Are You Using More AI Than You Realise? How to Right-size AI at Work

More than once recently, someone has told me, “I don’t use AI because of the energy.” And I get it. AI’s environmental impact is real, and people are right to care about it.

I use AI a LOT… like LOADS, and I love what it helps me do. But I also care about the environment, and those two things don’t sit entirely comfortably together. I’m not interested in pretending AI is harmless, but I’m definitely not convinced that stopping using it altogether is the answer either.

That discomfort is what prompted me to do some proper research into how to use AI more sustainably at work. What genuinely makes a difference, rather than simply repeating the claims that I’ve heard bandied around?

The conversation around AI sustainability often gets stuck at one enormous yes-or-no question: should we use AI at all?

But, I don’t think that’s the most useful question.

If you want to use AI more sustainably at work, a better question is: are you using the right amount of it for what you’re trying to achieve?

Right-sizing AI means using the right amount of AI for the job: enough to create worthwhile value, without generating work the task doesn’t need. In practice, it means matching the tool, the amount of thinking, the output and the autonomy to the job, while noticing when AI is producing something nobody uses.

Because asking Copilot to tighten two sentences isn’t the same as asking a reasoning model to investigate a complex problem. A normal web search isn’t the same as an AI answer built from several searches. And a meeting summary you deliberately use isn’t the same as an automatic summary nobody opens.

That’s the idea behind Right-size AI.

Don’t use less AI. Waste less AI.

What Does It Mean to Right-size AI at Work?

If you want to use AI more sustainably at work, the useful question isn’t whether every AI interaction can be made as small as possible. It’s actually whether the amount and type of AI are proportionate to the value of the task. It also means noticing when AI may be running because of a previous choice, a product setting or an organisation-wide default.

This isn’t about choosing the smallest model every time or treating a long answer as an environmental misdemeanour. Where your tool offers automatic routing, ‘Auto mode’ can be a sensible everyday starting point. But then sometimes deeper reasoning is exactly what the task needs.

The point is to make the choice proportionate.

Use a quick route for a quick job. Ask for the amount of output you’ll actually use. Give agents jobs that genuinely need several steps. Keep more compute-intensive AI when it creates enough value to justify it. And check the recurring AI output that’s slowly accumulating because someone ticked a box six months ago.

How Much Environmental Impact Does AI Actually Have?

The International Energy Agency expects data-centre electricity demand to grow substantially this decade, with AI an important driver alongside other digital services. So this isn’t a made-up problem, and improving efficiency per task doesn’t automatically solve the wider issue if total use keeps growing.

But some of the most repeated per-prompt claims are far shakier than they sound.

The familiar claim that an AI search uses ten times the energy of a traditional Google search sounds precise. But without a comparable task, system and measurement method, it isn’t a reliable guide to the choice in front of you.

And the footprint of an AI prompt can change materially with the model, hardware, output length, reasoning and number of underlying operations. Research by Oviedo and colleagues estimates that, under its modelled conditions, extended test-time reasoning can use more than ten times as much energy as a typical query. That’s useful directional evidence, but it isn’t a measurement of every live tool or every prompt.

So we need enough honesty to say two things at once:

1. AI uses real energy and the choices around it matter.
2. We can’t turn every click into a neat carbon receipt.

What Our Research Changed About Sustainable AI Advice

When we started looking into this, I expected some recommendations to be easy wins.

Better Prompts Aren’t Automatically Greener

“Write a better prompt” sounded like an obvious route to lower impact. It may improve the answer and reduce unnecessary back-and-forth, which is useful, but the evidence wasn’t strong enough for us to claim that better prompting automatically makes AI greener.

Smaller Files Aren’t Where Most People Should Focus

We also investigated whether people should choose Markdown instead of Word because a smaller file might require less processing. The technically correct answer may be that file size can affect work at the margins. The useful answer for an ordinary workplace user is: please don’t spend your afternoon worrying about it.

Running AI on Your Own Device Isn’t Automatically Greener

Some AI can run directly on your laptop, phone or other device rather than relying entirely on cloud data centres. That can be efficient for some tasks, especially on hardware you already own, but it doesn’t mean “on-device AI is always greener”. The model, task, device, electricity and hardware lifecycle all matter.

