Performance metrics

Performance metrics are about quality. Where the Overview tells you how much, this page tells you how well.

In this guide:

• Latency (response time)
• Accuracy (improve-rate)
• Resolution and escalation
• Agent SLAs
• Use the data

Latency: how fast does the bot reply?

Top section:

• p50 latency — median time from user message to bot reply.
• p95 latency — 95th percentile. The “long tail” people most notice.
• By model — broken down by which LLM was used.

Screenshot: Performance charts showing latency distributions per model.

If p95 is above 5 seconds, your bot feels sluggish. Common fixes:

• Switch to a faster model in Playground.
• Reduce context — prune unused knowledge sources.
• Check for slow AI actions inflating round-trip time.

Accuracy: how often is the bot right?

Three signals:

• Improve rate — flags per 100 conversations from Improve responses. Higher = bot wrong more often.
• CSAT correlation — average CSAT for bot-only vs. agent-handled conversations. Useful for justifying the bot’s value.
• Refusal rate — % of replies that say “I don’t know.” Healthy is 5–15%; over 25% means you need more knowledge.

Resolution and escalation

• Resolution rate — % of conversations closed without escalation.
• Escalation rate — % handed to humans.
• Reopen rate — % of solved conversations the customer reopens within 7 days. High reopens = answers weren’t actually solving.

A great chatbot:

• 60–80% resolution rate (rest escalate).
• Under 10% reopen rate.
• Improve rate under 3 per 100.

These are guidelines, not laws — your customer mix matters.

Agent SLAs (live conversations)

For escalated conversations:

• First response time — bot escalation to first agent message. SLA target configurable per plan.
• Full resolution time — escalation to closed.
• SLA hit rate — % of conversations meeting the SLA. Color-coded green/yellow/red.

Configure SLAs in Settings → SLAs. Track misses to identify staffing gaps.

Drill-downs

Each metric supports drill-down:

• Click a chart segment → filtered conversation list.
• Open a conversation → see what happened, learn from it.

Tips

• Look at p95, not just p50. Half your users have a worse experience than the median.
• Improve rate trends over weeks. Don’t react to a daily spike.
• Cross-reference accuracy with knowledge changes. A jump in improve-rate after a Guidelines edit means your edit hurt; revert.

What’s next