Somewhere in the middle of an ordinary Tuesday, a finance director opens the annual report draft and stops on a line she has not seen before. Someone in the sustainability team has flagged "AI usage" as a possible Scope 3 category, and asked whether it needs a number. The company has a Copilot licence for most of the office. A few teams use ChatGPT Enterprise. Marketing has a video tool that quietly runs a large model in the background. Nobody is entirely sure what to say.
That conversation is happening in a lot of British businesses right now. AI has arrived in day-to-day work faster than most carbon accounting frameworks were designed for, and the reporting question has caught up: does it count, where does it sit, and what should we do about it?
The short answer is that yes, it counts, and for almost every business using AI through a third-party provider, it sits inside Scope 3. Working out a defensible number is harder. Working out what to ask your suppliers is not.
This guide is for UK sustainability leads, finance directors, procurement managers and anyone else being asked those questions. It sets out where AI belongs in your business carbon footprint, what the current numbers actually say, what a sensible supplier request looks like, and how to write it up without overclaiming.
Table of contents
Where AI actually sits in your carbon footprint The three layers of AI-related emissions What UK businesses can reasonably report today A five-question supplier request for AI vendors What good and less-good answers look like How to write this up in a Carbon Reduction Plan or annual report The wider UK picture What CNB expects from certified partners on this
Where AI actually sits in your carbon footprint
For most businesses, AI usage is Scope 3, Category 1: purchased goods and services. If you buy a subscription to an AI tool, or your CRM quietly uses a large language model behind the scenes, the emissions from running that model are upstream of your business. They happen at the supplier's data centre, on the supplier's electricity contract, using the supplier's hardware. Under the GHG Protocol Corporate Value Chain (Scope 3) Standard, those upstream emissions belong in your Scope 3 inventory when you report on the products and services you buy.
There are two exceptions worth naming clearly:
If you own and operate the data centre, the direct fuel combustion sits in your Scope 1 emissions and the purchased electricity in your Scope 2.
If you run models on hardware you control but on a leased site, the picture depends on the terms of the lease and who holds the energy contract.
Almost no British business outside the largest banks and cloud providers falls into those categories. If you are using AI through a supplier, the emissions are Scope 3.
Evidence: GHG Protocol, Category 1: Purchased Goods and Services.
CNB view: treating AI as Scope 3 is not an accounting technicality. It is what makes the number honest. It also means the emissions belong in the same conversation as any other purchased service, which is where the practical response lives.
A Scope 3 figure that is roughly right and clearly explained is more valuable than a precise figure that hides its assumptions.
The three layers of AI-related emissions
When people talk about "the carbon cost of AI", they usually mean one of three different things. Keeping them separate makes the reporting question tractable.
Operational electricity. This is the electricity used to run inference (answering your prompts) and training (building the model in the first place). Training is a large upfront cost. Inference is smaller per query but happens billions of times a day, and now accounts for the majority of ongoing AI electricity use.
Embodied emissions. These are the emissions from making the hardware itself. Fabricating an advanced AI chip is energy-intensive, and the server, cooling and networking equipment around it all have carbon costs before they are ever switched on.
Network overhead. Every AI query travels through internet infrastructure. It is usually a small share of total emissions, but not zero.
For scale: the International Energy Agency estimates that global data centre electricity consumption was around 415 TWh in 2024, roughly 1.5% of global electricity, and projects it will roughly double by 2030, with AI-focused data centres tripling their consumption over that period. AI is not the whole of the data centre load, but it is the fastest-growing part of it.
None of this makes AI use uniquely bad. A commercial flight, a full year of office electricity, or a large steel order will typically dwarf a company's AI footprint. It does mean AI is now large enough to appear in a considered Scope 3 assessment, rather than sit invisibly in the "everything else" line.
What UK businesses can reasonably report today
Here is where honesty matters. The reporting question is easier than the measurement question.
The GHG Protocol lists four methods for calculating Scope 3 Category 1 emissions:
Supplier-specific. You ask the supplier for the emissions attached to what you bought. Best data, hardest to obtain.
Hybrid. You combine supplier-specific data where you can get it with industry averages where you cannot.
Average-data. You use published emissions factors for the type of service.
Spend-based. You multiply what you spent by a published spend-to-emissions factor.
Most UK businesses using AI today will end up on method three or four, sometimes moving to method two as suppliers improve their disclosure. That is not a failure. It is what reporting looks like at the edge of a new practice.
For UK-specific factors, the DESNZ 2026 greenhouse gas conversion factors remain the reference for electricity and general reporting. There is not yet a bespoke DESNZ factor for AI inference, so the honest approach for now is to state your method, cite the source, and note the limitations.
CNB view: a Scope 3 figure that is roughly right and clearly explained is more valuable than a precise figure that hides its assumptions. Auditors, procurement teams and readers can work with a documented estimate. They cannot work with a confident number that falls apart under a follow-up question.
A five-question supplier request for AI vendors
If your business is serious about reporting AI usage properly, the single most useful thing you can do is send the same short set of questions to every AI vendor you buy from. Five is enough. More becomes noise.
Where is our workload processed? Which data centre regions handle our account, and what is the electricity mix in each?
