A confidence score lets a system decide what needs a person and what does not. A document processing tool might read a printed invoice total with high confidence and a handwritten note with low confidence, sending only the second one for review. Without a score, every field would need the same level of manual checking, which defeats the purpose of automating the easy cases.
The limit is that a confidence score reflects how certain the model is, not whether it is actually right. A model can be confidently wrong, which is one reason confidence thresholds are set and tested against real examples rather than trusted at face value from the start.