I turned a scanned coal invoice into a spreadsheet — what worked, and what it got wrong

June 23, 2026 · USA Sky PDF

If you do your own books, you know the worst part of month-end: invoices that arrive as a scan or a phone photo. There's no text layer, so a normal PDF-to-Excel tool gives you nothing — and you end up retyping every cell (descriptions, HSN codes, rates, tax lines) into a spreadsheet by hand.

We built a tool that OCRs the scan and rebuilds the tables as Excel, so the time goes to reconciling instead of transcribing. Here's a real test — not a cherry-picked clean one.

The test: a real, ugly scan

This is a coal-haulage invoice a vendor sent — image-only, scanned slightly crooked, the kind that defeats most converters because there isn't a single selectable character on the page.

Scanned coal-haulage invoice — an image-only PDF with no selectable text — Before: the scanned invoice. No text layer, so every figure would otherwise be retyped by hand.

What came back

Run through “Read this scan → Excel” (the AI extraction), it returned a structured spreadsheet — a goods table plus a separate sheet of invoice metadata:

The extracted spreadsheet: a goods table with description, HSN, quantity, rate, and amount columns — After: the extracted goods table. A second sheet captured the invoice number, dates, despatch/destination, and vehicle number.

Description	HSN/SAC	Qty	Rate	Per	Amount
Indian Coal	27011200	27.92	8701.62	MT	260352.47
OUTPUT COST			2.5	%	6508.31
OUTPUT SGST			2.5	%	6508.31
CESS			400	MT	11968
Total		29.92			285338.09

Extracted goods table (verbatim output)

Where it got it wrong — named

It is not flawless, and anyone who'd use this will catch the slips on tie-out anyway — which is exactly the point. Three to flag honestly:

So it doesn't replace your review — it replaces the typing. You proofread the result against the original; you don't key it in from scratch.

When this actually saves time

On a page with zero selectable text, this is about a minute of reconciling against the original versus ten minutes of manual entry. For a clean, text-based PDF you wouldn't need it at all. For genuine scans and photos, it's the difference between “usable” and “do it by hand.”

Cost and privacy, straight

The scan-reading shown here is a paid tier — $15/month for 300 pages. There's a free tier for ordinary text-based PDFs plus basic OCR, but cleanly reading an ugly scan is the paid AI feature; we're not going to pretend otherwise.

This one is not in-browser: the file uploads over HTTPS, is processed on a server, auto-deletes within 24 hours, and is never used for training. (Our plain editor tools run in your browser; this one doesn't.) If you handle sensitive client documents, factor that in.

Try PDF to Excel