AI for Business
Scanned Invoice & Receipt Reader
Photograph or scan an invoice or receipt and get the vendor, numbers, dates, taxes and every line item back as a table you can download — the picture is read directly, so a scan with no text layer works.
This is a cloud tool. The picture of the invoice is sent to the model, resized on your device first; anything printed on it goes with it. Nothing is stored by us; the call is counted against your month’s allowance — 10 free, 300 on Pro. How the AI tools handle data.
Tips
- Try a sample fills the notes only — a sample cannot attach a picture. Add a photo or scan of any invoice, then press Read.
- Photograph the page straight on, in good light, with all of it in frame. A crooked or shadowed photo still reads; faint print and small totals are where mistakes creep in.
- The picture is resized on your device so its longer side is 1,600 pixels — enough for a full A4 page — and sent as a JPEG. A PDF is rendered here, first four pages.
- Check the totals and the GSTIN against the paper. legibility and unreadable_fields say where the model was unsure; a null is a field it could not read, not a zero.
- A typed or exported PDF is better served by the Invoice & Receipt Data Extractor, which reads the text layer and sends no picture at all.
Frequently asked questions
What is sent, and where?
The resized picture itself and a short instruction, through our gateway to the model — which is different from the text tools, where only text ever leaves your device. The personal-data shield masks numbers in the notes you type, but it cannot mask pixels; if the invoice shows something that should not travel, cover it before you photograph it. We store neither the picture nor the answer; only the fact of a call is counted.
How accurate is it on a photo?
Good on a clean, well-lit page: the number, the dates, the GSTIN and the line items come back correctly most of the time. Handwritten invoices, faded thermal receipts and blurred or crumpled photos read worse, and the model is told to return null rather than guess. Treat it as a first draft of data entry that you check against the paper.
Is this OCR?
Not in the classic sense: no text layer is produced. The model looks at the picture and writes the fields, which is why it copes with stamps, handwriting and odd layouts that OCR trips on — and why the numbers still need your eye.