An OCR vendor says their product achieves 99.9% accuracy. Sounds impressive. But what does that actually mean for your 50-page contract? And why does the same engine produce 98% on one document and 84% on another?
This guide explains how OCR accuracy is measured, what affects it, and what you should actually demand when evaluating document processing tools.
How OCR Accuracy Is Measured
Accuracy is typically measured at two levels: character accuracy rate (CAR) and word accuracy rate (WAR). CAR counts correct individual characters; WAR counts correct complete words. A single wrong letter drops a whole word from the WAR count.
A vendor claiming "99.9% accuracy" is almost certainly reporting CAR on a clean test dataset. On a 500-character page, 99.9% CAR means one wrong character — that sounds fine. But in practice, that character might turn "not liable" into "now liable" in a contract clause.
Key Insight
Always ask vendors: Is this character or word accuracy? What document types were in the test set? Were they scanned or digital PDFs?
What Confidence Scores Tell You
Most OCR engines return a confidence score (0–100) alongside each extracted word. A score of 85+ is generally reliable; below 70 means the engine was uncertain. Inkscribe AI surfaces low-confidence words in the editor so you can review them before exporting.
Confidence scores are more useful than a single aggregate accuracy number because they tell you where the problems are, not just how many there are.
The 6 Biggest Factors That Affect OCR Accuracy
- Scan resolution: 300 DPI is the minimum for reliable OCR. Below 200 DPI, accuracy drops sharply. Modern smartphones produce 200–400 DPI scans; dedicated scanners produce 300–600 DPI.
- Document age and condition: Yellowed paper, ink bleeding, fold lines, and water damage all reduce contrast between text and background. Pre-processing (contrast enhancement, de-speckling) helps significantly.
- Font and type size: Serif fonts in 10pt+ are easy. Decorative fonts, very small text (below 8pt), and very large display text can all confuse engines.
- Complex layouts: Multi-column documents, tables with merged cells, text overlaid on images — these require layout detection in addition to character recognition.
- Handwriting: Printed handwriting (clear block capitals) achieves 90–95% accuracy. Cursive handwriting is genuinely difficult; 80–88% is realistic with modern AI models.
- Language and script: Latin-script languages (English, French, Spanish) are the most reliable. Arabic, Chinese, Japanese, and other non-Latin scripts require specialised models.
Engine Differences: Tesseract vs PaddleOCR vs AI Post-Processing
| Engine | Best For | Weakness | Speed |
|---|---|---|---|
| Tesseract 5 | Clean digital PDFs, simple layouts | Poor on scans, no layout detection | Very fast |
| PaddleOCR | Scanned docs, complex layouts, non-Latin scripts | Slower than Tesseract | Moderate |
| Claude AI post-processing | Correcting OCR errors, handwriting, context-aware fixes | Cannot be used alone — needs OCR first | Adds 2–5s |
Inkscribe AI's hybrid pipeline runs Tesseract and PaddleOCR in parallel and uses Claude AI to reconcile differences and apply context — if the characters look like 'l' but the surrounding text makes 'I' (capital i) more plausible, the post-processor corrects it.
How to Benchmark OCR Accuracy for Your Documents
- Select 20 representative documents from your actual workflow — not the easiest ones.
- Manually transcribe a sample of each (or use a known ground truth).
- Run each engine and compute WER (word error rate) = substitutions + insertions + deletions / total words.
- Compare WER across engines on your specific document types.
- The engine with the lowest WER on your documents is the right one, regardless of vendor benchmarks.