I have spent a couple of years building carbon footprint software and, every time someone talks to me about "AI for sustainability", I get a little defensive. Not because I do not believe in AI. I use it every day and it is at the core of what we do at Grumbic. It is because it almost always gets sold from the wrong angle.
The fashionable promise is that AI will "solve" emissions calculation, as if the problem were the math. It is not. Multiplying a consumption figure by an emission factor is something any spreadsheet can do. The hard part, the one that eats up weeks, is something else: collecting and organizing the data.
Where AI actually changes things
A corporate footprint is built from hundreds of documents nobody designed for this. Electricity bills from three different retailers, fuel receipts, travel spreadsheets that each team fills in its own way, lease contracts with consumption buried in an annex. Each one with its own format, its own naming, its own units. That is the work that in most companies is still done by hand, copying numbers from a PDF into a spreadsheet for weeks.
This is where AI is great. Reading an invoice and pulling out the kWh, recognizing that a charge is gasoline and not diesel, classifying which scope and category each item belongs to. It is repetitive work, unglamorous, prone to human error, and a model does it in seconds. For me that is the real use case, and it is not flashy: taking the transcription of receipts off the sustainability team's plate.
The part I am most excited about comes after the calculation
If I am honest, extraction is not what excites me most. It is what you can do once the data is in and properly organized.
When the inventory is built on a methodology that holds up, you can start talking to it. Ask it directly what you are missing. An AI that knows the methodological structure knows which scopes and categories should be there, so it tells you, in plain language, where you have a gap. That you did not load the air travel, that a supplier's Scope 3 is missing, that one site's consumption looks off against the rest of the year. You do not have to guess or scroll through forty spreadsheet tabs looking for the figure that does not add up. You ask and it answers.
That completely changes how you review the information. Instead of reading a static hundred-page report, you explore. Show me emissions by site, compare them to last year, where is the biggest concentration, what happens if I drop this plant from the calculation. The footprint stops being a PDF you file away and becomes something you can interrogate, each person with their own questions and in their own way. For a sustainability lead that is worth more than any prebuilt dashboard, because the questions that matter are almost never the ones someone anticipated when designing the screen.
That said, all of this rests on the structure underneath being well built. And that brings me to my strongest opinion.
My strongest opinion on the subject
An AI on its own, without the methodology on top, produces wrong numbers with a lot of confidence. And that is worse than having no number.
I recently wrote about why guarantees of origin do not erase the emissions from transmission losses. It is a good example. A model that only learned to sound reasonable would put that line at zero without hesitating, because it sounds logical: you bought renewable, case closed. The standard says otherwise. And you only know that if the rule is written into the system, not if you left it to the model's intuition. AI does not understand the GHG Protocol. It reproduces patterns, and a plausible pattern can be completely wrong.
That is why I cannot stand products that treat the calculation as a black box where an invoice goes in and a number comes out, with no way to see what happened in between. In carbon accounting, "trust me" is not an answer.
How we build it at Grumbic
I separate the two on purpose. The AI does the grunt work: read, extract, classify, propose. The methodology is what rules the final number, and those rules are written out explicitly, not learned by memory. The GHG Protocol, MITECO where it applies, the treatment of each scope and category. The AI proposes, the standard disposes.
And everything stays traceable. When a verifier shows up, or when a company reports under CSRD, nobody accepts "the AI calculated it". You have to be able to show which invoice each figure came from, which factor was applied, and why. AI gets you to that point much faster, but traceability is what makes the number hold up when someone reviews it seriously.
What I actually believe
My bet is not AI instead of rigor. It is AI to wipe out the tedious part and to make reviewing the footprint stop being an act of faith. The sustainability team stops transcribing documents, can ask the inventory questions instead of deciphering it, and gets to focus on what matters: deciding what to do with the emissions the calculation puts in plain view.
The number is still just as serious. What changes is that getting to it no longer takes three months and no longer depends on nobody making a mistake copying a figure. That, to me, is the honest promise of AI in this field. Quite a bit less epic than the marketing version, and quite a bit more useful.
