AI Strategy

AI’s Carbon Footprint: What Small Businesses Need to Know (and Tell Clients)

Real per-prompt numbers, a ten-minute way to estimate your own footprint, and a two-sentence answer for the next client sustainability questionnaire.

B Biztrategy Published 5 November 2026 · 8 min read
Focused detail of a modern server rack with blue LED indicators in a data center.

Sooner or later, someone is going to ask you about the environmental cost of your AI use. It might be a large client sending a supplier sustainability questionnaire. It might be a tender form with a new box about "digital emissions". It might be a values-driven team member who has read that every ChatGPT question "boils a kettle". And if you have not thought about it, the honest answer — "I don't know" — does not land well.

The good news: for a typical small business, AI's footprint is real but small, it is measurable in an afternoon, and the steps that shrink it are mostly the same steps that cut your AI bill. This guide gives you the actual numbers the big providers have published, a worked calculation for a ten-person team, what the EU's revised sustainability reporting rules mean for SMB suppliers, and a ready-to-use disclosure paragraph you can paste into the next questionnaire.

Why this question is landing on your desk

Three forces are pushing AI emissions into SMB conversations in 2026:

  • Supply-chain reporting. Large companies have to account for emissions in their value chain, so they ask their suppliers. If you are a consultancy, agency or service provider to a corporate, you are part of their Scope 3.
  • Data-centre headlines. The International Energy Agency estimated that data centres used around 415 TWh of electricity in 2024 — roughly 1.5% of global demand — and projected that this could more than double to around 945 TWh by 2030, with AI the main driver. Numbers like that make people nervous about every prompt.
  • Client and staff values. Many owners we speak to describe the same tension: "I want the productivity, but I don't want to be a hypocrite about sustainability." That is a reasonable worry, and it deserves real numbers rather than vague guilt.

The trick is separating the global picture (where AI's growth is a genuine energy-policy issue) from your business's share of it (which, for most SMBs, is tiny compared with heating, travel and hardware).

What a single AI prompt actually costs

Until 2025 nobody had reliable per-prompt figures. Now three of the major providers have published their own estimates. They use different methods, so treat them as ranges rather than gospel:

  • Google (Gemini, median text prompt, August 2025): ~0.24 Wh of energy, ~0.03 g CO₂e and ~0.26 mL of water per prompt.
  • OpenAI (ChatGPT, average query, June 2025): ~0.34 Wh of energy and ~0.3 mL of water per query.
  • Mistral (Le Chat, 400-token reply, full lifecycle including hardware manufacturing, July 2025): ~1.14 g CO₂e and ~45 mL of water per reply.

The Mistral figure is higher because it counts the whole lifecycle — building the chips and data centres — not just the electricity at the moment you hit enter. That is the more complete picture, and it is the one to use if you want to be conservative.

Two caveats matter more than the table itself:

  • Not all prompts are equal. These are figures for a normal text answer. Long reasoning ("thinking") modes, deep-research runs and AI agents that work for minutes can use many times more computation per task. Image generation and especially video generation sit well above text.
  • The "kettle" myth is wrong by orders of magnitude. Boiling a litre of water takes roughly 100 Wh. A typical text prompt is closer to a third of a watt-hour — about the same as running an LED bulb for a couple of minutes.

A worked example for a ten-person team

Let's put the numbers to work. Take a ten-person agency where everyone uses an AI assistant daily.

  • 10 people × 50 prompts a day × 220 working days = 110,000 prompts a year
  • At ~0.3 Wh per prompt: 110,000 × 0.3 Wh ≈ 33 kWh a year
  • Apply a 10× "heavy use" multiplier for reasoning modes, research runs and some image generation: ~330 kWh a year
  • Using Mistral's lifecycle figure instead (~1.14 g per reply), 110,000 replies ≈ 125 kg CO₂e, or around 1.25 tonnes with the 10× multiplier

For context, a single laptop drawing 50 W for eight hours a day uses around 88 kWh a year. Ten laptops: roughly 880 kWh. One return economy flight from Madrid to London is typically estimated at a few hundred kilograms of CO₂e per passenger. So for most office-based SMBs, AI text use sits somewhere between "smaller than your laptops" and "about one or two business trips a year".

That is not zero, and it is growing as agentic tools spread. But it tells you where to focus: if you are serious about your footprint, travel, heating, and replacing hardware too often matter far more than whether you ask ChatGPT one more question.

