Does AI make a company more sustainable, or less? A question worth answering before you scale it

Lately, the moment anyone brings up rolling out AI at scale, someone asks a really good question. And every time, the room goes quiet for a second. Afterwards, over coffee, several people told me that was the question that stuck with them most.

Right — we’re going to adopt AI in a big way. But is that a plus or a minus for the company’s environmental footprint? I did my MBA in the Netherlands, one of the most sustainability-minded places there is, in a program built around sustainability and society. A lot of my friends care deeply about this and work in the field. And honestly, I’ve often felt a bit guilty, because I didn’t know how to answer that question. So every time I hear it, I can’t help but laugh. Ah — there it is again. No escaping it.

It’s a good question. And honestly, the first answer that popped into my own head was: we’re already a company that contributes to sustainability. Our business model itself reduces waste.

That instinct is true. But I later realized it doesn’t actually answer the question.

In my MBA, there was a course called Managing Sustainable Business. Professor Bart Dierynck said something I still remember: negative impact can’t be undone by dividends. You can’t say the harm you caused has been cancelled out just because the company made money, paid dividends, and did a lot of good things. Those are two separate ledgers.

Bring that into today, and the line has a new version: a footprint can’t be undone by a handprint.

A company’s environmental impact was always meant to be counted in two parts.

One is the handprint: the positive impact your product creates for society. The circular economy, resale, extending the life of things — those are beautiful handprints. (But a caution to self: for this kind of “avoided emissions” to count, you need a clear baseline and additionality. Otherwise it just becomes another pretty story — and that’s exactly the thing we should care about most.)

The other is the footprint: what your own operations consume. Electricity, compute, carbon. Whether adopting AI increases your environmental burden is a footprint question. Running models eats power, and that makes your own operations heavier.

“Our product is sustainable” is talking about the handprint.

By the way, the popular framing is actually the opposite: grow your handprint big enough and it offsets your footprint. But as you’ll see in a moment, proper carbon accounting doesn’t let you subtract one ledger from the other.

That course also discussed a classic case. Shell caused pollution. So should it donate money to disaster relief for some natural catastrophe, or should it deal with the CO₂ it emits? The answer is the latter. Because what a company is truly responsible for is the externality it caused itself — not doing some unrelated, feel-good good deed. An unrelated good deed, however lovely, can’t stand in for your responsibility over your own ledger.

That question I keep hearing is a version of this exact problem. What people are asking about is the new footprint we’re creating ourselves. What the instinct wants to answer with is the good thing we were already doing.

The issue is that these are two different ledgers, and you can’t net one against the other. Under the double materiality logic that the EU’s CSRD adopts, positive and negative impacts have to be disclosed separately — you’re not allowed to offset one against the other. If you stop at “we’re already sustainable,” you simply haven’t done the math on your own footprint yet.

And the more a company’s story is built on sustainability, the more it’s worth proactively doing that math. Because trust is the moat for this kind of company. The earlier you honestly lay the operational side open too, the deeper that moat gets. Honesty is simply what this kind of company should look like.

So how do you count it, and how do you offset it?

I’m no sustainability-business expert. But from what I learned, here’s how I’d guess the implementation could work (I genuinely dug up my old notes and coursework for this — hoping I didn’t get too much wrong 😅):

Counting

Three sources, from rough to fine: use public third-party estimates to get the order of magnitude; get a carbon footprint dashboard from your cloud provider; ask your AI vendor for the emission factor per million tokens. From what I found, the carbon of a single text query lands somewhere around a fraction of a gram of CO₂e — but that number can swing by one or two orders of magnitude depending on output length and the carbon intensity of the local grid, so it’s only good for grabbing the order of magnitude, not for precision accounting. And even the most frontier AI labs still disclose very little detail about their operational emissions. The measurement methods themselves are still taking shape.

For companies adopting AI via API or SaaS, this carbon almost all falls under Scope 3 — purchased services.

A quick primer: under the internationally used GHG Protocol, emissions are split into three tiers. Scope 1 is what you burn directly (company vehicles, on-site boilers). Scope 2 is the electricity and heat you buy and use. Scope 3 is every other indirect emission across your value chain — purchased goods and services, business travel, and yes, the cloud and AI services you buy. It’s usually the biggest of the three, and the hardest to measure.

There’s a line in my MBA notes that puts it bluntly: the escape routes are closing. Once CSRD kicks in, the number of companies in scope expands from just over ten thousand to fifty thousand — even small companies have to follow. And Scope 3 is exactly the tier most companies leave out of their sustainability reports, and the one that’s now hardest to keep dodging. When AI comes in, this is usually the part that gets missed.

Offsetting

Credible decarbonization is a ladder, and offsetting is the last step, not the first. First, don’t waste at the architecture level. Then switch to clean power — but there’s a quality difference here: a PPA that can prove it brings new renewable generation online is far more credible than buying unbundled renewable energy certificates, which in recent years have been criticized as having almost no additionality. Only the residual you genuinely can’t cut should be handled with carbon removal. What Anthropic bought when it joined Frontier is exactly this: carbon removal.

And here’s the key point that ties directly back to everything above: even after you buy offsets, proper carbon accounting still requires you to disclose your gross emissions honestly, and list the removals separately. You can’t subtract them down to zero and then declare “we’re net zero, don’t worry.” Buying carbon credits and declaring the problem solved comes from the same impulse as brushing it off with “we’re already sustainable”: both are a little too eager to balance the books early. But honest accounting is just slower by nature — and doing it beats not doing it.

But… don’t let sustainability become the brake on going AI-native

The most important thing when rolling out any system — AI-native included — is to get people to feel free to use it, and use it a lot. People only understand AI’s value once they’ve actually used it. If a company hands out the tools while also chanting “go easy, think about the carbon,” employees feel guilty, they hesitate, and adoption never takes off. And then, before you’ve even verified whether the net effect is positive or negative, you’ve already suppressed the value side. That’s picking up the fraction of a gram of carbon per query and dropping the entire leap in productivity.

I think the better approach is to move the responsibility for the footprint away from the individual employee’s usage decision and onto the system’s architecture decisions. And disclose those efforts at the same time — to build employees’ sense of ownership along with it. The fact that people ask this question at all tells me they genuinely care.

It’s not about saying how sustainable you are

We’re living through an industrial revolution. There’s no going back to life without the lightbulb. Nearly every company is going to adopt AI, so the real dividing line won’t be whether you use AI — it’ll be whether you take on the social responsibility that comes with it. My own view is that a company’s real sustainability maturity isn’t about whether it can say how sustainable it is. It’s about whether, when faced with a new question that doesn’t yet have a pretty answer, it’s willing to first say, honestly: that’s a good question. Let’s do the math.