🎟️Flaxxa Desk
How the AI copilot decides to auto-resolve a ticket
· March 10, 2026· 2 min read· 8 views
Desk's auto-resolve isn't magic — it's a confidence-threshold gate over Flaxxa AI's RAG pipeline. Understanding the inputs lets you tune it.
The confidence signal
Every AI-drafted reply carries a confidence score derived from three sources:
- Retrieval quality — the score of the top-k retrieved chunks from your KB. If the retrieved sources score below 0.7, response confidence drops.
- Token log-probabilities — the model's own confidence in the words it's emitting. Low-confidence tokens lower the response score.
- Optional self-consistency — sample n=3 responses and score by agreement. Off by default (costs 3×); recommended for high-stakes routes (refunds, compliance).
The threshold
You set the threshold per route under Desk → Settings → AI → Auto-resolve.
- Above the threshold — the bot sends the reply, closes the ticket, fires the CSAT survey.
- Below the threshold — the bot drafts the reply, attaches retrieved sources, and routes the ticket to an agent. The agent reviews + sends.
Recommended starting thresholds
- FAQ-shaped tickets (returns, shipping, business hours):
0.75 - Account-affecting (refunds, address changes, subscriptions):
0.92 - Compliance-affecting (KYC, account closure, security):
1.0(always escalate)
Tune downward gradually as you build confidence in the bot's judgment. Most teams start conservative and relax over the first 30 days.
What the audit log captures
Every AI action — drafted reply, sent reply, escalation, source citation — is captured in the audit log with the confidence score, the retrieved chunks, and the threshold at the time of the decision. Auditable + replayable.
Was this article helpful?