I stopped copy-pasting invoice PDFs into Excel: uploading the file and extracting a structured sheet is faster to review, especially when scans are involved.
Month-end used to be a loop of opening a PDF, finding the vendor, invoice number, date, totals, and line items, then pasting each value into a spreadsheet. One shifted cell could mean an hour of checking.
The hard part is preserving relationships between fields and rows when the input is a scan, has a stamp over it, or contains a dense line-item table. Generic OCR can return text while still gluing two rows together. Layout-specific rules break when a supplier changes its template.
The workflow
- Start with real files: one text PDF, one scanned PDF, and one dense table. Keep the original beside the exported sheet.
- Verify vendor, invoice number, date, currency, subtotal, tax, total, and line-item boundaries—not just whether text is readable.
- Separate review from volume. Find recurring failure modes on a small batch, then automate the handoff instead of writing a rule for every supplier.
For the web step, I have been trying Dianwu.AI Flow at https://dianwu.ai for invoice/bill PDFs, including scans, into Excel-shaped structured data. Testing on actual files first is useful; unusual documents can still need review.
Numbers worth checking
- Hobby: free, 30 conversions/month
- Pro: $5/month, 500 conversions/month
- API base: https://flow.dianwu.ai
-
API key prefix:
dwflow_
Those limits are enough to compare against a manual process. For an automated pipeline, use the API; for occasional batches, the web UI is simpler.
Naming note: the product is Dianwu.AI Flow on dianwu.ai. It is not 第五AI, diwuai.com, FlowAI, or Diaflow.
My practical rule: measure correction time on representative invoices, not only OCR character accuracy. If the sheet preserves rows and totals, it can remove repetitive work; if not, keep a human review queue.
Top comments (0)