1. Collect a sample that includes problems
Include regular suppliers, different taxes, credit notes, multi-page invoices and poor-quality documents. Add duplicates and forwarded invoices. One clean invoice does not represent the real workload.
2. Define fields and rules
- Supplier, invoice number, date and currency.
- Net amounts, taxes, totals and arithmetic checks.
- Order or delivery references when required.
- Duplicate rules and approval thresholds.
- What happens to missing or inconsistent data.
3. Separate extraction from validation
OCR recognises text; a model can help interpret formats. Neither replaces checks against your rules and data.
A document with an inconsistent total should remain pending with a visible reason. Recognising uncertainty is part of the workflow.
4. Test writing into your software
Before posting to production, test the connection with limited permissions and sample entries. Check failed submissions and retries: an invoice should not be recorded twice.
The original document, extracted fields, changes and final decision should remain linked. Define who can review and approve.
5. Evaluate the pilot
Compare time per document, detected errors and the share of exceptions against your starting point. Metrics should come from your documents and pilot, not a standard percentage.