How Insights metrics work

Insights metrics come from several sources and follow different counting rules. Read this page before you compare numbers across pages, because two pages can show different totals for the same period without either being wrong.

When metrics update

Metric type	When it updates
Platform counts, such as conversations, latency, errors, and LLM calls	As conversations happen.
LLM-scored metrics, such as quality, sentiment, intent, and outcomes	After a pipeline analyzes the finished conversation.

Metric cards show a dash until a pipeline evaluates enough conversations in the selected period. A dash means the measure has no data, not that it scored zero.

How metrics are calculated

Every metric in the Insights metrics reference carries a Calculated by value. It tells you where the number comes from, how much to trust it, and who can change it.

Calculated by	Meaning	Examples
Platform	Counted or computed directly from what the platform records: sessions, events, timings, and costs.	Conversations, E2E Latency, LLM Calls
LLM	After a conversation ends, an analysis pipeline asks an LLM to score or label it: quality, sentiment, intent, hallucination, knowledge gaps, guardrails, and context.	Quality Score, Avg Sentiment, Resolution Rate
LLM + platform	Platform rules decide the clear-cut cases and an LLM judge decides the rest. For containment, an escalation event or an abandoned session is a rule. Resolved, partial, or unresolved is the judge’s decision.	Containment Rate, Outcome Distribution
Platform estimate	A platform formula that approximates something it can’t measure directly. Treat it as a signal, not ground truth.	ASR Quality, TTS Quality
Platform + your settings	Platform counts combined with values your team enters, such as human cost per conversation, average handle time, or the billing interval.	Period Savings, FTE Equivalent, Billing Units
Customer ratings	What end users answered in a survey, scored with the method you pick.	CSAT, NPS
Telephony	Reported by the voice network or telephony provider for each call.	Avg MOS, outbound call outcomes
Your query	Whatever a custom dashboard widget’s query returns. Your team defines it.	Custom dashboard widgets, Query results

Compare numbers across pages

Not every page counts the same sessions

Page	Sessions counted
At a Glance	Production conversations only.
Analytics Explorer	Production sessions by default.
Quality Monitor, Agent Performance, Customer Insights	All evaluated sessions, including eval and synthetic sessions.
Voice Analytics	Every ended voice call, test calls included.
Billing & Usage	Billable sessions only. Test, eval, synthetic, and excluded-channel sessions are left out.

Expect small differences in totals between pages. Agent Performance also adds up per-agent counts, so a session that the pipelines evaluated under two agents counts twice.

Two different dates are used

Conversation counts use the date a session ended. Pipeline metrics, such as quality, sentiment, outcomes, and intents, use the date a session started. A session that crosses midnight can land on different days in each.

Score scales differ

Score	Scale
Quality (At a Glance, Agent Performance, Conversations tab)	1 to 5. The LLM judge scores 1 to 5.
Quality Monitor dimensions	0 to 1, shown as a percentage. Quality Monitor divides the 1 to 5 quality score by 5.
Context Score (Agent Performance)	0 to 5. The 0 to 1 context score multiplied by 5.
Sentiment	-1 to +1. The Customer Insights sentiment index rescales it to 0 to 100.
Avg MOS	1 to 5.
ASR Quality	0 to 100.

Safety Score means two things

On Quality Monitor, Safety Score is the average guardrail score: 1 minus the larger of the false-positive and false-negative scores. On Agent Performance, it’s the guardrail pass rate: the share of conversations that guardrails didn’t flag. The two values won’t match.

Containment isn’t resolution

Containment Rate on At a Glance is the share of conversations the AI resolved without a person. Resolution Rate on Customer Insights asks whether the customer’s main intent was resolved, whatever happened along the way. A conversation can be resolved and still involve a person, so Resolution Rate is usually higher.

Sentiment includes the agent’s messages

Average sentiment and the frustration flag cover every message in the conversation, not only the customer’s. The colour bands also differ: At a Glance turns green at 0.50, and Customer Insights at 0.30.

Voice vs prev is a relative change

On Voice Analytics, vs prev shows a relative change. A barge-in rate that moves from 1.0% to 0.5% shows as down 50%, not down 0.5 points.

Thresholds and colour bands

The healthy, watch, and critical bands on At a Glance come from Settings → Insights Config, and you can change them per project. The ROI cost inputs, such as human cost per conversation, sit in the same place. Other pages use fixed bands, which each page guide lists.

Pipeline rules that affect dashboards

• Quality, Agent Performance, Customer Insights, and parts of At a Glance depend on the built-in pipelines. Enable them in Insights → Pipelines.
• Pipelines don’t backfill older conversations.
• Draft pipelines don’t run.
• Custom pipelines skip eval and synthetic sessions.
• Never processed on an enabled pipeline means the dashboards it feeds get no new data.