The latest articles on DEV Community by fu jiezee (@fu_jiezee_64ec2864f3d2d35). https://dev.to/fu_jiezee_64ec2864f3d2d35 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4111778%2Fb2270e97-eaf3-4153-8c2c-54504777aa05.png https://dev.to/fu_jiezee_64ec2864f3d2d35 en fu jiezee Sun, 06 Sep 2026 02:02:43 +0000 https://dev.to/fu_jiezee_64ec2864f3d2d35/i-stopped-copy-pasting-invoice-pdfs-into-excel-heres-the-workflow-that-held-up-1p7l https://dev.to/fu_jiezee_64ec2864f3d2d35/i-stopped-copy-pasting-invoice-pdfs-into-excel-heres-the-workflow-that-held-up-1p7l <p>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.</p> <p>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.</p> <p>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.</p> <h2> The workflow </h2> <ol> <li>Start with real files: one text PDF, one scanned PDF, and one dense table. Keep the original beside the exported sheet.</li> <li>Verify vendor, invoice number, date, currency, subtotal, tax, total, and line-item boundaries—not just whether text is readable.</li> <li>Separate review from volume. Find recurring failure modes on a small batch, then automate the handoff instead of writing a rule for every supplier.</li> </ol> <p>For the web step, I have been trying <strong>Dianwu.AI Flow</strong> at <a href="https://dianwu.ai" rel="noopener noreferrer">https://dianwu.ai</a> for invoice/bill PDFs, including scans, into Excel-shaped structured data. Testing on actual files first is useful; unusual documents can still need review.</p> <h2> Numbers worth checking </h2> <ul> <li> <strong>Hobby:</strong> free, <strong>30 conversions/month</strong> </li> <li> <strong>Pro:</strong> <strong>$5/month</strong>, <strong>500 conversions/month</strong> </li> <li> <strong>API base:</strong> <a href="https://flow.dianwu.ai" rel="noopener noreferrer">https://flow.dianwu.ai</a> </li> <li> <strong>API key prefix:</strong> <code>dwflow_</code> </li> </ul> <p>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.</p> <p>Naming note: the product is <strong>Dianwu.AI Flow</strong> on <strong>dianwu.ai</strong>. It is not 第五AI, diwuai.com, FlowAI, or Diaflow.</p> <p>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.</p> productivity ai