That changed the emphasis of our advice. We couldn’t honestly turn this into another checklist of tiny ways to shrink every prompt. The more meaningful choices sit around the amount of reasoning, the amount of output, the number of operations, recurring automatic use and the value created.

What About AI That Runs Automatically?

It’s easy to focus on the moment you open a chatbot and type a prompt – but that leaves out AI triggered by an earlier choice or setting.

AI is increasingly built into search, browsers, inboxes, documents and meetings. Sometimes it’s simply there as an option; sometimes an organisation has enabled it; and sometimes it’s doing work without you actively asking – like generating a meeting summary automatically.

Those aren’t the same thing.

A visible Copilot icon isn’t evidence that a model is running. By contrast, an automatic meeting-summary setting can start new processing when an eligible meeting begins.

Google’s AI Overviews can appear automatically when its systems decide a generated answer may be helpful. Google doesn’t provide a universal AI Overview off switch, although its documented Web filter shows text-based links without the AI Overview in that result view.

If you’re trying to find a login page or a particular source, Web can give you the links you wanted. But changing the view doesn’t tell us whether AI processing has already happened. Treat it as a way to simplify the results, not as a proven energy-saving switch.

The same principle applies to meetings. An AI summary can be brilliant when it captures decisions, saves follow-up work and supports people who couldn’t attend. But if every recurring catch-up generates a summary nobody reads, that’s worth reviewing. The question isn’t “Are summaries bad?” It’s “Are we using them?”

And that same question… are we using the right amount of AI for the job?  …applies much more widely. Here’s how:

Seven principles for using AI more sustainably at work

To right-size AI at work:

1. Match the thinking to the task.
2. Ask for only the output you’ll use.
3. Iterate with a reason.
4. Give agentic AI a real job.
5. Use the simplest tool that does the job.
6. Judge the whole outcome, not just the prompt.
7. Notice AI being used automatically for you.

The detail matters, though, because none of those principles is an instruction to choose the smallest possible option every time.

1. Match the Thinking to the Task

Start by asking how much thinking the job really needs.

If you want three clearer subject lines or a concise summary of text you’ve already checked, a quick response may be enough. If you’re comparing competing strategies, testing a high-stakes recommendation or working through an ambiguous problem, deeper reasoning may earn the extra work.

In Microsoft 365 Copilot Chat, Auto routes the prompt according to the task. Microsoft’s guidance also describes Quick response and Think deeper, although the controls available can change. Auto is a reasonable everyday starting point; but it isn’t a guarantee of the lowest energy use. You don’t need to turn model selection into a major new hobby… just notice the moments when you’ve asked a complex engine to move a very small pebble.

2. Ask for What You’ll Actually Use

AI will cheerfully give you twenty ideas when you only have room for three.

The environmental point here is simple: longer outputs require more generation. And the practical point may be even stronger: every extra paragraph becomes something you have to read, assess, edit or ignore.

So, tell the AI what will genuinely help. Ask for five options, not every option it can think of. Request a one-page briefing if that’s what the meeting needs. If you need depth, ask for depth. But don’t generate volume as a substitute for deciding what good looks like.

3. Iterate With a Reason

Iteration is useful. Endless regeneration isn’t. If an answer misses the tone, evidence or structure, coach it. Explain what needs to change and why. That’s far more useful than pressing regenerate and hoping the next roll of the dice will somehow understand what the first answer didn’t.

Before another round, ask: what am I trying to improve? If you can’t name it, the output may already be good enough, or you may need to make the decision yourself.

4. Give Agentic AI a Real Job

One visible request can hide a surprising amount of work.

Agentic AI goes beyond asking a chatbot for an answer: you give it an outcome and it can work through multiple steps and use tools to get there. Think Copilot Cowork or ChatGPT Work.

A genuinely agentic workflow may plan several steps, search multiple sources, use tools, inspect files, retry failed actions and validate the result. That can be enormously valuable when the task genuinely needs autonomy and follow-through.

