AI's energy footprint: myth, fact, and what SMEs should actually measure

The idea that AI is simply too energy-hungry to use responsibly is not unreasonable. Headlines about data centre water consumption and the carbon cost of training large models are real and worth taking seriously. So when an Operations Manager, Sustainability Lead, or Managing Director decides to pause an AI project on environmental grounds, that is a thoughtful call, not a naive one.

The problem is that the concern is often shaped by the most extreme examples rather than by the tools a small business would actually use. September 2026 research reported by Money.co.uk found that 22% of small businesses not currently using AI cited energy and natural-resource concerns as a reason. That is a meaningful proportion. It also suggests those businesses may be making decisions based on worst-case assumptions rather than measured data about the specific tools on their shortlist.

This article does not argue that AI has no environmental cost. It does. The argument is that a blanket rejection based on the sector's most energy-intensive activities is likely to deprive your business of genuine productivity gains, while doing little to reduce global emissions. The right response is not to reject AI categorically. It is to ask the right questions about each tool you are considering.

Why this concern is entirely reasonable

The sustainability concern about AI did not come from nowhere. Training a large language model from scratch is genuinely energy-intensive. Published estimates show that training a single frontier model produces substantial carbon emissions. Water usage at data centres has also drawn scrutiny, particularly during periods of high demand or drought.

For a business with published sustainability commitments, those headlines land heavily. Membership bodies and professional firms that report on net zero to members or clients will recognise this. Sustainability Leads in particular are right to ask whether adopting AI sits comfortably alongside net-zero targets or supply-chain environmental standards. The concern is legitimate. The leap from "training frontier models is expensive" to "every AI task my team runs is irresponsible" is where the reasoning starts to break down.

Most SMEs are not training models. They are using them. Running a query through an existing model, generating a document, or summarising a report is a different activity entirely from building the model in the first place. The energy cost is orders of magnitude smaller. That distinction matters when you are weighing the environmental case.

What the evidence actually shows

The honest answer is that published, task-level energy data for consumer AI tools is still limited. Vendors do not consistently disclose per-query energy consumption figures in a form that allows direct comparison. That transparency gap is itself a problem, and it is one reason the concern persists. When information is absent, people reasonably assume the worst.

What the evidence does show is that the framing matters enormously. Not all AI tasks carry the same footprint. A lightweight model handling a structured summarisation task uses substantially less energy than a large multimodal model generating video or processing thousands of images. The size of the model, the volume of queries, and the energy source of the data centre hosting it are all variables that affect real-world impact.

Hosting arrangements are particularly relevant. Several major AI providers now publish data centre renewable energy targets or current renewable energy percentages. Whether those commitments translate into the specific server running your query at a given moment is a harder question, and one worth asking suppliers directly. The point is not that the answer will always be satisfactory. The point is that the question is answerable, and the answer varies by provider.

For UK businesses, there is also no general regulatory requirement to avoid AI on energy grounds. As independent legal commentary confirms, environmental claims should be evidence-based and consistent with your wider sustainability reporting or procurement commitments, but no rule currently prohibits AI use because of its energy footprint. That may change as reporting standards tighten, which is another reason to build the habit of measuring now rather than guessing.

What believing this myth costs in practice

For Operations Managers and Managing Directors, the cost of blanket avoidance is not abstract. If a team is spending fifteen hours a week on manual document processing that a well-configured AI Assistant could handle in two, that is thirteen hours of capacity returned to higher-value work. Choosing not to pursue that gain for environmental reasons is a legitimate decision, but only if it is based on a fair assessment of the actual footprint involved, not on assumptions drawn from data centre infrastructure reporting.

Research published in September 2026 found that sustainability is emerging alongside cost, trust, and security as a meaningful adoption barrier for UK SMEs. That tells us the concern is not going away. It also tells us that businesses which develop a rigorous, evidence-based approach to environmental assessment will have a competitive advantage over those making decisions on vague grounds in either direction.

Sustainability Leads are in a position to turn this into a genuine asset. A clear internal framework for evaluating AI tools on environmental criteria, applied consistently to every procurement decision, is more defensible than either uncritical adoption or blanket refusal. It also positions the business well as supplier and customer expectations around environmental governance continue to rise.

A practical framework for assessing AI's footprint

Rather than asking "is AI bad for the environment", the more useful question is "what is the environmental impact of this specific tool, for this specific use case, at our expected volume of use?"

Three areas worth examining before committing to any AI tool:

1. Model size and task fit. Smaller models performing narrow, well-defined tasks use less energy than large generalist models applied to open-ended generation. If your use case is structured, use the smallest model that does the job reliably.

2. Hosting and energy sourcing. Ask suppliers directly about their data centre energy mix and published sustainability targets. Many now publish this information. If a supplier cannot or will not answer, factor that into your assessment.

3. Displacement versus addition. The most defensible environmental case for any AI deployment is that it replaces a material amount of manual work rather than simply generating additional output. An AI tool that reduces a team's document handling by half has a very different net footprint calculation than one that doubles the volume of content produced without reducing the effort required elsewhere.

None of this is a guarantee that every AI deployment will be net-positive on environmental grounds. Some will not be, and it is right to say so when the evidence points that way. The goal is to make that assessment on actual data rather than on the assumption that all AI carries the footprint of its most intensive applications.

The honest limitation here

The transparency gap in this space is real. If you ask a major AI provider for the per-query energy consumption of a specific model running on a specific data centre on a given day, you will not get a clear answer from most of them today. Independent auditing of AI energy claims is still nascent.

That means any environmental assessment you conduct now will be partly based on proxy indicators: data centre energy mix, model size, usage volume, and provider sustainability reporting. Those are reasonable proxies. They are not the same as verified, task-level data.

The practical response is to build your assessment framework now, document your reasoning, and revisit it as better data becomes available. A Sustainability Lead who can show the board that AI procurement decisions followed a consistent, evidence-based methodology is in a stronger position than one who either ignored the question or blocked all AI deployment without analysis.

A realistic approach for your team

For teams that want to move from gut feeling to a structured evaluation, the most useful first step is building shared competence around what questions to ask. gecco's Training and consultancy programme helps SMEs evaluate AI tools with clear criteria rather than assumptions. The vendor-neutral approach means businesses can assess specific platforms and use cases on their own terms, including environmental considerations.

Where to take this next

If your business is weighing AI adoption against sustainability commitments, the AI Readiness survey is a practical place to start. It is designed to surface exactly this kind of question, not just whether you are technically ready for AI, but whether your team has the criteria and confidence to make informed decisions about specific tools.

Take the AI Readiness survey. You will get access to 65+ free resources and a custom AI Readiness report. We then offer a free 30-minute showcase to walk through your results.

gecco's Training and consultancy programme is designed for teams that want to build that evaluative capability for themselves, applying it to AI tools across any function and any vendor, including the environmental dimension.

Website · LinkedIn · Case Studies · Newsletter