GUIDE · GUIDES
Invoice scanning: what it captures, where it fails, and how to verify it
Invoice scanning turns a PDF or photo into draft AP data. Here is which fields it should capture, the failures that still need a person, and a verification checklist before approval.
Invoice scanning is the first step in most AP automation: take the document a vendor sent — PDF, email attachment, or phone photo — and turn it into structured data you can approve and pay. It saves typing. It does not, by itself, make the bill correct.
This guide covers what scanning should capture, where it still fails, and how to verify the result before it enters an approval workflow.
What invoice scanning actually does
Modern capture combines three things:
- Imaging — renders the PDF or photo into a readable page image, handling rotation and multi-page documents.
- Text recognition (OCR) — reads characters from the page.
- Field extraction — a model maps that text to AP fields: who sent it, how much, for what, and when it is due.
The output should always be a draft linked to the original document. If you cannot click from a field back to where it came from on the page, you cannot verify it — and verification is the whole point.
Fields a scanner should capture
| Field | Why it matters |
|---|---|
| Vendor name and remittance details | Matches the bill to the right vendor record |
| Invoice number | Your duplicate check — the same number from the same vendor should warn, not pay twice |
| Invoice date and due date | Starts the payment-terms clock and the approval deadline |
| Line items, quantities, unit prices | Lets an approver see what was actually bought |
| Subtotal, tax, shipping, total | The arithmetic must reconcile: lines + tax + shipping = total |
| Purchase order reference, if present | Connects the bill to what was ordered |
| Currency | Prevents a USD workflow silently paying a EUR amount |
In the Payvy Rust source today, email and upload intake store the original in S3-compatible storage, extraction creates a draft, and the draft is checked for total arithmetic and matched against vendor records. Extraction quality depends on the document and the model you configure — we do not publish an accuracy guarantee. See current invoice-processing capabilities.
Where scanning still fails
Plan your workflow around these failure modes instead of hoping they will not happen:
- Scanned and photographed paper — skew, shadows, folds, and low resolution degrade recognition.
- Handwriting and stamps — “Paid,” approval stamps, and handwritten corrections are routinely missed or misread.
- Multi-page and bundled documents — an invoice, a statement, and a credit memo in one PDF get merged into one draft.
- Tables without clear lines — quantities and prices shift columns, so line totals stop matching the invoice total.
- Look-alike characters —
0/O,1/l, and5/Serrors land in invoice numbers and amounts, which breaks duplicate detection. - Novel layouts — a vendor who redesigns their invoice can break a template that worked for a year.
None of these is a reason to avoid scanning. They are reasons the draft needs checks, not blind trust.
A verification checklist before approval
Run these in order. The first three catch most costly errors:
- Open the original next to the draft. Never approve from extracted fields alone.
- Reconcile the arithmetic. Line items sum to the subtotal; subtotal + tax + shipping equals the total. If it does not reconcile, stop.
- Check the invoice number against history. Same vendor + same number + similar amount is a duplicate until proven otherwise.
- Confirm the vendor and destination. A correct total paid to changed bank details is still a loss. Treat bank-detail changes as a separate, verified event.
- Read the due date and terms. Early-payment discounts and late fees both live here.
- Match to the purchase order or receipt, if you use them. Price and quantity variances should be visible exceptions, not silent edits.
- Record who verified it. The audit trail should show the document, the draft, any corrections, and the approver.
Rule of thumb: the scanner proposes, a person disposes. Any field you would not retype without looking, do not let software fill without looking either.
Choosing capture software
Ask these questions of any tool, including ours:
- Does every extracted field link back to the original document?
- Are arithmetic and duplicate checks built in, or is it your job to notice?
- Where do originals live, and who controls retention and deletion?
- Which model processes the document, and does document content leave your infrastructure?
- What does it cost when it works — and what work remains when it fails?
How this fits an AP workflow
Scanning is step one of five: capture, verify, approve, pay through your bank, and reconcile back to your books. Payvy is being built around exactly that chain, with agents preparing the draft and evidence and people authorizing payment. The hosted web experience is still in development; the intake and extraction above are in the open-source core.
Try the sample approval workflow, inspect the Rust source, or email support@payvy.com about your invoice volume — it helps us prioritize capture work.
Updated November 10, 2026. Product status reflects the reviewed Rust build; see Features for what is implemented vs. planned.