What is your published Scope 1, 2 and 3 emissions inventory? For the most recent reporting year, with the assurance status.
Do you offer a product carbon footprint (PCF) figure? In kg CO2e per unit of usage (per query, per million tokens, per compute hour), or a customer-allocated total.
What is your renewable energy sourcing method? Matched hourly, annual, unbundled certificates, or grid-average? This changes the meaning of any "carbon-free" claim.
How can we access this data in future reporting years? A one-off answer is helpful; a repeatable one lets you improve the estimate over time.
CNB view: these five questions do two jobs. They give you a defensible basis for your own reporting, and they signal to the supplier that AI-related emissions are now a normal procurement question. Both matter. The market improves when buyers start asking.
What good and less-good answers look like
A strong answer names the data centre region, cites an audited inventory, distinguishes hourly-matched renewables from annual certificates, and offers either a PCF or a customer allocation method. A weaker answer points at the corporate sustainability page and stops there. Neither is a reason to switch supplier immediately, but the pattern of answers you collect will tell you which parts of your AI supply chain are ready for scrutiny and which are not.
How to write this up in a Carbon Reduction Plan or annual report
Reporting language for AI does not need to be complicated. It needs to be accurate about method and honest about limits.
A useful pattern, in three parts:
What we did. We identified AI tools used across the business as part of our Scope 3 Category 1 inventory for the reporting year, following the GHG Protocol Corporate Value Chain Standard.
How we estimated it. Where suppliers provided a product carbon footprint, we used it directly. Where they did not, we applied an average-data method using published data-centre electricity intensity figures and the DESNZ 2026 conversion factors, and treated the result as an estimate.
What we plan to improve. We are asking our AI suppliers for more granular figures for the next reporting year, and will move to a supplier-specific method for the tools where that data becomes available.
Avoid: any statement that a specific AI tool is "carbon neutral" because the provider says so, unless you can point to independent verification and the retirement records that back the claim. Avoid: describing your AI emissions as "offset" through the provider's contract, unless you have the underlying evidence yourself. These are the areas where the Advertising Standards Authority is currently active on green claims across the wider tech and services sector.
The wider UK picture
The AI energy debate in the UK is now a real policy conversation, not a forecast. In July 2026, Ofgem announced measures to free up grid capacity by tackling speculative data centre projects, after demand connection applications rose from 41 GW to 125 GW in under a year, with data centres accounting for at least 80 GW of that increase. The regulator's consultation, open until 16 September 2026, proposes a commitment fee for large data centre developments to distinguish serious projects from ones that are simply holding a queue place.
The point for a UK business is not that AI is about to be regulated as an emissions category. It is that AI's electricity demand has become large enough for the grid regulator to write policy about it, which means AI's carbon footprint is no longer an abstract question for reporting. It is a live one.
What CNB expects from certified partners on this
For businesses working towards Carbon Neutral Britain certification, AI usage is now something we look at as part of the Scope 3 review. We do not expect a perfect figure. We do expect a documented method, a named source for any published factors, and evidence that the business has at least asked its AI suppliers the five questions above.
This is consistent with how we treat any emerging Scope 3 category. Certification means we can show the reader the basis for the claim, not that every corner of the value chain has been measured with laboratory precision. AI is a good example of why that matters. A defensible estimate today, improved each reporting year, is worth more than a confident silence.
Scrutiny of AI's carbon impact is going to sharpen over the next couple of years. That is a good thing. It is the same pressure test that has been applied to flights, supply chains and offset markets, and it will produce the same result: better data, clearer language, and businesses that can answer the follow-up question without hesitating.
Rob Hebden
Sustainability Consultant | Carbon Expert | Helping UK Businesses on the Journey to Net-Zero
Is my AI usage Scope 1, 2 or 3?
For almost every business using AI through a third-party provider, it is Scope 3 Category 1 (purchased goods and services). Scope 1 and 2 only apply if you own or operate the data centre yourself.
Do I need to include ChatGPT or Copilot in my carbon footprint report?
If you are reporting Scope 3 Category 1 in full, yes. If you are reporting a limited subset of Scope 3, the honest choice is to state the exclusion and note that AI usage will be included in a future reporting year.
How do you calculate Scope 2 emissions?
Scope 2 emissions are calculated by multiplying electricity consumption (kWh) by the relevant greenhouse gas emission factor (kg CO₂e per kWh). UK organisations typically use official government conversion factors and must report both location-based and market-based results.
What if my AI vendor will not provide emissions data?
Use an average-data or spend-based method, cite it clearly, and treat the vendor's non-disclosure as itself useful information for future procurement decisions. Ask again next year.
How much CO2e does a typical AI query produce?
Published estimates vary widely, from a fraction of a gram to several grams per prompt, depending on model size, hardware and grid mix. Any headline number quoted without those variables should be treated cautiously. For reporting purposes, a total based on your electricity share is more defensible than a per-query multiplier.
Does buying "green" cloud offset my AI emissions?
Not automatically. A supplier's renewable energy claim, an offset arrangement or a "carbon neutral" cloud badge can be part of the picture, but they need underlying evidence: matched-hourly renewables, verified retirement records, independent assurance. Without that, the safer position is to report the gross figure and note the supplier's own claims separately.
