, 250+ employees
Group 31 July 2027Revenue $50m+, gross assets $25m+, 100+ employees

Emissions-based and asset-owner thresholds also apply. Check RG 280 for the full criteria. Group 3 entities with no material climate-related risks or opportunities can make a statement to that effect instead of full climate disclosures. Thresholds checked as at 2 November 2026.

Four rules shape how AI fits in:

How big is AI likely to be in your emissions inventory?

For businesses using AI for everyday office work, it’s likely to be small. Tunley Environmental concludes that for most businesses, the direct carbon impact of moderate AI use is likely to be small and easily accounted for within wider carbon reporting. Our worked example below puts a 50-person team’s AI use at between about 7 kilograms and 1.9 tonnes of CO₂e a year, depending on the method. Compare your own estimate with your total inventory to see where it sits. Size isn’t the only test, though. Whether information is material depends on what users of your report need to know, and AI is attracting strong investor and customer interest. Some businesses may choose to explain how they treat AI emissions even when the figure is small.

AI becomes more significant for businesses that run large API workloads, host AI models in the cloud, or build AI into their own products. Cloud platforms are also where provider emissions data is most available (see the provider table below), so these businesses can often measure with supplier data.

If AI is a small share of your emissions, a simple estimate is usually proportionate. AASB S2 asks for reasonable and supportable information available without undue cost or effort. Your estimate still needs to use one of the GHG Protocol’s calculation methods, which we cover below. For most direct AI subscriptions, that means a spend-based estimate: your subscription fees multiplied by the relevant emission factor. It’s the easiest method to apply until providers publish better activity data. AI costs are often bundled into wider software or IT spend, so check whether your inventory captures them at all. Record the method, emission factors and assumptions you used, so your assurance provider can follow how you reached the figure. 

If you build AI into your own products

Software companies with AI features face a second question: what to tell their own customers, who will start asking for the same data for their Scope 3 inventories. Start by measuring the AI compute behind your product, ideally with your cloud provider’s carbon tool. Then allocate it to customers on a clear basis, such as usage or revenue, and share the method alongside the figure so customers can see what’s included. Expect these requests to grow as more of your customers report.

How Sustainability Tracker uses AI: Tracker AI uses a mix of Relevance AI and OpenAI models. Relevance AI manages the workflow, and OpenAI generates the answers. To save resources, it picks a model based on how hard the question is. Simple questions go to a smaller, faster model. Harder questions that need more thinking go to a larger, more powerful one. This way, the big models are only used when they’re needed. 

What emissions data do AI providers give their customers?

This is where most businesses hit a wall. We checked each major provider’s customer reporting tools as of 02 November 2026.

How you buy AICustomer emissions toolDoes it cover AI?
Microsoft 365 CopilotEmissions Impact Dashboard for Microsoft 365Copilot is not in Microsoft’s list of covered apps
Gemini in Google WorkspaceWorkspace carbon footprint reportNo. Google states Gemini is excluded
Gemini API or Vertex AI on Google CloudGoogle Cloud Carbon FootprintYes. Both are listed as covered services
Models on AWS (including Amazon Bedrock)AWS Customer Carbon Footprint ToolCovers AWS services broadly. Bedrock is not named specifically
ChatGPT or OpenAI API (direct)None foundNot applicable
Claude or Anthropic API (direct)None foundNot applicable
Grok or xAI API (direct)None foundNot applicable

Microsoft

The Emissions Impact Dashboard for Microsoft 365 reports emissions for Exchange Online, SharePoint, OneDrive, Teams, Word, Excel, PowerPoint and Outlook. Copilot does not appear in that list. A Microsoft community moderator has also stated that Copilot and other integrated AI services are excluded, though that answer comes from a forum volunteer rather than Microsoft itself. Confirm with your Microsoft account team. You may also need to purchase the Power BI add on if your existing account does not already include it.

Google Workspace

Google’s admin help for the Workspace carbon footprint report states plainly that the calculations do not currently include Gemini for Workspace features. 

Google Cloud

Google’s list of covered services includes both the Gemini API and Vertex AI. If your developers access Gemini this way, you have supplier-specific data available.

AWS

In October 2025, AWS added Scope 3 emissions to its Customer Carbon Footprint Tool, covering server manufacturing, facility power and equipment transport. AWS’s methodology document says the model covers EC2, S3, Lambda, SageMaker and the rest of AWS’s services. It does not name Bedrock, so ask AWS whether third-party model usage is included.

OpenAI, Anthropic and xAI

We could not find a customer-level emissions report from any of the three. We cover what each company discloses publicly in our AI platform comparison.

How to estimate AI emissions when your provider gives you nothing

The GHG Protocol’s calculation guidance lists four methods in order of specificity: supplier-specific, hybrid, average-data and spend-based. For most direct AI subscriptions, two are realistic. 

