Whatever your business — if an invoice lands on your desk and you need the figures off it, Extractor reads them. It lifts the parties, line items and totals off any invoice — any vendor, any layout, PDF or scan — entirely on your machine, and anything it can't read is flagged, never invented.
| Item | Description | Qty | Amount |
|---|---|---|---|
| STL-14 | Shelf bracket | 240 | 2,880.00 |
| OAK-32 | Oak panel 32" | 60 | 5,400.00 |
| FIN-08 | Matte finish kit | 45 | 1,012.50 |
| LBR-01 | Assembly labour | 32 | 2,240.00 |
Every figure sits beside the page it came from. The read is deterministic — numbers are lifted off the invoice, never generated — and an on-device AI re-reads the source to flag a misread or a line that doesn't add up. It can only flag, never write: no invented value reaches your books.
A transposed figure, a value read from the neighbouring column, a total that doesn't add up — surfaced before it reaches your books.
Extraction is rules, not a language model. It can misread, but it cannot invent — and the AI only ever flags, so no made-up figure slips through.
Reading, recognition and the AI check all run locally, offline. No invoice, no figure, no line item ever leaves the machine.
Most readers need a template per supplier, tuned once and brittle forever. Extractor finds the table by its headings and reads each field by its label — so a vendor you've never billed with works on the first try, in any layout, in many languages.
Measured on 170+ real invoices it had never seen — digital and scanned alike — Extractor read the figures that matter at about 99.7%. No template, no per-vendor tuning, and nothing trained on your documents. The numbers are read; they are never generated.
There's no market it's "built for." Every invoice carries a seller, a buyer, line items and a total — so whatever your business and whatever the layout, Extractor reads it. No template, no setup, nothing sent anywhere.
Finds the table by its headings and reads each field by its label — a supplier it has never seen works on the first try.
Anything the page doesn't state goes to Issues for a person to settle — blank-with-a-warning beats plausible-but-wrong.
Clean exports and camera scans run through the same engine — the accuracy barely moves between them.
Reading, recognition and the AI check all happen locally. No upload, ever — the confidentiality question is short.
Every field and line item written to a tidy Excel workbook you can file, reuse or feed into your own system.
A discrete graphics card speeds recognition several-fold; without one, it falls back to the processor automatically.
Seller and buyer details are learned on every run; the next invoice from the same party fills in what the page leaves out.
A subscription code unlocks the AI check and export — checked on the machine, no internet and no restart.
Reading and the side-by-side review are always free. A licence unlocks the on-device AI cross-check and Excel export — one code for your whole company, every machine, a flat monthly price for your volume — no per-page bills, no cloud, and your invoices never leave the building.
Reading and review are always free — no licence needed. Download Extractor →
Install in a click — no administrator rights. Reading and the side-by-side review are free; a subscription code unlocks the on-device AI cross-check and export.
It reads any invoice — the seller and buyer, every line item, and the totals — and lays each field beside the source page for review, entirely on your machine. A subscription adds an on-device AI cross-check and export to a clean workbook.
Any of them. There are no per-supplier templates: it finds the table by its headings and reads each field by its label, so a layout it has never seen works on the first try — digital PDF or scanned image, in many languages.
Yes. Reading, OCR and the AI cross-check all run locally on the machine. No invoice, figure or line item is ever uploaded.
Figures are read straight off the page by rules, not generated by a model — so a value the page doesn't state is never made up. Anything doubtful is flagged for a person to settle, and the AI check can only flag a mismatch, never write a value.