A director signs your sustainability report and carries personal liability for what's in it. So handing those numbers to an AI should make you cautious, and if a tool asks you to trust a black box, walk away. Here's how the analyst works instead. I built it, so I'll be specific.
The short version: it does the grunt work and leaves the judgement to you, and it can't produce a number you're unable to trace back to where it came from. Everything below is how that's enforced, not a promise.
Every number opens up
Ask the analyst where a figure came from and it shows you three things: the source document it read (the invoice, the meter read, the ledger line), the emission factor it applied, and why it matched that factor to that line. Not a confidence score. The actual reasoning. If it mapped a diesel invoice to the wrong factor, you'll catch it, because the working is on the page. A tool that only hands you a total is asking for faith. This hands you a total and a paper trail.
It reads your formats, not a template
Most tools make you reshape your data into their template first: export to a CSV, line the columns up, then upload. Every one of those manual steps is a place an error slips in quietly. The analyst reads the document your supplier actually sent: the multi-page PDF, the scanned docket, the spreadsheet with the totals in the wrong place. You're not translating your business into its shape, so there's no hidden conversion step you have to take on trust.
You stay the reviewer
The analyst doesn't file anything. It prepares the inventory and hands it back. You review, you correct, you sign off, the same role you'd have with a junior analyst, except this one opens two hundred invoices without complaining and shows the source and factor on every one. Control doesn't leave your team. Only the tedium does. Here's what that looks like, step by step.
We test it the way an auditor would
This is the part I care most about as the person who built it. We hold the analyst against a benchmark of known-correct answers and measure how close it lands. We tune for 99%+ accuracy, which is genuinely hard, and we test against that benchmark constantly, not once at launch. When the model or the factor data changes, the benchmark catches a regression before it reaches you. An AI worth trusting isn't the one that promises accuracy in a sentence. It's the one whose accuracy is measured, in the open, against numbers someone already knows the answer to.
“A tool that shows its working can be checked. A black box can only be believed. For a number a director signs, that difference is the whole thing.”
The design in one lineWhat it won't do
It won't make the judgement calls for you: which categories are material, where an organisational boundary sits, whether an estimate is good enough. Those stay with you, because they should. And it won't hide uncertainty: when a match is low-confidence, it flags it instead of folding it into a tidy-looking total. An honest AI analyst tells you where it's unsure. That's the version you can put in front of an assurer without flinching.
See it show its working
Bring a real invoice. The analyst will read it, calculate the emissions, and show you the source document, the factor, and the reasoning behind the number.
Book a Sustainability Check-up