Method 1: Spend-based

Multiply what you spend by an emission factor for the relevant industry. A widely used free dataset is the US EPA’s Supply Chain GHG Emission Factors v1.3, published in July 2024. The EPA created these factors for reporting Scope 3 Category 1 and Category 2 emissions.

The EPA groups its factors by NAICS code, the standard US system for classifying industries. Two categories fit AI subscriptions. In the v1.3 dataset, software publishers (NAICS 511210) carry a factor of 0.080 kg CO₂e per US dollar (80 grams of CO₂e for every US dollar spent), and data processing and hosting services (NAICS 518210) carry 0.093 kg CO₂e per US dollar. The dataset gives each factor with and without wholesale and retail margins. These figures use the “with margins” column, which reflects purchaser prices.

Australian alternatives

Australian practitioners commonly use the free spend-based factors in the University of Melbourne’s EPiC database (2019), adjusted for inflation. The database focuses on construction, but it includes general AUD spend factors. Paid Australian factors are available from IELab. Australian factors reflect the Australian economy. US factors may better reflect AI services delivered by US providers from offshore data centres. Either choice can be defended if you disclose it.

Know the limitations of the US factors before you use them:

On flat-rate plans, price is a poor proxy for AI compute. Subscriptions do cap usage, but within those caps a user sending ten prompts a month pays the same as one sending thousands. Usage-based pricing, now common for heavier business use, ties spend more closely to compute.

Supplier-specific spend factors: If a provider publishes its total emissions and annual revenue, you can divide one by the other to get a factor based on that supplier’s own data. Use Scope 1, Scope 2 and upstream Scope 3, since downstream emissions (such as customers using the provider’s devices) fall outside your purchase. Microsoft, Google and Amazon publish both figures, though the result is an average across each whole company. OpenAI, Anthropic and xAI don’t publish the figures needed.

Method 2: Activity-based

Multiply a usage measure, such as prompt count or tokens (a token is a chunk of text, roughly three-quarters of a word), by a matching emissions figure. Google publishes the most detailed per-prompt figure: the median Gemini Apps text prompt emits 0.03 grams of CO₂e and uses 0.24 watt-hours of energy. Mistral publishes a figure per 400-token response (1.14 grams of CO₂e for its Large 2 model). The two figures use different methods and boundaries, so they can’t be compared directly. Google calculated its carbon figure from its own data centres’ emissions. If you start from the energy figure instead, you need an emission factor for the grid where the processing happens, which providers rarely disclose. 

This method has its own gaps. Google’s figure covers text prompts only, excluding image and video generation. It describes a median prompt, and Hugging Face’s AI Energy Score found reasoning models used on average 30 times more energy than models without reasoning. Applying a Gemini figure to other providers’ models is an assumption you would need to disclose.

Usage data is increasingly available on business plans. Microsoft 365 Copilot’s usage report shows prompts submitted per user. ChatGPT Enterprise and Edu analytics show messages sent per user, broken down by model. Claude Team and Enterprise analytics show chats and tokens. Google Workspace admins can see Gemini usage by app, including how many messages users exchanged with the Gemini app. API usage is reported in tokens. Match your usage measure to the emissions figure you apply. 

Worked example: a 50-person team

A business with 50 staff uses AI chat tools every working day. Here is how the two methods compare. 

Spend-basedActivity-based
InputUSD 20,000 annual AI spend (2022 dollars)50 staff × 20 prompts × 230 working days = 230,000 prompts
Factor0.080 kg CO₂e/USD (software publishers) or 0.093 (data processing and hosting)0.03 g CO₂e per prompt (Google’s median Gemini text prompt)
Result1.6 to 1.86 tonnes CO₂e6.9 kg CO₂e

Illustrative only. Figures are rounded and assume every prompt is a median text prompt.

The spend-based estimate is more than 200 times higher. Per person, that’s about 32 to 37 kg of CO₂e a year on the spend-based method, or about 140 grams on the activity-based method. The classification choice alone (software versus data services) moves the spend-based figure by about 16%.

Spend-based estimates often come out higher than activity-based ones, and the GHG Protocol ranks spend-based as the least specific of its four methods. The two figures here also measure different things. The EPA factor covers the whole supply chain of an average US software or hosting business. Google’s figure covers serving one median Gemini text prompt in Google’s own infrastructure. Many AI tools now run reasoning steps, searches or agent tasks behind a single prompt, so a median prompt is likely to understate current use. The spend-based result is likely to overstate emissions for light users on flat-rate plans, though usage-based pricing narrows that gap for heavier users. For mandatory climate reporting, choose a defensible method, apply it consistently and disclose your assumptions.

If you’re heading towards assurance, lean conservative. Spend-based figures tend to overstate emissions for light users, and market-based per-prompt figures like Google’s tend to understate them. Where you’re unsure which better reflects your use, choose the method that gives the higher figure and explain why.

