A three-stage n8n workflow that monitors Gmail for incoming supplier invoices, extracts structured data using Claude’s vision API, and books the purchase mutation directly into e-Boekhouden — replacing 45 minutes of manual data entry per week with a 30-second fully automated process.

---

The Problem

A Dutch ZZP’er I work with receives around 40 supplier invoices per month. Hosting bills, software subscriptions, subcontractor invoices, the occasional printed PDF from a local supplier. They all arrive in the same Gmail inbox, in different formats, from different senders.

The previous process: open the email, download the attachment, read the PDF, manually type vendor name, invoice date, invoice number, amount ex-VAT, VAT amount, and ledger category into e-Boekhouden. Then file the PDF in a Google Drive folder with a consistent naming convention. Repeat 40 times a month.

That’s roughly 4–5 minutes per invoice, or about 3 hours a month spent on pure data entry. Low-value, error-prone work that nobody wants to do.

The client asked if it could just happen by itself.

---

What I Built

Three workflows in n8n, running on a self-hosted instance on a VPS:

Workflow 1 — Ingestion: Monitors Gmail every 10 minutes for emails with PDF or image attachments that look like invoices (subject-line filter + sender domain whitelist). Decodes the attachment, uploads it to a staging folder in Google Drive, and passes the file ID to Workflow 2 via a webhook.

Workflow 2 — Extraction: Receives the file ID, fetches the PDF from Drive, converts the first page to a base64-encoded image using a self-hosted Gotenberg instance, and sends it to Claude’s vision API with a structured extraction prompt. The prompt asks for a JSON object with these fields: vendor_name, vendor_kvk, invoice_number, invoice_date, amount_ex_vat, vat_amount, vat_rate, currency, ledger_hint, and a confidence score between 0 and 1. Claude returns structured JSON in about 1–2 seconds.

If confidence is below 0.85, the workflow sends a Slack message to the client with the extracted data pre-filled and a link to the original PDF. They can correct it in a Google Sheet and trigger Workflow 3 manually. If confidence is 0.85 or above, Workflow 3 fires automatically.

Workflow 3 — Booking: Takes the validated JSON, maps the fields to the e-Boekhouden REST API format, and POSTs a new purchase mutation (/v1/Mutatie) with the correct ledger account, relation ID (looked up from a mapping table), and VAT code. On success, it renames the PDF in Google Drive to YYYY-MM-DD_VendorName_InvoiceNumber.pdf and moves it to the archived folder. It also writes a log row to Google Sheets for the client’s own records.

---

How the Extraction Prompt Works

This part took the most iteration. A generic “extract invoice data” prompt gives inconsistent results, especially for Dutch invoices where VAT is called BTW, ledger accounts vary by business, and some invoices only show the total-inclusive amount.

The final prompt structure:

You are processing a Dutch supplier invoice. Extract the following fields as a JSON object.
If a field is not clearly visible or you are not confident, set it to null and reduce the confidence score accordingly.

Fields: vendor_name, vendor_kvk (KVK number if present), invoice_number, invoice_date (ISO 8601), 
amount_ex_vat (numeric, no currency symbol), vat_amount (numeric), vat_rate (9 or 21 as integer, 
or 0 for exempt), currency (ISO code), ledger_hint (your best guess at the expense category in Dutch: 
e.g. "software abonnement", "hosting", "kantoorkosten", "onderaannemer"), confidence (0.0 to 1.0).

Return ONLY valid JSON. No explanation.

The ledger_hint field feeds a simple lookup table that maps free-text categories to actual e-Boekhouden ledger account numbers. Not perfect, but correct about 85% of the time based on testing with 3 months of historical invoices.

---

The e-Boekhouden Integration

e-Boekhouden has a REST API (replacing the old SOAP interface as of 2026) with a Swagger UI at api.e-boekhouden.nl. Authentication is Bearer token, obtained by POSTing your API key to /v1/Authenticatie.

The mutation POST body looks like this:

{
  "Soort": "FactuurOntvangen",
  "Datum": "2026-04-15",
  "Omschrijving": "Hetzner Online GmbH - INV-2026-04123",
  "Betalingstermijn": 30,
  "Regels": [
    {
      "TegenrekeningCode": "4210",
      "BtwCode": "HOOG_VERK_21",
      "Bedrag": 82.64,
      "Omschrijving": "Hosting april 2026"
    }
  ],
  "RelatieCode": "HETZNER"
}

The RelatieCode and TegenrekeningCode values come from the lookup table. New vendors that aren’t in the table trigger a Slack alert and get queued for manual review.

---

Results

After running for 6 weeks with 40–45 invoices per month:

• 87% of invoices processed fully automatically, end to end, with no human involvement
• 13% flagged for review (mostly new vendors or handwritten-style PDFs from small local suppliers)
• Average processing time: 28 seconds from email arrival to booking in e-Boekhouden
• Manual time saved: from ~3 hours/month to about 20 minutes of exception review
• Zero missed invoices during the test period (previous workflow occasionally missed invoices buried in email threads)

The client now has a Slack channel called #invoices that’s either silent (everything processed fine) or sends a single message with a pre-filled correction form. Most weeks it’s silent.

---

What I’d Do Differently

The Google Drive staging step adds latency and complexity. A cleaner setup would stream the PDF directly from Gmail to Gotenberg without the intermediate Drive step. I kept Drive in because the client wanted a browsable archive of all processed PDFs, but if archiving isn’t a requirement, you can cut two nodes and about 3 seconds of processing time.

The confidence threshold of 0.85 was set conservatively. After seeing the actual error distribution, 0.80 would probably be fine for most invoice types and would reduce the manual review queue by half.

---

Stack

• n8n (self-hosted, VPS)
• Claude API — claude-opus-4-5 with vision input
• Gotenberg — PDF to image conversion (self-hosted, Docker)
• e-Boekhouden REST API — purchase mutation creation
• Gmail API — trigger and attachment extraction
• Google Drive — PDF archive
• Google Sheets — processing log and vendor mapping table
• Slack — exception notifications

---

Want to automate your invoice processing? Get in touch.