Confidence thresholds
Control when Liya shows AI drafts, when it routes to humans, and how to tune accuracy over time.
Every AI draft Liya generates comes with a confidence score — a value from 0 to 100 representing how certain Liya is that its reply is accurate, grounded, and appropriate for the specific conversation. Confidence thresholds let you decide what happens at different confidence levels.
How confidence is calculated
Liya's confidence score is a composite of several signals:
- Retrieval confidence — how closely the retrieved knowledge base articles match the customer's question
- Grounding score — how well the generated reply is supported by the retrieved articles
- Intent clarity — how unambiguous the customer's request is
- Conversation history signal — whether prior messages in the thread provide useful context
- Safety check — whether the reply passes content safety evaluation
The final score is a weighted average. A score of 80+ generally means Liya is highly confident. Scores below 50 indicate a reply should be reviewed carefully or not shown at all.
Threshold modes
| Mode | Threshold range | Behaviour |
|---|---|---|
| Draft always shown | 0–100 | Liya always shows a draft, regardless of confidence. Agents review everything. |
| Draft shown above threshold | e.g. ≥60 | Draft is shown only when confidence meets or exceeds the threshold. Below it, agents see an empty composer with a flag. |
| Auto-send above threshold | e.g. ≥92 | Replies meeting the threshold are sent automatically without agent review. Requires Enterprise plan. |
Configuring thresholds
Go to Settings → AI → Confidence settings. You can set global defaults or per-intent overrides.
Global threshold
The global threshold applies to all conversations not covered by an intent-specific rule. The default is 65 — drafts with confidence ≥ 65 are shown; below 65, the composer is empty with a note explaining the draft was withheld.
Per-intent thresholds
Set stricter thresholds for sensitive intents. For example, you might want billing and refund replies to require ≥ 85 confidence, while shipping enquiries are fine at ≥ 60.
Navigate to Settings → AI → Intent thresholds, select an intent category, and enter a minimum confidence value.
What agents see at low confidence
When a draft is withheld due to low confidence, agents see a flag in the composer area:
- The reason for low confidence (e.g. "No matching knowledge article found" or "Ambiguous customer intent")
- The knowledge base articles that were considered (even if none matched well)
- A prompt to either write a manual reply or select a canned response
- An option to "Request draft anyway" — which generates the best available draft with a warning banner
Low-confidence conversations are also automatically surfaced in the QA Dashboardso managers can review patterns and identify knowledge gaps.
Tuning thresholds over time
Start with the default threshold (65) and adjust based on observed patterns after 1–2 weeks of traffic:
- Many drafts shown but agents frequently editing heavily: Your threshold may be too low. Raise it to 70–75 to filter out weaker drafts.
- Too few drafts shown, agents mostly writing from scratch: Your threshold may be too high for your current knowledge base coverage. Lower it slightly or add more articles to improve coverage.
- High edit rate on a specific intent: Add more articles for that topic or set a higher per-intent threshold.
Confidence score visibility for agents
By default, the confidence score is shown as a percentage badge on every draft. This helps agents decide how much to scrutinise a reply before sending.
You can hide the score from agents (while keeping the threshold behaviour) under Settings → AI → Display → Show confidence score to agents. Some teams find that scores bias agents toward trusting high-confidence replies without reading them carefully.