Most marketing agencies now use AI somewhere in their work. Very few can tell you how much, what it costs, or what it costs the environment.
Little Green Agency, as the name suggests, intends to be an ethical and environmentally conscious brand, so this is an attempt to set out the challenge honestly, including the numerous aspects that do not yet have a clear answer.
This is not an article about which AI tool is best, or a claim that we have solved anything. I’m just writing about the elephant in the room: AI is useful, sometimes incredibly productive, but it carries real costs that are easy to ignore precisely because they seem intangible.
Out of sight out of mind
When you run a query through an AI model, you don’t see any of the workings in the background. There is no noise or emissions, and no meter turning. That is the first of many reasons that the environmental impact is not always top of mind of the user. The processing is happening in a data centre far away, drawing electricity and, as we’re told, lots and lots of water for cooling.
Looking at best estimates online, the per-query cost is actually small. A single text query uses somewhere in the region of a fraction of a watt-hour of electricity, and approximately tens of millilitres of water once you count the water used to generate that electricity and not just the on-site cooling.
Mistral’s lifecycle assessment puts a typical 400-token response at about 45 ml of water and 1.14 grams of CO2 equivalent. Providers quoting on-site cooling only report far less, under a millilitre, which tells you how much the answer depends on what you decide to count.
The trouble is that these numbers come with wide margin error and unstated assumptions, and they climb sharply for longer, heavier tasks. A quick question and a long document analysis are not remotely the same size.
An independent academic study put a single short query at around 0.42 watt-hours, roughly 40% more than a Google search, and long prompts running to thousands of words use many times that. So anyone quoting you a single tidy figure for “the cost of an AI query” is overly simplifying things.
Multiply an individual’s modest use by hundreds of millions of users and hundreds of billions of queries a year, and you get the real issue: data centre expansion, strain of national grids, and water demand being skewed in specific locations.
That is a systemic problem, not one any single agency solves by being careful. As data researcher Hannah Ritchie points out, consumer-level AI use is only a fraction of total data centre demand. I want to be clear about that up front, because the alternative is the kind of inflated claim this agency exists not to make. Being deliberate about our own use is worth doing. It is not the same as fixing the industry.
Green is our biggest factor. It may not be yours.
For LGA, the environmental angle is one of the biggest considerations in how we use AI. That is a genuine position, not a marketing line, and it shapes the choices set out below.
But it is worth saying plainly that the environmental cost is only one of several reasons people raise concerns about AI. Depending on what you care about, your list might be led by data privacy and how your inputs are stored or used for training, by the treatment of the outsourced workforce that labels and moderates the data these models are built on, by the use of AI in military and state surveillance settings, by copyright and the training of models on creative work without consent, or by the documented cases of personal and psychological harm.
These are separate arguments with separate evidence, and reasonable people weight them very differently.
I mention this because there is no single “ethical AI” that clears every one of those bars. The tools most often held up as the cleaner options tend to be smaller providers scoring well on privacy and refusing military work, while the mainstream models most of us actually use for serious work score poorly across several of those axes at once.
Ethical Consumer’s guide to AI rates the tools across exactly these separate measures, and on the environmental one it found essentially everyone failing, with even the best-rated providers only beginning to disclose their impact. So this is not a field with a hero product. It is a field of tradeoffs, and the responsible move is to be honest about which ones you are making rather than to pretend you found the exception.
For transparency: we were using ChatGPT and now use Claude. I am not going to turn this article into a case for that switch, because the honest version of that decision involved capability and privacy considerations alongside the ethical ones, and I would rather write about the wider challenge than dress up one agency’s tooling choice as a moral verdict.
What we are doing about it: measuring, and maybe capping
You cannot manage what you do not measure, and most agencies are not measuring their AI use in any real sense. We are working on changing that for Little Green Agency.
The plan is to count token use, the unit AI models are billed and measured in, per client. Where our work runs through the AI provider’s API rather than a flat-fee chat subscription, that count is exact rather than estimated.
This gives us two things. First, honest cost accounting: we can see which work is AI-heavy, price it accurately, and spot where an automated process is quietly getting out of hand. Secondly, an actual basis for restraint, because a number you can see is a number you can set a limit against.
