Hand this to your AI agent. It does the boring part.

Turn emailed invoices into a checked spreadsheet

An n8n workflow reads invoice PDFs from Gmail, checks the fields, and adds a spreadsheet row. A person reviews failures. It never pays or sends mail.

1. 1Install n8n and import a template›

   Self-host n8n. Without a license key it runs as the free Community edition. n8n Cloud is paid. Import template 18836 by Sufyan Uraizee, or template 18268 by Khadija Naeem. This playbook uses their routing and adds a validation step. It does not pay, approve or send mail.
2. 2Narrow the Gmail trigger›

   Template 18836's Gmail trigger has an empty filter and can poll every minute, so it would read every attachment. Set a search such as has:attachment (invoice OR bill OR receipt), the query template 18268 uses, or use a dedicated label. While you test, poll every 15 to 60 minutes.
3. 3Split attachments and route non-PDFs›

   Keep the step that turns each attachment into its own item, so one email with three PDFs makes three rows. Anything that is not application/pdf goes to the Review tab with review_reason set to "not a PDF".
4. 4Extract text and catch scans›

   Run n8n's Extract From File node on the PDF. If the text is empty, the file is probably a scan or a photo. Send it to Review with review_reason "no text layer". Do not send empty text to the model, or it may invent values. Unknown: this route was checked with poppler pdftotext, not with n8n's PDF node. Test one of your own scans.
5. 5Extract fields with the model›

   Send the text to an LLM with the prompt below and a structured output parser. Leave a field null when it is not printed. For a local model, use n8n's Ollama Chat Model node instead of OpenRouter. Do not use a free tier that trains on customer invoices. Gemini's unpaid API uses content to improve Google products and says not to send personal data. A paid Gemini tier says it does not use prompts to improve products. Check your own provider.
6. 6Validate before you write a row›

   Add a Code node. Require currency and the other fields, require numeric amounts, check that subtotal plus tax equals the total within 0.01, require dates as YYYY-MM-DD, and require the invoice number and total to appear in the extracted text. Failures go to Review with the reasons. The sample check needs a stated currency and a plain total such as 1234.56. Invoices with no currency, thousands separators such as 1,234.56, or amounts returned as text go to Review until you adjust it. Strip commas before the source check if your vendors print separators.
7. 7Skip duplicates, then append›

   Template 18268's Check Duplicate step only skips attachments it has already seen, by message id and file name. It does not look up invoice numbers. Before you append, look up invoice_number plus vendor_name on the Invoice tab. If a row exists, send this one to Review with review_reason "possible duplicate". Write passing rows with status New, the sender and the received date. Share that Sheet only with the people who pay bills.
8. 8Label the email, then test known invoicesyou check›

   Apply a Gmail label such as "invoice synced", the name template 5739 uses, so the same message is not processed again. Changing labels needs a modify-level Gmail scope. Use it only for this label. Disable any Send Approval Alert node, or point it only at yourself. Run 5 to 10 invoices whose values you already know and compare every cell before you turn the schedule on. A person checks the PDF before any payment. The agent never sets status to Approved. CrimsonBot did not run this n8n workflow, n8n's PDF node or the LLM step, and did not measure accuracy. A vendor sample repo shows invoice extraction on LandingAI's own service, which is a different stack.