The Invoice Pipeline

Description

GOAL
Build a workflow that automatically extracts invoice data from incoming PDF invoices, matches it against purchase order data and goods receipt in the ERP, and triggers the booking in the accounting system on a match. On discrepancies, the workflow notifies the finance team with full details and flags the case for manual review instead of booking it through automatically.

Adapt the structure to the concepts of your automation tool (for example nodes, steps, zaps, or agent actions). While doing so, preserve the trigger logic, the sequence, the branching, and the data flow.

TRIGGER
Type: new email with a PDF attachment in the dedicated invoice inbox, for example [email protected].
Filter: only attachments of type PDF, other emails are ignored.

STEPS
1. Read the PDF: obtain the text via OCR or extraction from the attachment.
2. AI step: data extraction. Use exactly the Invoice Data Extraction working prompt (see the Working Prompts section). Output: JSON with all invoice fields.
3. Duplicate check: check invoice number and supplier together as a unique key against already processed invoices. On a duplicate, jump directly to step 7 and flag the case as a duplicate.
4. Load purchase order data: load the purchase order data via the purchase order number from the ERP (SAP, Dynamics, or Odoo). If the purchase order number is missing, use matching by supplier, amount, and date as a fallback.
5. AI step: matching. Use exactly the Discrepancy Analysis working prompt (see the Working Prompts section) with the invoice JSON and the purchase order JSON as input. Output: JSON with a recommendation.
6. Branch: if the recommendation is book_automatically, trigger the booking in the accounting system (DATEV, Lexoffice, or comparable) with cost center and account mapping.
7. Otherwise, or on a duplicate or a missing purchase order: escalate to the finance team via ticket or a message in Slack or Teams, with invoice data, purchase order data, and all discrepancy details. Flag the case for manual review.
8. Both branches: log the incoming email, the extracted JSON, the match result, and the final status in the audit log.

PROMPTS
Take over both working prompts from the section of the same name unchanged and in full. Do not alter the placeholders they contain, and map them as described in the data flow.

DATA FLOW
Step 1 to step 2: extracted invoice text.
Step 2 to steps 3, 4, and 5: invoice JSON.
Step 4 to step 5: purchase order data as purchase order JSON.
Step 5 to steps 6 and 7: the recommendation, discrepancies found, details, and rationale fields from the match result.

ERROR HANDLING
Incomplete extraction, for example a missing invoice number or gross amount: do not book, escalate instead, and state the cause in the review note.
No purchase order found, even with fallback matching: do not book, escalate instead.
Duplicate detected: do not book, escalate as a duplicate instead.
Foreign currency or an unusual format: escalate for manual review.
Principle: when in doubt, never book automatically, escalate instead.

PLACEHOLDER CONVENTION
All placeholders in square brackets inside the two working prompts are mapped dynamically, as described in the data flow. The tolerance percentage in the matching prompt is not dynamic: you set it yourself before activation, calibrated against real invoices from the past. List all placeholders that require manual configuration separately at the end.

SELF CHECK BEFORE COMPLETION
Check, before you report the workflow as finished:
1. Are all steps connected correctly, in particular the branch between booking and escalation and the audit log in both branches?
2. Are both working prompts inserted in full and unchanged, and is the tolerance percentage mapped?
3. Do the duplicate check and the fallback matching work as described?
4. Are all points listed that require manual configuration: invoice inbox, ERP and accounting system access, cost center and account mapping, tolerance percentage, and escalation channel?

Working prompts

You are an expert in invoice analysis. Analyze the following invoice and extract all relevant data into a structured JSON format. If a detail is not unambiguously stated in the invoice text, mark it as an assumption rather than presenting it as certain.

INVOICE:
[INVOICE_TEXT]

Extract the following information:
- Invoice number
- Supplier/issuer (name and address)
- Invoice date
- Due date
- Net amount
- VAT rate and amount
- Gross amount (total)
- Description of goods or services/line items
- Purchase order number (if available)
- IBAN/bank details

OUTPUT FORMAT (JSON):
{
  "invoice_number": "...",
  "supplier": {"name": "...", "address": "..."},
  "invoice_date": "YYYY-MM-DD",
  "due_date": "YYYY-MM-DD",
  "net_amount": 0.00,
  "vat_rate": 0.00,
  "vat_amount": 0.00,
  "gross_amount": 0.00,
  "line_items": ["Item 1", "Item 2"],
  "purchase_order_number": "...",
  "iban": "..."
}

If information is not available, set the value to null. Use correct number formats with a period as the decimal separator.

You are an expert in invoice auditing. Compare the extracted invoice data with the purchase order data and identify discrepancies. Explicitly mark any assessment for which relevant information is missing as an assumption.

INVOICE DATA:
[INVOICE_DATA_JSON]

PURCHASE ORDER DATA:
[PURCHASE_ORDER_DATA_JSON]

CHECK CRITERIA:
- Supplier matches
- Amount deviates by no more than [TOLERANCE_PERCENT]%
- Description of goods or services matches the purchase order
- Purchase order number is correct (if available)

OUTPUT:
Create a structured analysis in the following JSON format:
{
  "discrepancies_found": true/false,
  "details": [
    {"category": "Amount/Supplier/Service", "description": "...", "severity": "critical/medium/low"}
  ],
  "recommendation": "book_automatically" or "manual_review_required",
  "rationale": "..."
}

Be precise and state concrete figures for discrepancies. Amounts with a deviation below [TOLERANCE_PERCENT]% are considered correct.

JSON export