How to Understand Confidence Scores
Learn what confidence percentages mean and why some fields are flagged for review.
CargoLint assigns a confidence percentage to each extracted field, indicating how certain the AI is about that value’s accuracy. Understanding these scores helps you prioritize review work and trust the platform’s results.
Before you start
- You have documents that have been processed
- You understand basic document types (invoices, packing lists, etc.)
- You know how to access the review queue
Steps
1. Know the 78% review threshold
CargoLint flags any field with confidence below 78% for review, and routes a document to the Review status when its overall score falls below 78%. This threshold balances accuracy with efficiency - fields above 78% are typically reliable enough to use without close checking, while fields below require human verification.
2. Understand what confidence means
A confidence score reflects how certain the AI is that a value is correct, based on how clearly the value appears in the document and how well it matches expected patterns. A field with 95% confidence was extracted from clear, unambiguous text. A field with 55% confidence comes from faint, obscured, or ambiguous text that could be misread.
3. Know that every document type uses the same threshold
Commercial invoices, packing lists, bills of lading, and certificates of origin all use the same 78% review threshold, and issues of the same severity cost the same on every type. An 85% invoice and an 85% packing list mean the same thing.
4. Identify high confidence fields in the UI
High-confidence fields (85% and above) appear with a green badge. You can usually approve these without detailed review, though spot-checking is always wise.
5. Identify medium confidence fields in the UI
Fields between 78% and 84% show an amber badge. Review these carefully, but they’re usually correct. Compare them against the original document image.
6. Identify low confidence fields in the UI
Fields below 78% display a red badge. CargoLint automatically queues these for review. Always verify low-confidence fields against the original document before approving.
7. Know why a high-scoring document can still be in Review
The confidence score is one input to review routing, so the score’s color and the document’s status can differ. Three conditions hold a document in Review regardless of how high it scores:
- HS codes waiting for confirmation - suggested codes are never applied automatically; a reviewer confirms each one
- A degraded scan - the scan quality was too low for reliable reading
- A shipment finding - a consistency check found conflicting information between this document and another in the same shipment
A green 90% document held for HS confirmation was read cleanly; it is waiting on a human decision the score does not measure. The document view shows the specific hold reason. Conversely, an amber document in Pending Approval cleared the threshold with no holds, so one-click approval is all it needs.
8. Investigate why confidence is low
Low confidence often occurs when:
- Text is handwritten or printed at an unusual angle
- The document is blurry, faded, or has poor contrast
- A field contains special characters or formatting
- Multiple interpretations exist (for example, date formats like “03/14” could be March 14 or the 14th of March depending on regional convention)
Improving the scan quality (see Scanning Tips) often resolves low-confidence extraction.
Tip: Don’t automatically distrust low-confidence fields. The AI is simply indicating uncertainty, not that the value is wrong. Always compare against the original document.
Tip: High confidence doesn’t mean 100% accuracy. Even 95% confident fields can occasionally be incorrect, especially with unusual document layouts. Spot-check a few high-confidence fields each day.
What’s next
- How Confidence Scoring Works - How the score is calculated and how it routes documents
- Document Statuses - What Pending Approval and Review mean and how documents move between them
- Reviewing Documents - Correct fields in the review queue
- Scanning Tips - Improve scan quality to boost confidence scores
- Troubleshooting - Resolve issues with consistently low confidence