Docs
AI Agent

Confidence Thresholds

How Ticket0 decides between auto-send, human review, and human-required handling.

Ticket0 computes a confidence score for each AI draft using a rule-based system. Thresholds control which path the conversation takes.

Decision bands

Score is 0–100 (persisted internally as 0.00–1.00; the UI uses the 0–100 form).

  • At or above the auto-send threshold (workspace-configurable, default 85): draft is ready to send (auto_send). Only actually auto-sent when autoReply is enabled; with it off, scores in this band fall into human_review instead.
  • 60 to (threshold − 1): draft requires operator review (human_review). The composer opens with the draft pre-populated for one-click approval.
  • Below 60: draft suppressed and ticket routed to a human with an empty composer (human_required). The 60 floor is hardcoded and can't be lowered.

What affects confidence

Confidence is not one signal. It combines:

  • Knowledge base match quality
  • Classification confidence
  • Customer sentiment
  • Topic risk penalties (for example billing/account)
  • Similar historical resolution patterns
  • Draft quality indicators (for example very short drafts are penalized)

Adjusting threshold

Go to AI Agent → Features → Auto-reply confidence threshold.

Practical starting points:

  • General support (default): 85
  • Low-risk FAQ / order status: 75–80
  • Billing / account-sensitive flows: 90+

Lowering the threshold increases automation volume (more replies sent without review) and the risk of a bad auto-send. Raise the threshold if auto-sent replies are being followed up with corrections, flagged by customers, or misrepresenting policy. If operators are heavily editing drafts before sending (i.e. in human_review), the fix is prompt/KB quality — the threshold doesn't move those drafts out of the composer.

On this page