The Confidence threshold validator sets the minimum confidence the model must have in a field for it to be auto-approved. Anything below the threshold is flagged, so a human looks at the values the model was unsure about and the rest pass through untouched.
This is the main dial between automation rate and accuracy. A high threshold catches more mistakes but sends more documents to review; a low one automates more and lets more mistakes through.
Options
Rather than guessing, run a batch of documents through and look at how often the model is right at each confidence level. Start strict, then lower the threshold on fields that prove reliable.
How it behaves
This validator always reports a warning, so the reviewer can accept a low-confidence value as it stands. The warning/error setting shown in the panel has no effect here — if a field must be corrected before the review can be completed, enforce it with a validator that supports errors, such as RegExp or Cannot be empty. 0 makes every run stop for human review, which is occasionally useful when you want a person to see everything. 