Which method should you use?

Expect your method to change. As providers publish better data, you may move from spend-based to activity-based or supplier data, and your reported figure may shift sharply as a result. AASB S2 asks you to disclose changes to your measurement approach and the reasons for them. Where you can, recalculate earlier years on the new basis, so a change of method isn’t mistaken for a real reduction.

Questions to ask your AI vendors

Supplier engagement is how Scope 3 data improves over time. Add these to your procurement and renewal conversations. If you’re short on time, start with the first three: they tell you whether you can use supplier data or need to estimate.

  1. Do you provide customer-level emissions data for our usage? If so, which scopes does it cover?
  2. If not, can you provide a supplier-specific emission factor per dollar spent? This should cover your Scope 1, Scope 2 and upstream Scope 3 emissions.
  3. Is AI usage (including assistants, agents and API calls) included in that data, or reported separately?
  4. What methodology do you use, and has it been independently verified?
  5. Do you report location-based and market-based figures?
  6. Which data centre regions process our requests, and what energy powers them?
  7. Do you publish corporate Scope 1, 2 and 3 emissions? Are they independently assured?
  8. Do you publish per-request energy, carbon or water figures for the models we use?
  9. When will you provide any data you can’t provide today?

How to talk about your AI use without greenwashing

The ACCC’s Making environmental claims guide sets out eight principles for trustworthy claims, including having evidence to back up claims, not hiding important information, and explaining conditions or qualifications.

Applied to AI, that means:

Our guide on how to avoid greenwashing covers these principles in more depth.

Should AI’s social impacts be in your procurement policy?

AASB S2 covers climate only, but your procurement policy can go further. Several social issues sit in AI supply chains:

If your organisation reports under the Modern Slavery Act (consolidated revenue of at least $100 million), consider including AI suppliers when you assess supply chain risks. Our AI platform comparison covers these issues in more detail.

What we don’t know about AI Scope 3 emissions yet

Some questions can’t be answered from public information as at November 2026:

What you can do now

  1. Map your AI use. Find out which AI tools staff use, how often and on which plans, including free and personal accounts. Usage reports on business plans make this quicker.
  2. Ask at renewal. Ask each AI vendor whether they provide emissions data for your usage, including AI. Our vendor questions above go further.
  3. Put AI on the agenda. Add AI to your sustainability and procurement policies, and record your estimation method so your figure improves each year.

We’ll update this guide as providers publish more.

Frequently asked questions

Do AI tools count as Scope 3 emissions?

Yes. Emissions from AI subscriptions and cloud AI services you buy sit in Scope 3, usually Category 1 (purchased goods and services) under the GHG Protocol. Cloud providers such as Google allocate their own Scope 1, 2 and 3 emissions to customers, who report them as Scope 3.

When do Australian businesses need to report Scope 3 under mandatory climate reporting?

From each entity’s second reporting year. Group 1 began with annual reporting periods starting on or after 1 January 2025, Group 2 from 1 July 2026 and Group 3 from 1 July 2027. Transition relief lets entities leave Scope 3 out of their first report, and they don’t need a year-one comparative figure in year two.

Does the Microsoft 365 Emissions Impact Dashboard include Copilot?

Microsoft’s documentation lists the applications the dashboard covers (Exchange Online, SharePoint, OneDrive, Teams, Word, Excel, PowerPoint and Outlook), and Copilot is not among them. Confirm with Microsoft before assuming Copilot usage is captured.

Does Google Workspace’s carbon footprint report include Gemini?

No. Google’s Workspace admin help states that the carbon footprint calculations do not currently account for Gemini for Workspace features.

How do I estimate emissions from ChatGPT or Claude subscriptions?

We could not find a customer emissions report from OpenAI or Anthropic, so you will need to estimate. A spend-based estimate multiplies what you spend by an emission factor for software or data services. An activity-based estimate multiplies your prompt or token count by a matching published figure. Business plans for both tools give admins usage data. Document which method and factors you used and why.

How big are AI emissions likely to be for my business?

For businesses using AI for everyday office work, AI is likely to be a small share of total emissions. Our worked example for a 50-person team produces between about 7 kg and 1.9 tonnes of CO₂e a year, depending on the method. Businesses running heavy API, cloud or AI-product workloads should measure more carefully. Whether AI emissions are material for disclosure is a separate judgement to make with your sustainability adviser or assurance provider.

How do I account for free AI accounts?

Free accounts have no spend, so a spend-based estimate doesn’t work. Use an activity-based estimate instead. Free accounts give admins no usage data, so estimate prompt numbers (a short staff survey works), multiply by a published per-prompt figure, and record your assumptions. Free accounts come with no provider emissions reporting, so an estimate is usually the only option.

Update log

DateActivity
17 November 2026First published.

by Sustainability Tracker

This article was contributed by Sustainability Tracker.