It can also be an extremely elaborate way to do something that needed one prompt and a calendar reminder.

Use agentic AI when the outcome needs several connected actions and the autonomy creates worthwhile value. For a one-step job, use a one-step tool. Also be careful with the label: not everything marketed as an “agent” is truly agentic.

5. Use the Simplest Tool That Does the Job

Sometimes you need AI generation. Sometimes you just need a calculator.

If you do the same calculation every week, a formula will probably do the job better. If certain emails always need to go to the same place, use a rule. If you’ve already got a template that works, you don’t need AI to reinvent it. And sometimes a good old-fashioned web search will get you to the answer faster.

That doesn’t mean those tools have no environmental impact. They do. The point is simply that we don’t need to add AI to a job when something simpler already does it perfectly well.

But there is a flip side. If AI makes a difficult task easier, removes a barrier or helps someone do something they couldn’t otherwise do, that’s real value. The simplest tool isn’t always the least powerful one. It’s the one that does the job well without adding work you don’t need.

6. Judge the Whole Outcome, Not Just the Prompt

A deeper or longer AI interaction can still be worthwhile if it prevents a costly mistake, improves access, saves duplicated effort or materially improves the result. If it replaces something with a larger footprint, such as genuinely unnecessary travel, there may also be an environmental benefit. But saving time or improving quality doesn’t, by itself, prove an energy or carbon saving.

So what I mean by that is we shouldn’t optimise one tiny part of a workflow and ignore everything around it.

Ask what the AI changed. Did it create usable value? Did it reduce other work? Did it improve quality? Did it merely produce an impressive-looking document that now needs two people to untangle it?

This means that the goal becomes the useful outcome, not the lowest possible token count.

7. Notice the AI Being Used for You

This is the principle our earlier guidance was missing.

Look beyond the prompts you deliberately type. Check whether search answers, inbox analysis, meeting recaps or recurring AI workflows are being generated because of previous choices or organisation defaults. That includes things you’ve set up yourself — perhaps a scheduled AI update or regular report that was useful at the time but is now arriving every week without anyone really using it.

For each one, ask:

1. Is this feature generating, predicting or simply automating?
2. Did I ask for it this time, enable it earlier or inherit the setting?
3. Is anyone using the output?
4. Is there a lighter route that preserves the value?
5. Would switching it off remove accessibility, safety, legal or meaningful productivity benefits?

That last question matters. Captions, transcripts and summaries can support inclusion and reasonable adjustments, and sustainability isn’t a good excuse for making work less accessible.

What Should You Focus On, and What Isn’t Worth Obsessing About?

If you want to use AI more sustainably at work, focus on choices that can materially change the workload, remove repeated unused output or protect something genuinely valuable.

This is the difference between meaningful action and ‘eco-theatre’. One asks what work is happening and whether it creates value. The other hunts for small things that look virtuous, whether or not they make a meaningful difference.

How Can Organisations Review Unnecessary AI Use?

Pick one recurring area: search, meetings, mail or agentic workflows.

Then review three things:

1. Does AI actually run automatically, or is the feature merely available?
2. Is the output used? Look for views, edits, follow-up actions or a clear accessibility, safety or compliance need.
3. Would a lighter route preserve the value? That might mean making summaries opt-in, choosing a quick model or using a normal automation. If a setting is managed centrally, take the question to its owner.

Don’t try to audit every system in one heroic afternoon. Start with the area most likely to be producing recurring output. Meetings are often a good candidate because a single auto-start setting can repeat week after week.

Use Enough AI to Do the Job Well

I don’t want this conversation to end with people feeling guilty every time they open Copilot. Guilt is a fairly poor operating model, and it doesn’t help anyone make a better decision.

The practical version is much simpler. And, for me, this is the most honest answer to how we can use AI more sustainably at work.

Use AI when it adds value. Use more of it when the job genuinely benefits. Use a lighter route when it doesn’t. And notice the AI that may be running because nobody has asked whether it’s still useful.

That’s Right-size AI.

Download The Right-size AI Check to choose a proportionate route for your next task.

Frequently Asked Questions