Can AI accurately calculate my carbon footprint from invoices?
AI can read quantities and units off invoices well enough to build a footprint, but accuracy depends on the controls around it, not the model alone. The arithmetic should be fixed rules (quantity times a named factor), and every figure should trace back to its source file so a person can check it.
Where AI fits and where it should not
There are three steps in an invoice footprint: read the line (supplier, item, quantity, unit, date), classify it (diesel, grid power, steel, freight) and calculate (quantity times emission factor). AI is useful for the first step and can help with the second. The third should never be left to a model's judgement. It is plain multiplication with a published factor, and it should give the same answer every time.
The ways invoice reading goes wrong
- Digit and decimal errors: 4,200 read as 42,000, or a comma read as a decimal point on a blurred photo.
- Wrong unit: kVAh taken as kWh, litres taken as kilograms, or a pack size taken as a quantity.
- Wrong line: the amount in rupees read as the quantity, or a tax line read as an item.
- Wrong classification: a lubricant classed as diesel, or a generator rental classed as fuel.
- Missing pages or lines in multi-page PDFs, and handwritten or stamped figures that are hard to read.
- The same invoice uploaded twice, or a scan and a photo of the same paper bill both counted.
Controls that make the result trustworthy
The GHG Protocol asks for an inventory that is complete, consistent, transparent and accurate, and for procedures that catch errors. These controls are how an AI-read inventory meets that, whatever tool you use.
- Fixed output: the same fields for every invoice, with blanks left empty rather than guessed.
- Classification by tax code first (HSN), with keyword matching used only when the unit also fits.
- No factor, no figure: a line with no matching published factor is shown as unmatched, not estimated.
- Range checks: flag quantities far outside normal for that supplier or month.
- Recalculation: recompute quantity times factor and compare with the stored figure.
- Human review of flagged lines, and a link from every number back to the file it came from.
Should you trust a claimed accuracy percentage?
Be careful with any single accuracy figure for AI extraction. It depends on the document mix, scan quality and how errors are counted. Ask instead what happens to a line the system is unsure about, whether you can open the source for any figure, and whether the same file always gives the same number.
Worked example: what one misread digit does
- Correct reading: 42,000 kWh x 0.710 kg CO2 per kWh = 29,820 kg CO2
- Misread as 4,200 kWh: 4,200 x 0.710 = 2,982 kg CO2
- Error: 26,838 kg CO2, about 27 tonnes understated for one month
- A range check against the previous month's units catches this before it reaches a report
What Senseible does here
Senseible reads invoices with a vision model at temperature 0 into a fixed schema, classifies lines by HSN first, applies only published factors and recomputes every line. Its Verified status means these rule checks passed; it is not an independent audit, and it publishes no accuracy percentage.
Sources: GHG Protocol Corporate Standard, Chapter 7 (managing inventory quality), ISO 14064-3:2019, verification and validation of GHG statements
Last reviewed 2026-10-15.