When an AI makes a prediction or decision, it can output a confidence score alongside the result — typically a number between 0 and 1. A high score means the model is sure. A low score signals uncertainty.
Systems use confidence scores to decide whether to proceed automatically, ask for clarification, or escalate to a human. An invoice extraction agent that extracts the vendor name with 97% confidence and the invoice total with 63% confidence can auto-populate the high-confidence field and flag the low-confidence field for human review — rather than treating all outputs equally. This is a simple but powerful way to manage risk in production AI systems.
