How AI draft confidence works

Every AI draft carries a confidence score (e.g. "AI 72%"). It's the model's honest self-assessment of how sure it is the reply is correct and complete — not a guess at how nice it sounds.

What raises confidence

• A grounded answer. When the draft is based on a published Help Center article (or other workspace context), confidence is higher because the answer is anchored to your real documentation.
• A clear, single-topic question that the knowledge base covers directly.

What lowers confidence

• No matching knowledge. If your workspace has no article covering the question, the model answers from general knowledge and deliberately keeps confidence at or below ~50% — it has no policy reference to confirm the answer fits your product.
• Ambiguity, missing attachments, or multiple questions in one message.

The takeaway

If common questions are scoring low, the fix is usually to add a Help Center article on that topic. Once the AI can cite your own documentation, the same question scores noticeably higher.