That second point is where we are considering capping token use per client: a budget that the work stays within, rather than open-ended consumption. However, counting tokens is not the same as reducing them, and a cap only means something if it changes behaviour rather than just recording it.
Measurement on its own would be fairly useless. But measurement that feeds an actual decisions (e.g. use less here, use a lighter tool there, do not use AI for this at all) I think is potentially really valuable. That is what I’m aiming for, but it would be best described as work in progress than claim it as a finished credential.
Where AI fits in our SEO work, and where it does not
The most open thing I can offer here is a map of our own process, showing where AI is genuinely required, where it is an option we could take or leave, and where a fast and simple model does the job.
One principle is that the lowest-impact query is the one you do not process. A good deal of SEO work does not need AI at all, and reaching for it by default is the habit worth breaking.
Keyword position tracking, technical crawls, log file review, analytics, backlink data: these run on dedicated tools and structured data, not language models.
Pulling a ranking report or auditing crawl errors is a job for the right software, and adding AI to it would be slower, less accurate, and a waste of resource. So a large share of what we do sits outside AI entirely, by choice.
Where AI is genuinely required, or close to it, is the work that is language-heavy and pattern-heavy at a scale that manual effort cannot match. Analysing a large set of search queries for intent and grouping them into themes is a real example, as is working through a big content inventory to spot gaps, overlaps, and cannibalisation across dozens or hundreds of pages.
That is the kind of task where the model is doing something that would otherwise take days and where the output genuinely improves the work.
Then there is a broad middle ground where AI is an option rather than a necessity. Drafting a first pass of a page, producing outline structures, summarising research, generating variations of a title or description to react to: AI can do these, and often quite well, but a skilled person can do them too.
The honest question here is whether the AI version is actually better or just faster, and whether faster is worth the cost on that particular task. Sometimes yes, sometimes no. Treating this middle ground as automatic is how AI use inflates without anyone deciding it should.
Within the work where we do use AI, the choice of model matters as much as the decision to use one, because a light model and a heavy one are not close in what they consume. A great deal of the routine work, classifying and tagging, extracting structured data from text, straightforward summarising, simple drafting, runs perfectly well on a fast, lightweight model. Defaulting to the most powerful model for these is the single most common way to burn resource for no benefit. We treat the light model as the default and make the heavier one earn its place.
The heavier reasoning models are worth their cost on a narrower set of tasks: genuine analysis that has to hold a lot of context at once, strategic work that involves weighing competing considerations, complex problem solving where the quality difference is real and matters to the outcome.
Reserving the powerful model for the work that actually needs it, rather than using it for everything because it is there, is both the better environmental choice and, usually, the better professional one, because it forces you to be clear about what each task actually requires.
Energy Tiers by Model Class (according to AI Overviews)
Haiku Models (Light / Fast): Consume the least energy, estimated well under 0.5 Wh per standard request, optimised for high-speed, low-complexity tasks. source
Sonnet Models (Balanced / Mid-Range): Standard middle-tier usage falls roughly in the 1–3 Wh range per typical text exchange, balancing capability and electrical footprint. source
Opus / Flagship Models (High-End): Large-scale frontier models (such as Opus variants) consume significantly more processing power, with estimates starting around 4+ Wh per baseline exchange and scaling upward based on token length. source
Extended Thinking / Effort Modes: Utilising high reasoning “effort” levels or extended thinking configurations multiplies the active compute load, increasing the per-query energy consumption substantially over base text generation. source
Will we be using AI moving forward then?
I do not think there is a clean way to use AI, and I would distrust any agency that told you there was. What I think is defensible is to use it deliberately: to ask whether a task needs it at all, to measure what you use, to reach for the lightest tool that does the job, and to be honest with clients about the real and modest cost rather than either ignoring it or overstating your virtue in managing it.
For Little Green Agency the environmental factor weighs heavily in those choices. It may not be the factor that weighs most for you, and that is a reasonable position too. Either way, the challenge is the same for every agency and most businesses now: AI is genuinely useful, its costs are genuinely real, and the responsible path runs between pretending neither of those things is true.
If you would like to talk about how AI does and does not fit into your own marketing, that is a conversation I am always happy to have.
Leave a Reply