What the confidence score measures
Every extraction receives a confidence score between 0% and 100%. This reflects how clearly and unambiguously the extracted values appeared in the source document — not whether the values are factually correct, but how certain the AI is about what it read.
A score of 95% means the AI found clean, well-formatted text. A score of 40% might mean the document was a low-resolution scan, fields were partially obscured, or the document structure was unusual.
Score ranges and what they mean
- 80–100% (green) — High confidence. Values extracted cleanly. Suitable for automated processing without human review.
- 50–79% (amber) — Moderate confidence. Most values are likely correct but spot-checking is recommended for financial amounts or dates.
- 0–49% (red) — Low confidence. Document may be a poor-quality scan, hand-written, or in an unusual format. Human review strongly recommended.
How to improve low confidence scores
- Use higher-resolution scans (300 DPI minimum)
- Ensure good contrast between text and background — avoid shadows and glare
- Create a custom template with detailed field descriptions to guide the AI
- Make corrections to extracted fields — each correction improves future accuracy on similar documents
よくある質問
100%のスコアはデータが正しいことを保証しますか？
高い信頼度はAIが文書中の値を明確に識別したことを意味しますが、事実の正確性を検証するものではありません。
同じ文書でスコアが毎回違うのはなぜですか？
AIモデルにはわずかな非決定的なばらつきがあります。同じ文書の抽出間で±5%の差が生じる場合があります。