I think the sustainability debate around AI becomes much more useful when we bring it down to the level of everyday behaviour. It is easy to talk about data centres, emissions and electricity demand as if they sit far away from ordinary users. In reality, every prompt is part of that system. When someone asks a chatbot to write, summarise, analyse or generate something, the request is processed through physical infrastructure that needs electricity, cooling, hardware and network capacity.
One prompt is not the main problem. Google’s 2025 analysis estimated that the median Gemini Apps text prompt used around 0.24 watt-hours of electricity and 0.26 millilitres of water, while Epoch AI estimated that a typical ChatGPT query using GPT-4o may use roughly 0.3 watt-hours. Those figures are lower than some older estimates, and they should stop us from making exaggerated claims about one short text prompt. The wider issue is scale. The International Energy Agency projects that global data centre electricity consumption could more than double to around 945 TWh by 2030, with AI one of the main drivers of that growth.
This is where prompt quality starts to matter. A useful prompt that saves time, improves accessibility or supports better work may be worth its cost. A vague prompt that produces a weak answer, leads to five follow-up attempts, and is then mostly rewritten by a person, is harder to defend. The environmental impact of one inefficient exchange may be tiny, but poor prompting habits across large teams can create unnecessary compute, cost and rework. For data professionals, that makes prompting part of responsible AI practice, not just a writing trick.
What happens when a prompt becomes energy demand
A prompt begins as text on a screen, but the main work usually happens in a data centre. The prompt is sent to a model, broken into tokens, processed through specialist hardware, and returned as a response. If the answer is short, the work may be relatively light. If the user asks for a long document summary, a detailed code review, an image, or a multi-step agent workflow, the system may need to process more input, generate more output, or call tools several times before the task is complete.
The electricity is not only used by the chip producing the answer. Servers create heat, and that heat must be managed through cooling systems. Data has to move through networks, storage systems need to remain available, and the data centre itself needs supporting infrastructure. The carbon impact then depends on when and where the electricity is used. A request processed in a facility powered by cleaner electricity has a different emissions profile from the same request processed in a region with a more carbon-intensive grid. Water can also become part of the picture where cooling systems rely on it, although estimates vary and provider transparency is still uneven.
This is why the environmental impact of AI is best understood as a pipeline. A prompt becomes compute, compute becomes electricity demand, electricity demand creates emissions depending on the grid, and heat creates a cooling requirement. The user does not see that chain, but it still exists. The practical point is not that every prompt is harmful. The point is that prompts should be purposeful, especially when AI is being used repeatedly across an organisation.
How vague prompts create waste
Vague prompting is wasteful because it increases the chance of rework. A prompt such as “summarise this report” gives the model too little direction. The answer may be too long for the intended reader, too general for the decision being made, or too focused on details that do not matter. The user then asks for it to be shorter, then more practical, then more relevant to the business context, and eventually rewrites it anyway. Each step may feel small, but the workflow is inefficient because the task was not framed clearly at the start.
This is very similar to poor requirements gathering in analytics. If a stakeholder asks for a dashboard without explaining the decision it needs to support, the first version is unlikely to be right. The analyst then has to rebuild, reframe and clarify. The same principle applies to AI prompts. The model can only respond well to the task it has been given. If the task is unclear, the output will often be unclear too.
In my view, this is one of the most practical sustainability points for everyday AI users. Most people cannot choose the exact data centre their prompt runs in, and they cannot directly control the electricity mix behind a model. But they can reduce avoidable prompting. They can make the first prompt more specific, give better context, use the right source material, and choose the right tool for the task. That is not a complete solution to AI’s environmental impact, but it is a behaviour that scales.
A structured guide to lower-waste prompting
A good prompt does not need to be long. It needs to be clear enough to reduce unnecessary back-and-forth. The first step is to state the task precisely. Instead of asking the model to “help with a report”, ask it to “summarise this customer feedback report for a regional retail manager in five short paragraphs, focusing on delivery delays, stock availability and recurring service issues.” That gives the model a job, an audience and a boundary.
