Confidence Scoring

Not every answer an AI produces is equally reliable, and the dangerous ones are the confident-sounding guesses. Wonderchat scores its own confidence on every answer before it goes out, so a shaky response never reaches your customer as if it were certain.

The agent evaluates how well your content actually supports each answer. When confidence is high - the answer is clearly grounded in your docs - it responds. When it's low, the answer escalates to a human or falls back, instead of being delivered as fact. You set the threshold, so you decide how cautious the agent is: tighter for high-stakes support, looser where a best-effort answer is fine.

This turns "is the AI accurate?" from a hope into a control. Rather than trusting that every answer is right, you get a system that flags its own uncertainty and routes the doubtful ones to a person - the same judgment a good support rep uses when they say "let me double-check that."

Combined with grounding and citation, it's a third layer of safety: answer only from your content, show the source, and score the confidence before sending. It's part of a SOC 2 Type II certified, GDPR-compliant platform. Ship answers you can stand behind, and escalate the ones you can't.