Extracting data from invoice PDFs: methods, common errors and how to avoid them

The invoice PDF is the most automated document of all — yet also where errors cost the most: a wrong total propagates into accounting, a wrong VAT code into tax filing, a missing line into inventory. Understanding extraction methods and typical failure points avoids swapping manual data entry for «automatic» data entry that still needs fixing.

Four approaches, from slowest to most scalable

• Manual — operator reads and types: flexible, does not scale, human errors
• Template / fixed coordinates — rules per known supplier: fast until the layout changes
• OCR + rules — extracted text and patterns searched in the flow: fine on repeatable layouts
• AI / IDP — interpretation of new layouts, tables and semantic fields: scales on variability

Native PDF vs scan: not the same invoice

A PDF generated by the supplier's ERP often has a text layer or electronic structure: extraction can be nearly instant. A printed and scanned invoice is an image: OCR is required, with all DPI and quality constraints. A serious workflow detects file type and chooses the method — it does not treat everything as a scan.

Frequent errors — and why they happen

• Totals — decimal separator (comma vs dot), discounts at page bottom, VAT rounding not aligned with lines
• VAT numbers and codes — OCR confuses 0/O, 1/l; fields split across two lines
• Line items — tables with multi-line descriptions, rows split across pages, quantity in a narrow column
• Duplicates — same invoice from email and upload, different protocol numbers

Automating extraction without validating totals just moves the error from typing time to accounting posting time.

What to ask before automating

Bring a representative sample: mix of suppliers, at least some scans, «messy» cases. Ask for first-pass field accuracy, not generic «accuracy». Verify cross-checks: document total vs sum of lines, allowed VAT rates, supplier VAT number in master data.

Targeted review, not review everything

The goal is not zero human clicks on every file, but zero repetitive typing: the system extracts, flags anomalies, the operator intervenes only there. A workflow that forces you to recheck every field has little advantage over manual work.

For accounts payable with many suppliers, LOCRAI extracts fields and line items with built-in validation and queues only exceptions — so you measure savings on the time you spend copying today, not on abstract promises.