GuidesUnderstanding AI
OCR and AI: two ways to read scanned documents
Two routes
OCR then a language model
Vision model
The tools in each family
| Family | Examples |
|---|---|
| Open-source OCR | Tesseract |
| Cloud OCR services | Google Document AI, AWS Textract, Azure Document Intelligence |
| Models that read images and PDFs | Claude, and the other major multimodal models |
How to choose
| Situation | Prefer |
|---|---|
| High volume, always the same form | Specialised OCR |
| Varied documents, tables, free layouts | Vision model |
| Badly degraded scan | OCR with pre-processing, then a model |
| Handwriting | Vision model, with review |
What doesn’t change
- A fixed output schema, validated by code.
- Business checks: totals, dates, known numbers.
- A review queue for doubtful cases, and numbers on the error rate.
A full example: extracting data from PDF invoices. Against errors: preventing AI hallucinations.
Frequently asked questions
- Does a vision model replace OCR?
- Often, for moderate volumes and varied layouts. For millions of identical pages, specialised OCR is still cheaper.
- What about handwriting?
- Recent vision models read a lot of handwriting, without being infallible. Plan a review for the important fields.
- What resolution should I send?
- A sharp image where the text is readable by eye. A blurry or skewed photo does more damage than a modest resolution.