Start hereintermediate~25 min
Extract PDF Invoice Data Into a Spreadsheet Automatically (No Typing)
For a one-off, paste the invoice PDF into ChatGPT or Claude and ask for a clean table of the fields you need, then copy it into your spreadsheet. For invoices that arrive every week, build a pipeline in n8n, Make, or Power Automate that watches a folder or inbox, reads each PDF, and appends a row to Sheets or Excel. Always check the total and tax against the PDF before trusting the row.
Extract PDF Invoice Data Into a Spreadsheet Automatically (No Typing)
For a one-off, paste the invoice PDF into ChatGPT or Claude and ask for a clean table of the fields you need, then copy it into your spreadsheet. For invoices that arrive every week, build a pipeline in n8n, Make, or Power Automate that watches a folder or inbox, reads each PDF, and appends a row to Sheets or Excel. Always check the total and tax against the PDF before trusting the row.
What you'll end up with
A spreadsheet where each invoice is one tidy row: supplier, invoice number, date, due date, subtotal, tax, and total. No more opening a PDF, squinting at it, and retyping numbers into cells. You get two routes here. One is a 30-second paste for the invoice sitting in front of you right now. The other is a pipeline that watches a folder or inbox and fills the sheet on its own, so invoices that arrive every week land as rows without you touching them.
The one-off: paste and ask
For a single invoice, you don't need any setup. Open ChatGPT or Claude, attach the PDF, and ask for exactly the columns you want back. Keep the prompt specific so the output drops straight into your sheet: `Read this invoice and return a table with: supplier, invoice number, invoice date, due date, subtotal, tax, total. One row. Use YYYY-MM-DD dates and plain numbers, no currency symbols.` You get a clean table in seconds. Copy it, paste it into your spreadsheet, and you're done. Have ten invoices this month? Attach them all at once (both tools take multiple files) and add one line: "Return one row per invoice, plus a column for the source filename." Now ten minutes of retyping is one paste, with a filename column to audit each row against its PDF.
When to build the real pipeline
The paste method is great until invoices show up constantly and the copy-paste itself becomes the chore. Once you're handling them weekly, or several suppliers email PDFs to one address, it's worth wiring a pipeline that runs without you. The shape is always the same three moves: a **trigger** (a new file in a folder, or a new email with a PDF), a **read** step (AI pulls the fields), and an **append** step (a new row in Sheets or Excel). Build it once and every future invoice files itself.
Step 1: pick your trigger
Decide where invoices land. Two common setups: - **A folder** (Google Drive, OneDrive, Dropbox): you drop or save PDFs there and the workflow fires on each new file. - **An inbox**: suppliers email invoices to one address, and the workflow fires when mail with a PDF attachment arrives. In n8n, Make, or Power Automate, the first node is this trigger. Power Automate fits well if you're on Microsoft 365, since it reads Outlook and OneDrive natively. n8n fits if you want to own the workflow and point it anywhere. Start with whichever storage your invoices already use.
Step 2: read the PDF with AI
Add a step that sends the PDF to an AI model and asks for the fields. This is the same instruction as the one-off, written once and reused for every invoice: `Extract supplier, invoice number, invoice date, due date, subtotal, tax, total. Return only JSON with those keys. Dates as YYYY-MM-DD, numbers with no symbols. If a field is missing, use null.` Asking for **JSON** (a simple structured format of key-value pairs) matters here: the next step needs named fields it can map to columns, not a paragraph. If your PDFs are text-based, the model reads them directly. For scans, add an OCR step before this one so the AI receives clean text.
Step 3: append the row
Add the final step: **Append row** to a Google Sheet (Sheets node / Power Automate "Add a row") or **Add row** to an Excel table on OneDrive. Map each JSON field from Step 2 to its column: supplier to Supplier, total to Total, and so on. Include the filename or email subject in a column so every row traces back to its source. Run a test with one real invoice and watch the row appear. That single successful run is the moment the chore turns into infrastructure.
Step 4: build in the accuracy check
AI reads invoices well, but "well" is not "always," so never let a number reach your books unverified. Add a lightweight check before you trust a row: - Add a **flag column**. In the prompt, tell the AI to set it to "review" when the PDF is blurry, a field is missing, or subtotal plus tax doesn't equal the total. - In your sheet, **filter for "review"** and eyeball just those against the PDF. - Spot-check the **total and tax** on the rest, the two fields that matter most and slip most often. The quiet failure mode here isn't a wrong supplier, it's a number read in the wrong format: an invoice from Germany may write the total as `1.234,56` while your sheet expects `1234.56`. Tell the AI your format explicitly ("decimal point for cents, no thousands separators") and flag any total far larger or smaller than typical. This keeps human eyes only where judgment is needed.
Try it on one invoice now
Grab the most recent invoice PDF in your inbox. Open ChatGPT or Claude, attach it, and run the one-off prompt above. Paste the result into a fresh spreadsheet with your headers. That single round trip shows you exactly how clean the output is and which fields need a second look. Once you trust it on a handful of real invoices, you'll know precisely what to wire into the n8n, Make, or Power Automate pipeline, and the typing is gone for good.
Try this now
Your turn: open chatgpt and pick your trigger. Just do step one now — the rest takes minutes. Save this guide to pick up where you left off.
FAQ
How accurate is AI at reading invoice PDFs?
On clear, text-based PDFs it reads supplier, date, and total correctly almost every time. It slips on scanned or photographed invoices, odd layouts, and number formats like 1.234,56. Treat the total and tax as the two fields to verify on every invoice, and check supplier and date when the layout looks unusual. A two-second glance per invoice catches the rare miss.
Do I need to know how to code to build the recurring pipeline?
No. n8n, Make, and Power Automate are visual builders where you add a trigger and a few action steps, then fill in fields. You connect a "new file" or "new email" trigger to a step that reads the PDF, an AI step that returns the fields, and a step that appends a row. The only typing is the prompt and the column mapping.
Which tool should I use for the pipeline?
If your business runs on Microsoft 365, Power Automate is the easy path because it already reaches your Outlook, OneDrive, and Excel with sign-ins IT manages. If you want to own the workflow, point it at any storage, and keep AI costs low, use n8n. Make sits in between as a hosted visual builder. All three build the same watch-read-append shape.
What about scanned invoices or photos instead of clean PDFs?
Those need OCR (optical character recognition) to turn the image into text first. Modern AI models read many scans directly when you send the file as an image, but accuracy drops on low-quality photos. For a reliable pipeline on scanned documents, add a dedicated OCR or document-extraction step before the AI, and verify more carefully.