The second step is to provide context and evidence. If the answer should be based on a report, dataset, transcript or policy document, give the model that material and tell it not to go beyond it. This reduces the chance of generic output and helps prevent unsupported claims. For analytical work, this is especially important. A model asked to explain a metric without seeing the underlying definition may produce something plausible, but not necessarily correct.
The third step is to specify the output format and quality bar. A briefing note, a blog section, an executive summary and a technical explanation should not look the same. It is also worth stating what to avoid. For example, a useful prompt might say: “Avoid unsupported claims, do not introduce new statistics, and explain any uncertainty clearly.” That kind of instruction reduces the need for later correction because it tells the model what good looks like before it starts, although it cannot reduce the risks entirely.
A lower-waste prompt therefore has four parts: the Goal, the Context, the Sources and the Expectations of the output (GCSE). For example: “Using the customer feedback notes below, write a 400-word briefing for a store operations manager. Focus on the three issues most likely to affect customer satisfaction this month. Use only the information provided, explain where the evidence is weak, and end with two practical actions the manager could take.” This is more efficient than asking for a broad summary and correcting it five times afterwards.
Choosing when not to use AI
Optimising prompts also means knowing when a prompt is the wrong tool. If the task is to calculate a percentage, filter a dataset, retrieve a known figure, apply a fixed rule or run a repeatable transformation, a spreadsheet formula, SQL query or small script may be more appropriate. Using a large generative model for deterministic work can be inefficient, and in some cases it also creates more risk because the output may be less reproducible.
This matters for organisations because AI is often adopted through convenience. Once a tool is available, people begin using it for tasks that feel adjacent to its strengths, even when simpler methods would work better. A customer comment summary may benefit from a language model. A monthly sales total does not need one. A first draft of a communication may be a sensible use case. A basic database lookup usually is not.
There is also a model selection issue. Not every task needs the largest general-purpose model. Smaller models, retrieval systems, templates and conventional analytics workflows can often solve routine tasks with lower cost and less computation. The most mature organisations will not measure AI success by how much it is used. They will measure whether the right level of AI is used for the right kind of task.
What organisations can do about prompt waste
The first organisational step is to make AI usage visible. Teams already monitor cloud spend, software licences and storage. AI usage should be treated in a similar way. A basic dashboard could show how many model calls are being made, which teams use AI most heavily, which workflows generate long outputs, and where repeated prompting is common. This does not require perfect carbon accounting. It creates enough visibility to identify obvious waste.
The second step is to build prompt standards for common tasks. If teams regularly use AI to summarise reports, draft member communications, classify feedback or prepare meeting notes, they should not all be improvising from scratch. A small library of well-tested prompts can improve quality and reduce repeated attempts. It also makes review easier because teams can see how the output was produced.
The third step is to connect AI usage to value. Heavy usage may be justified if it saves time, improves accessibility, reduces manual rework or supports better decisions. Light usage may still be wasteful if the outputs are not used. The question should not be “how do we use less AI?” It should be “where does AI create enough value to justify the compute, cost and governance burden?”
For analysts, this creates a clear role. They can help organisations measure prompt volume, identify inefficient workflows, compare model usage with outcomes, and design better reporting around AI adoption. This is where sustainability, cost control and responsible AI meet. A better prompt is not only better communication. At scale, it becomes better resource management.
Final thought
The environmental impact of AI is not best understood through guilt about individual prompts. It is better understood through patterns of use. A single prompt may have a small footprint, but vague prompting, unnecessary retries, oversized models and automated workflows can create avoidable waste when repeated across teams and systems.
Good prompting is therefore a practical sustainability habit. It gives the model a clearer task, reduces rework, improves output quality and helps people use AI more deliberately. It does not solve the whole environmental challenge of AI, but it is one of the areas where everyday users and data teams can make a difference immediately.
Want to learn more? Our Copilot Expert course shows how to go beyond the chat and automate workflows across your entire digital ecosystem. No coding required.
Find your IoA membership package here
