The environmental footprint of AI in an SME: measuring without kidding yourself

The blind spot of the carbon footprint

SMEs that measure their footprint count their buildings, their vehicles, their travel. Almost none counts its uses of artificial intelligence (AI) — even as those uses spread: assisted writing, translation, document sorting, chatbots. Every query consumes electricity in a data centre, often abroad, on an electricity mix more carbon-intensive than the Swiss one.

This is not a problem of conscience, it is a problem of method: nobody spontaneously knows how much a query to a large language model (LLM) “weighs”, and the figures in circulation contradict one another. This guide gives sourced orders of magnitude, an estimation method within reach of an SME, and the reduction moves that cost nothing.

The honest orders of magnitude

Public estimates for 2025-2026 place a query to a general-purpose LLM between a fraction of a watt-hour and roughly 2 to 5 Wh at the high end, depending on the model, the length of the answers and the infrastructure. The uncertainty is real — providers publish little — and the viral figures circulating on social media are often out of date or stripped of context. Remember the range, not the decimal.

Per unit, it is little: a few minutes of an LED bulb. What weighs is volume and normalisation — when AI is built into every e-mail, every search, every document, the meter runs continuously. To that are added two less visible items: the manufacture of hardware (servers, chips — often half the total footprint of digital technology) and the cooling water of data centres.

The training of large models, for its part, is spectacular but shared: spread over billions of queries, it weighs less, for a user, than daily inference. For an SME deployer, it is indeed usage that is the item to track.

The method: an order of magnitude in two hours

The formula fits on one line: number of users × queries per day × working days × energy per query × carbon intensity of electricity. The aim is not a precise figure — it would be wrong — but a documented range, with every assumption written down in black and white.

Example: an SME of 50 employees, half of whom use AI at a rate of 20 queries a day. 25 × 20 × 220 days × 3.5 Wh ≈ 385 kWh a year. On a US electricity mix (~0.4 kg CO₂e/kWh, the most frequent case for the large AI providers), that gives roughly 150 kg CO₂e a year — the order of magnitude of 1,000 km by car. Little, as things stand. But that figure triples if usage becomes general, and it ignores hardware.

First step before any calculation: the inventory. You do not measure what you have not listed — and most SMEs discover on that occasion AI uses nobody had declared. The register of AI systems, which you should be keeping anyway for compliance, serves exactly that purpose: our free tool generates it from your use cases. And for the formula itself, an order-of-magnitude calculator — the same assumptions, adjustable — is available on our Footprint page.

Reducing: the frugality that costs nothing

The right tool for the task. A frontier model to rephrase an e-mail is a lorry to deliver a letter: lightweight models or deterministic functions (document templates, rules) are enough for a large share of uses. The first frugality move is a move of economic common sense — useless queries also cost money.

Choose providers that document. Published energy and climate commitments, efficient data centres, transparency about the electricity mix: build the criterion into your purchasing grid, as you do (or should do) for data protection. Hosting in Switzerland or Europe on renewable electricity cuts carbon intensity by a factor of two to five compared with the average US mix.

Extend the life of hardware. The hidden item in an SME's digital footprint remains the manufacture of equipment: keeping a computer five years instead of three often does more for your balance sheet than all the AI usage optimisations combined. Measure first, optimise afterwards — in that order.

What the law requires (and does not yet require)

No text today obliges a Swiss SME deployer to measure the footprint of its AI uses. The AI Act encourages voluntary codes of conduct including environmental sustainability (art. 95) and imposes energy documentation obligations on providers of large models — not on their customers.

The pressure arrives through another channel: the value chain. Large companies subject to non-financial reporting — Swiss (art. 964a et seq. of the Code of Obligations) or European — pass their obligations down to their suppliers through ESG (environmental, social and governance) questionnaires. An SME that can quantify the footprint of its digital operations, AI included, answers in two lines where its competitors improvise. It is a commercial advantage before it is a virtue.

We apply the method to ourselves

The Environmental dimension of the KarmaScore assesses precisely that: measurement of the footprint, frugality by design, purchasing criteria, dated targets. And because a method is worth something only if you apply it to yourself, we publish the footprint of our own product — assumptions, ranges and targets included — on our Footprint page. Roughly 10 to 15 kg CO₂e a year at the current stage: negligible today, tracked quarterly so that it stays so in proportion to usage.

Fifteen minutes of self-assessment place you across the 24 indicators, environmental dimension included, and the action plan costs each move. Free, no account, and your answers never leave your browser.

Frequently asked questions