A wrong load record is worse than no load record.
LogAI measures extraction confidence at the field level and routes uncertain extractions to a human review queue before they reach the load record. The goal is not to extract everything — it is to never let a wrong value enter the system silently.
How extraction accuracy works
Field-level confidence, review queues, and measurement per document type and source.
Confidence score per field, not per document
Extraction confidence is measured at the field level — not as a single pass/fail score for the whole document. The origin city may extract with high confidence while the declared weight extracts with low confidence on the same BOL. Each field is reported and routed independently.
Low-confidence fields go to a review queue
When a field's confidence score is below the configured threshold, it goes to the extraction review queue — not silently into the load record. A dispatcher sees the raw document, the extracted value, and the confidence score, and corrects the field in one step.
Review queue does not block the load
A low-confidence extraction on a non-critical field does not stop the load from moving forward. The load record is created with the confident fields populated. The uncertain field is flagged for review and the load is held at the confirmation step until a human verifies it.
Accuracy reported per document type
LogAI tracks extraction accuracy separately for rate confirmations, BOLs, carrier invoices, and EDI 204s. A brokerage using primarily emailed PDFs from a specific shipper will see accuracy metrics for that shipper's documents specifically — not an industry average.
Accuracy reported per source
Documents from email, fax-to-email, EDI, and carrier portals have different accuracy characteristics. LogAI reports them separately so operations managers know which ingestion channel is producing the most review-queue items and can prioritize accordingly.
Accuracy improves with corrections
Each dispatcher correction in the review queue is a labeled training signal. Field extractions that consistently receive the same correction from a specific shipper's document format are flagged for model improvement. Accuracy on known document formats improves over time.
How extraction accuracy is calculated
Accuracy is not self-reported by the extraction model. It is measured from dispatcher behavior in the review queue — the fields a dispatcher accepted versus the fields they corrected.
What happens at each confidence level
High confidence
Auto-accepted
Field is written to the load record without routing to the review queue. Dispatcher sees the extracted value and can override.
Medium confidence
Flagged for review
Field is written tentatively and flagged in the review queue. Dispatcher confirms or corrects before the load is confirmed.
Low confidence
Held for manual entry
Field is left blank. The load record is created without the field. Dispatcher sees the raw document and fills in the correct value.