Overview
OCR converts pictures of text into actual text. Modern OCR goes beyond character shapes: it recovers reading order, table structure, handwriting and layout, which is what makes downstream document automation possible.
How it works
A detection stage finds text regions, a recognition stage reads each region, and a layout stage reassembles reading order and structure. Vision-language models increasingly do all three at once and can answer questions about the page directly.
Use cases
Invoice data capture
Pull totals, dates and line items from supplier invoices into accounting software.
FinanceArchive digitisation
Make scanned historical records full-text searchable.
Public sectorID and form processing
Read identity documents and application forms during onboarding.
BankingAccessibility
Convert scanned course material into text a screen reader can announce.
EducationBenefits
- Unlocks decades of paper and scanned archives for search.
- Removes manual data entry from invoice and form workflows.
- Makes scanned material accessible to screen readers.
- Feeds structured data into systems that never accepted PDFs.
Limitations
- Handwriting accuracy varies enormously by writer.
- Low-resolution scans, skew and shadows degrade results sharply.
- Complex multi-column layouts and nested tables still trip parsers.
- Specialist notation (maths, chemistry, music) needs purpose-built engines.
What to look for when choosing a tool
- Layout and table reconstruction, not just raw text
- Handwriting support if you need it
- Per-field confidence scores for human-in-the-loop review
- On-premise or on-device option for sensitive documents
- Throughput and per-page cost at your volume
FAQ
Does OCR work on handwriting?
Modern engines handle neat handwriting reasonably well and struggle with cursive or messy writing. For high-stakes data, route low-confidence fields to a human check.
What image quality does OCR need?
Around 300 DPI, straight, evenly lit and in focus is the practical target. Most accuracy problems trace back to the capture step rather than the engine.
Can OCR preserve tables?
Layout-aware engines can reconstruct table structure. Plain text-only OCR flattens tables into a stream of words, which usually breaks downstream parsing.