Confidence Thresholds
Tune the auto-reply and triage thresholds to match your support quality bar.
The confidence threshold controls when TR-AI auto-resolves a ticket vs escalates it to triage. Lowering the threshold increases the automation rate but also increases the risk of incorrect replies reaching customers.
The default threshold
The default auto-send threshold is 0.82 (composite confidence), with a review threshold of 0.60: replies between the two are held as drafts for an agent, replies below the review threshold go to triage. Assisted mode (off by default) holds every reply, including high-confidence ones, until an agent approves it.
Threshold guide
| Threshold | Automation rate | Risk level | Recommended for |
|---|---|---|---|
| 0.95 | Low | Very low | New deployments, sensitive industries |
| 0.90 | Medium | Low | Conservative teams, regulated ticket categories |
| 0.82 | High | Medium | Default — most teams |
| 0.75 | Very high | High | Mature deployments with large, high-quality KBs only |
How to change the threshold
Thresholds are configured per trial account in Dashboard → Settings → Resolution engine. You can set one auto-send threshold for all tickets and override it per ticket category (how-to, troubleshooting, outage, integration, new version / release, billing, account access, sales, feature request, general). The engine classifies every ticket before answering and applies the category's threshold when one is set.
The minimum similarity floor
Separate from the confidence threshold, TR-AI has a hard floor at best_similarity = 0.30. If the vector search returns no KB chunk with similarity ≥ 0.30, the ticket goes directly to triage without calling the LLM — regardless of the confidence threshold setting.
This floor cannot be changed. It prevents the LLM from generating hallucinated replies when there is no relevant KB content.
Monitoring threshold performance
Use the Dashboard → Recent resolutions feed to monitor confidence scores over time. Look for:
- Tickets scoring 0.75–0.81 — these are close to the default threshold and may be auto-resolvable with KB improvements
- Tickets scoring below 0.60 — these represent KB gaps that need new articles
- The ratio of resolved to triage outcomes — your effective automation rate