A quick calculation prompt

You can get your own estimate in ten minutes. Export usage counts from your AI admin dashboard (most business plans show messages or tokens per user), then paste this into your assistant:

"Here is our monthly AI usage by user and tool: [paste]. Estimate our annual electricity use and CO₂e using 0.3 Wh and 1.14 g CO₂e per standard text prompt, and a 10× multiplier for reasoning, research and image generation. Show your workings in a table, state every assumption, and give a low, central and high estimate."

Keep the output. It becomes the evidence behind your disclosure.

What EU reporting rules mean for SMB suppliers

The Corporate Sustainability Reporting Directive (CSRD) was cut back sharply by the EU's 2025 "Omnibus" simplification package. Under the political agreement reached at the end of 2025, mandatory reporting is aimed at very large companies — more than 1,000 employees and more than €450 million turnover. Almost no SMB is directly in scope.

What does still reach you is the value-chain effect. Large companies still need supplier data, so the questionnaires keep coming. The revised rules include a "value-chain cap": in-scope companies should not demand more from smaller suppliers than what is covered by the voluntary SME standard (known as the VSME). In practice, that means you can reasonably answer from VSME-level information rather than building a corporate-grade carbon inventory.

The VSME does not ask you to break out AI separately. AI electricity sits inside your purchased cloud and software services. But clients increasingly ask about it directly, and being able to answer in two sentences with a number signals maturity. (This is general information, not legal advice — check the specifics with your accountant or sustainability adviser if a contract depends on it.)

Seven ways to cut AI emissions (and your bill)

Almost every lever that reduces AI's footprint also reduces cost, which makes this an easy sell to any owner who says "too many tools, too much spend".

  1. Right-size the model. Use the fast, small model for rewrites, summaries and email replies. Save reasoning modes for genuinely hard problems. Most assistants let you choose; agree a team default.
  2. Stop the regenerate loop. Five "try again" clicks is five prompts. A better first prompt — role, context, format, length — usually gets you there in one. Our guide to prompt engineering for small business covers the structure.
  3. Be deliberate with images and video. Generating 40 image variants to pick one is the single most energy-hungry habit in most marketing teams. Brief tightly and generate fewer.
  4. Switch off AI you don't use. Many tools now run AI in the background on every document or email. If nobody uses the feature, turn it off — see the AI features already in your software.
  5. Cancel duplicate subscriptions. Six overlapping tools means six sets of background processing and six invoices. Run the exercise in how to audit your AI tool stack once a quarter.
  6. Prefer providers that publish data. Choose vendors that disclose energy figures and renewable-energy commitments, and, where offered, EU data-centre regions. Add one question about it to your vendor checklist.
  7. Batch agent and automation runs. An automation that re-processes your whole CRM every hour does far more work than one triggered only when a record changes.

What to tell clients: a ready-to-use disclosure

When a questionnaire or client asks about AI and sustainability, a short, specific answer beats a long, defensive one. Adapt this:

"We use AI assistants for drafting, analysis and research. Based on provider-published figures and our own usage data, we estimate our AI use at approximately [X] kWh and [Y] kg CO₂e per year, which is included within our purchased digital services. We default to smaller models for routine tasks, limit image and video generation, review our tool stack quarterly, and prefer vendors that publish energy and emissions data. All client deliverables are reviewed by a person before delivery."

That last sentence does double duty — it answers the sustainability question and the quality question that usually sits behind it. If you are still deciding how much to say about AI use in general, read should you tell clients you use AI? first.

What not to say

  • "AI is carbon-neutral for us." You cannot know that, and provider offsets are not your reduction.
  • "We don't use AI." If your staff do, unofficially, this is both untrue and a risk.
  • A number with no method. Always say where the figure comes from and that it is an estimate.

The bottom line

AI's environmental impact is a real issue at a global scale, but for most small businesses it is a modest line in the footprint — smaller than your laptops, smaller than your travel. That is not a reason to ignore it; it is a reason to handle it calmly. Measure it once, adopt the habits that cut waste (they also cut cost), and write a two-sentence disclosure you can reuse. Then put your energy into the decisions that matter more: which workflows AI should handle, and which ones it should not.

Where does your business stand on AI?

Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.

Take the Free Quiz →