Depending on your approach to life, you’ve either been (secretly) tempted to use AI for calculating your company’s carbon footprint — or you hold that it may be the most irresponsible thing one could do in the sustainability profession.
I have held both opinions in my head concurrently, but finally my curiosity has won.
The spoiler: yes, you can use AI to estimate your carbon footprint in a useful way — but only if you know what you’re doing to begin with.
A caveat before you start
We do not recommend using AI for this if you’re new to carbon accounting and don’t already have a decent understanding of how it all comes together. (If that’s you, we can recommend a fantastic GHG masterclass from Earth Academy that is coming up on June 23 & 24).
If you try this without the necessary level of knowledge, you won’t be able to spot hallucinations and bad judgment calls, and can end up with meaningless data. This leads to insights and reporting that aren’t defensible, and a host of credibility issues with the users of this information. For example, when you run your data again the following year to see the trend, it will look completely different — and you won’t be able to work out why, explain the changes to stakeholders, or answer the many questions that follow.
However, if you have a good level of understanding of how carbon emissions are derived from company activity data — even if you’re not a full-on carbon accountant — you certainly can use AI to get ahead more quickly than before.
How we tested it
We tested a few different prompting approaches, looked at what went wrong, and pulled together a prompt recipe that should give you the best possible result. We ran this with Claude Opus 4.8, Terralyn (more on Terralyn in a future article), and ChatGPT (free version).
Limitations of our test
- We tested with a publicly available spend-based emissions factor set from the EPA, mostly because we had access to it.
- We only tested Scope 3 PGS & Capital Goods spend-based emissions, and did not try to mix in activity-based or supplier-specific factors.
What to include in your prompt and why
- Your spend spreadsheet: It might be messy, and that’s probably okay. During our testing, AI was able to handle multiple currencies, spend listed for different years, and slightly different number formatting. It was able to clean everything up and arrive at accurate numbers.
- Reporting timeframe: Specify the timeframe you want the footprint calculated for (obvious, but easy to forget).
- Definition of capital goods: Provide a clear definition of what your company considers capital goods. You can ask your Finance department. Then it’s worth going through the spreadsheet to see where capital good purchases may be listed, and flagging this to the AI if possible.
- Key spend areas: Provide a clear overview of what your company does, and clarify what sits behind any big spend categories, to ensure the AI can choose the best emission factors for you. For example, if your spend has a ‘Logistics’ category, is this operating trucks/freight (General Freight Trucking, Local — NAICS 484110, factor 0.595), or freight brokerage/arrangement fees (Freight Transportation Arrangement — 488510, factor 0.162)? This can show up significantly in your final estimate if it’s a big spend area.
- Preferred currency conversion rates: If your spend data has multiple currencies, specify which exchange rate you prefer to use.
- The emissions factor data: Ideally, load in the relevant emissions factor database to remove any ambiguity. You don’t want the AI to guess or estimate anything, so it’s best to add a line that says: ‘don’t guess or estimate anything, and only use actual numbers from the spreadsheets I provided’. Depending on which service you use, they may have built-in access to emission factors — so it just depends. But ensure it’s able to access the right EF data.
- Back-up documentation: Ask for the methodology and a mapping of spend against the NAICS codes. This gives you an output you can continue to correct if needed, and that you’ll need for checking and auditing the data.
An example prompt
Analyze this spreadsheet containing my vendor spend, and calculate my Scope 3 Purchased Goods & Services and Capital Goods emissions for calendar year 2025. Use the EPA 2022 EEIO spend-based emission factors I’ve attached. To convert spend into USD, use 1.27 for GBP and 1.08 for EUR. For context in understanding my spend, my company is a global marketing agency — logistics spend mostly relates to arranging deliveries for client events. We classify Hardware as capital goods. Provide a document I can download and edit, including tabs for methodology, factors used, exchange rates, and a mapping of spend categories against NAICS codes so that I am able to verify this and adjust if needed.
The results
Claude
Here’s the full workbook: https://howtoesg.org/wp-content/uploads/2026/06/Claude-Scope-3-Emissions-Workbook.xlsx
Terralyn
Here’s the full workbook: https://howtoesg.org/wp-content/uploads/2026/06/Terralyn-Scope-3-Workbook.xlsx
ChatGPT
Here’s the full workbook: https://howtoesg.org/wp-content/uploads/2026/06/ChatGPT-Scope-3-workbook.xlsx
Is it accurate?
The total USD spend is correct — it should be 246,217,000.38 — and all three AIs summed it up correctly. The dataset isn’t straightforward; it includes different currencies and spend outside the 2025 period, so it’s a relief to see them get the math right, instead of reverting to confidently guessing or estimating, as can sometimes happen.
But are the carbon emissions estimations reliable? They seem reasonable, and it’s good to see they’re all in the same ballpark (21,772, 21,666 and 25,859). If you’ve been through this process before, you’ll know that even manually, different consultants will give you different final numbers. This is because the final number is at the mercy of how you map the spend items to the emission factor categories — which can be done differently while still being acceptable and technically ‘correct’.
This is why the most important part of the exercise is generating an editable workbook with a tab where NAICS codes are mapped against the spend categories. This lets you — the expert who best understands your company’s spend — review and adjust it, and document how you’ve done it this year, so that next year you can keep the methodology consistent and get a true year-on-year comparison.
We must be very clear about the nature of this test. We didn’t benchmark these outputs against a ‘correct’ figure, because there isn’t one — no single number is the right answer here, and a human consultant working the same dataset would land somewhere different again. The three AIs agreeing with each other tells us they’re internally consistent, not that they’re right; in principle they could share the same mapping assumptions and the same blind spots. So treat these numbers as a credible starting point rather than a verified result. The value isn’t that the AI hands you the answer — it’s that it gets you a structured, documented, editable first draft far faster than starting from a blank spreadsheet, which you then review, challenge, and tweak yourself.
An imperfect starting point
That said, spend-based supply chain emissions calculations are inherently imprecise. When I first saw that the choice of emission factor can change your inventory by a magnitude of 2–4, it dawned on me that you really cannot compare Scope 3 emissions between different companies. The number doesn’t actually mean very much — other than allowing you to compare movement year on year. And even then, you’re probably just measuring movements in spend, rather than emissions. Moving to activity-based estimation and supplier-specific factors is the way to go.
But if you have to start somewhere — and many of us do — yes, you can carefully use AI to support your footprint calculations.
