AI is now part of everyday business, with tools such as Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google) and Copilot (Microsoft) supporting research, writing, coding, data analysis and customer service.
But these advances come with an environmental cost. Although the emissions associated with individual AI queries are usually very small, companies with carbon-reduction plans, net-zero targets or Scope 3 reporting requirements should not treat the tech as impact free.
Limited data
Businesses face challenges in accurately reporting AI’s impact in part because of the lack of provider data. Anthropic, for example, has not yet published a full corporate sustainability report with audited Scope 1, 2 and 3 emissions.
The more important issue is not cost, but the quality and transparency of the data
Generally, where supplier emissions data is unavailable, organisations will often use spend-based emissions factors for Scope 3 reporting. This method uses industry average data rather than specific company figures.
For example, using a spend-based proxy of 0.1177 kg CO₂e (carbon dioxide equivalent) per pound spent, a company spending £10,000 (US$13,504) per year on an AI/software service would estimate:
- 1,177 kg CO₂e, or 1.177 tonnes CO₂e, in annual emissions
- an offset cost of around £17.66 (US$23.84) per year, assuming £15 (US$20.25) per tonne.
For comparison, using organisation-level emissions intensity from public reporting, the same annual spend could be estimated at around 0.627 tCO₂e for Copilot or 0.513 tCO₂e for Gemini. At £15 per tonne, this would equate to offset costs of approximately £9.41 (US$12.70) and £7.70 (US$10.39) respectively.
The difference is small in financial terms. Anthropic may add roughly £8-10 (US$10.80-13.50) per year in offset costs compared with Microsoft or Google equivalents. The more important issue is not cost, but the quality and transparency of the data available.
Scaling issue
Spend-based reporting is useful for carbon accounting, but it doesn’t always reflect the real impact of individual AI use. On a per-query basis, text-based AI is already relatively efficient.
Indicative 2025-26 estimates for a typical text query include:
- Gemini: around 0.03g CO₂e per text query
- Open AI’s GPT-4o: around 0.13g–0.19g CO₂e per query
- Anthropic Claude: around 0.2g–0.4g CO₂e for standard models.
These figures vary depending on model size, query complexity, output length, data centre efficiency and electricity mix. However, for moderate business use, even tens of thousands of text prompts per month may only add up to a few kilograms of CO₂e per year. This means the carbon cost of day-to-day AI use is often small compared with larger emissions sources such as energy, transport, purchased goods or supply-chain activity.
AI can be efficient at the individual query level while still contributing to rising electricity demand
The concern around AI is not about one employee using Claude to draft a document. The issue is scale. AI adoption is growing rapidly across businesses, public bodies and households, while more advanced applications such as AI agents, large-scale automation, complex reasoning, image generation and video generation can require significantly more computing power than straightforward text prompts.
AI can therefore be efficient at the individual query level while still contributing to rising global electricity demand. The overall impact will depend on how quickly models become more efficient, how data centres are powered and cooled, how quickly electricity grids decarbonise and whether AI is used to support wider environmental progress.
Practical approach
A practical approach is to understand which AI tools are being used across the business; whether usage is limited to text prompts or includes more energy-intensive applications; and how these tools should be included within Scope 3 reporting.
Where supplier data is unavailable, businesses can use a suitable spend-based factor as a transparent proxy while asking suppliers for better information on emissions, renewable energy use, data-centre efficiency, water use and carbon-reduction plans.
Credible sustainability reporting depends on clear boundaries
It is also important to keep the impact in proportion. For most organisations, AI subscriptions will be a small part of the overall footprint, so carbon reduction efforts should still focus on the most material emissions sources. If AI emissions are estimated using spend-based data, this should be communicated clearly and proportionately.
The main concern, however, is transparency. Because Anthropic does not currently publish detailed organisational emissions data in the same way as Google or Microsoft, businesses may need to rely on spend-based emissions factors when calculating Scope 3 impacts. (See the example calculations above.)
The figures are not large, but it is still worth measuring because credible sustainability reporting depends on clear boundaries, consistent methodology and honest communication.
The responsible position is not to reject AI, but to use it thoughtfully, measure its impact where possible, ask providers for better transparency and ensure AI supports wider environmental progress rather than distracting from it.
