Receipt OCR: read what is written, nothing else.
When the merchant replies, their invoice arrives as an attachment in any state: a crooked photo, a till PDF, a sheet of glued receipts. The platform reads, structures and cleans, then lets a human decide whenever a character will not read.
What the reading extracts
Amounts, in cents
Totals, VAT per rate, unit prices: everything is converted to integer cents at read time. No floating-point rounding will ever shift a cent of VAT.
Dates, converted once
A receipt date is local to the shop. It is converted exactly once, at the same place in the pipeline, to universal time: no timezone drift can move an expense across days.
Product lines
Each line is read with its quantity and price, then compared against what the platform expects from that type of merchant. Deviations go to correction.
Several receipts on one sheet
A scanned sheet bundling several receipts is split into individual receipts, each read separately, each tied to its own expense.
The stance: no invented data
Character recognition is sometimes wrong, especially on a crumpled receipt. When a piece of information will not read, the platform says so: the value stays empty and the document goes to human review. Slower on one receipt — but the only rule compatible with a VAT deduction that must survive an audit. Full detail in receipt or invoice: what French law admits (FR).
Frequently asked questions
Do you handle receipt photos and PDFs?
Yes. Attachments received by email are processed in any common format: photo, PDF, scan. Sheets bundling several glued receipts are split receipt by receipt before reading.
What happens to the card number printed on a receipt?
It is purged at read time: only the last four digits are kept, because that is what matching needs. A dedicated test verifies that no full number ever enters the database.
And if the document is unreadable?
The platform invents nothing. An unread value stays unread and the document goes to human review: an operator decides. A guessed digit would be worse than a missing one.
Once read, the document is confronted with its expense: that is invoice ↔ expense matching.