Answering with AI2 min read
Understanding confidence scores
What the percentage means and how to prioritize review
What the score measures
The confidence score reflects how well your knowledge base supports the generated answer — how closely the question matched past answers and how relevant the retrieved document passages were. It is not a guarantee of correctness; it's a prioritization signal.
Reading the bands
- 70%+ — strong support. Skim for tone and accuracy.
- 40–69% — partial support. Read carefully; the draft likely mixes solid material with generalization.
- Below 50% — a visible low-confidence warning appears. Limited knowledge-base support for this question — review thoroughly or rewrite.
Using scores well
- Triage: handle high-confidence answers in a fast first pass, then budget real time for the low end.
- A cluster of low scores on one topic is a signal your knowledge base is missing a document — upload the relevant policy and regenerate.
- Never let a deadline talk you into accepting low-confidence answers unread. The score is telling you something.
