Introduction
Manual invoice validation consumes significant time in any finance department: comparing order lines, verifying taxes, detecting duplicates, checking totals. A well-designed AI automation workflow can substantially reduce human effort on these tasks, but it requires attention to specific steps and typical failure points.
The 5 Steps of an Automated Validation Workflow
1. Ingestion and Normalisation
Invoices arrive in heterogeneous formats: PDF, XML (SAF-T), email, supplier platforms. The first step is to capture and normalise:
- OCR for PDFs: Tools like Google Document AI or Azure Form Recognizer extract text from scanned or image-format invoices.
- Structured parsing: Electronic invoices (XML) are parsed directly, but you need to map different supplier schemas.
- Document classification: An AI model (or simple heuristic) identifies whether the file is an invoice, credit note, or receipt.
Common pitfall: Assuming all PDFs have selectable text. Many suppliers send low-quality scans. Test OCR with real samples before implementation.
2. Extraction of Critical Fields
Once you have the text, the system needs to identify:
- Supplier VAT number
- Issue date
- Invoice number
- Line items (description, quantity, unit price)
- Subtotal, VAT, total
Named Entity Recognition (NER) models trained for invoices (or fine-tuned on your layouts) do this. Alternatives include:
- Regex + heuristics for predictable fields (Portuguese VAT numbers have a fixed pattern, dates follow known formats).
- Large Language Models (LLMs) with prompt engineering: you pass the invoice to GPT-4 or Claude and request structured JSON. Works well with varying layouts, but has per-call costs.
Common pitfall: Trusting extraction 100% without cross-validation. A misrecognised VAT number can associate the invoice with the wrong supplier. Implement sanity checks (e.g., does the VAT number exist in your supplier database? Is the date after the last invoice from that supplier?).
3. Business Rule Validation
With fields extracted, you apply accounting logic:
- VAT calculation: Verify that
subtotal × vat_rate = vat_amountwithin acceptable margin (±€0.02 for rounding). - Total coherence:
subtotal + vat = total. - Duplicates: Compare invoice number + VAT number against existing records.
- Matching with purchase orders: If you have a procurement system, cross-reference invoice lines with purchase order (PO) lines. Detect discrepancies in quantity or price.
- Approval thresholds: Invoices above threshold (e.g., €5,000) require additional approval.
Common pitfall: Rules too rigid generate false positives. An invoice may have a line discount that changes the expected calculation. Design the workflow to flag exceptions rather than block them outright.
4. Exception Handling
Not all invoices pass automatically. Those that fail validation go to a human review queue. AI can help here with:
- Correction suggestions: "The recognised VAT number was X, but doesn't exist. Similar supplier: Y (95% confidence)".
- Prioritisation: Invoices with minor discrepancies (€1 VAT difference) go to quick review; invoices without an associated PO go to manager approval.
Common pitfall: Exception queue becomes a bottleneck if automation rate is low (<70%). Measure why failures occur (OCR issue? Poorly calibrated rule?) and adjust the system iteratively.
5. Accounting Integration
Approved invoices are exported to the ERP:
- Format: Usually CSV, XML, or direct API (e.g., integration with Sage, Primavera, SAP).
- Account mapping: AI can suggest the accounting category based on history ("Invoices from supplier X always go to account 622001 - Specialised Services").
- Audit trail: Keep logs of all automated decisions (which field was corrected, which rule applied) for compliance.
Common pitfall: Not testing integration with real data before go-live. A mapping error could post hundreds of invoices to the wrong accounts.
Practical Tools on the Market
OCR and Extraction
- Google Document AI: Pay-per-use API, good for Portuguese layouts.
- Azure Form Recognizer: Trainable for company-specific layouts.
- Open-source: Tesseract OCR + Python scripts (low accuracy, but zero cost).
Validation and Workflow
- Zapier/Make: For simple workflows (email → OCR → Google Sheets).
- n8n (open-source): Workflow automation with API integrations.
- Specialised platforms: Some offer an end-to-end package (OCR + validation + integration), but check whether they support Portuguese requirements (e.g., VAT number validation, Portuguese VAT rates).
LLMs for Structured Extraction
- OpenAI GPT-4 Turbo with Vision: You can send the invoice image directly and request JSON. Cost ~$0.01 per invoice.
- Claude 3.5 Sonnet: Anthropic alternative, strong at following structured instructions.
When to use LLM: Highly varied layouts, low volume (hundreds/month), need for context (e.g., interpreting handwritten notes in margins).
When NOT to use: High volume (thousands/day), critical latency (<1s), cost is a limiting factor. In those cases, local models or heuristics work better.
Implementation Checklist
Before putting the system into production:
- [ ] Have you tested with a sample of 50+ real invoices from your main suppliers?
- [ ] Critical field extraction rate (VAT number, total) >95%?
- [ ] Exception queue has a clear interface for the accountant to correct without leaving the system?
- [ ] ERP integration tested with dummy data and validated by someone who knows the system?
- [ ] Do you have a baseline of current time (how long to manually validate 100 invoices) to measure gains?
- [ ] Is it documented who approves which type of exception (e.g., invoices >€5k go to CFO)?
Next Steps
Start small: choose a supplier with predictable layout (e.g., a telecom supplier that always sends the same format) and automate just that workflow. Measure time saved, adjust rules, and expand gradually.
Complete automation (0% human intervention) is rarely the goal. The value lies in reducing manual review of 100% of invoices to 20-30%, freeing your team to focus on real discrepancies and business decisions.