Insights metrics reference

This page lists every metric on the Insights pages, grouped by page. For each metric, it gives the definition, how the platform calculates it, and how to read it. For the rules that apply across pages, such as which sessions each page counts, see How Insights metrics work.

Calculated by values

Calculated by	Meaning
LLM	An LLM reads the conversation and scores or labels it, in an analysis pipeline.
LLM + platform	Platform rules decide part of the answer (for example, an escalation event). An LLM judge decides the rest.
Platform	Counted or computed directly from what the platform recorded: sessions, events, timings, and costs.
Platform estimate	A platform formula that approximates something it can’t measure directly. Treat it as a signal, not ground truth.
Platform + your settings	Platform counts combined with values your team enters, such as human cost per conversation or the billing policy.
Customer ratings	What end users answered in a survey, scored with the method you pick.
Telephony	Reported by the voice network or telephony provider for each call.
Your query	Whatever the widget’s query returns. Your team defines it.

At a Glance

Page guide: At a Glance.

KPI row

Metric	Calculated by	Definition	How it’s calculated	How to read
Conversations	Platform	Total conversations with your agents in the selected period.	Finished production sessions whose end time falls in the window. A supervisor’s hand-offs to child agents count as one conversation. Live sessions are not counted.	The volume behind every other number on the page. The “N evaluated” line shows how many have analysis results; a big gap means pipelines are behind or switched off.
Containment Rate	LLM + platform	Share of conversations the AI fully resolved without a human.	Conversations with the outcome “contained resolved” ÷ all conversations with an outcome. Platform rules mark escalated (an escalation event) and abandoned (timeout or the customer left); an LLM judge labels the rest resolved, partial or unresolved.	Higher is better. Healthy at 70% or more, watch from 50%, critical below 50%. Bands can be changed in Insights Config.
Quality Score	LLM	Average answer quality, on a 1 to 5 scale.	Average of each conversation’s latest overall score from the quality evaluation pipeline. The overall score weights helpfulness and accuracy 1.0, professionalism and instruction following 0.8.	Healthy at 3.5 or more, critical below 2.5. Shows a dash when the quality pipeline has not run.
Avg Sentiment	LLM	How customers felt, from -1 (negative) to +1 (positive).	From the Sentiment Analysis pipeline: an LLM scores every message from -1 to +1. Each conversation takes the mean of its messages; the card averages the conversations.	Healthy at 0.50 or more, watch from 0, critical below 0. Agent messages are included, so a polite agent pulls the number up.
Period Savings	Platform + your settings	Estimated money saved in the selected period.	Contained conversations × human cost per conversation, minus the actual LLM cost of those conversations.	Green above $0, red below. Only as accurate as the human cost entered in Cost settings.
Escalation Rate	Platform	Share of conversations handed to a person.	Conversations that raised an escalation event ÷ all conversations with an outcome. Needs the quality pipeline to assign outcomes.	Lower is better. Healthy up to 15%, watch to 25%, critical above 25%.

Overview tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Conversation Volume & Containment Rate	LLM + platform	Daily conversations and daily containment on one chart.	Volume per day by session end date. Containment per day by session start date: resolved ÷ all outcomes that day.	Look for days where volume rises and containment falls together. Today’s partial day plots near zero.
Outcome Distribution	LLM + platform	How evaluated conversations ended: resolved, partial, unresolved, escalated or abandoned.	Count and share of each outcome among conversations with an outcome.	Partial and unresolved are where the agent can improve. Escalated went to a person.
Volume (per agent)	Platform	Conversations the pipelines evaluated for each agent.	The largest single-pipeline count for the agent, so a conversation is not counted twice.	Evaluated volume, not total traffic. A supervisor usually shows the most.
Confidence (per agent)	LLM	How sure the intent classifier was about what customers wanted.	Average intent classification confidence, 0 to 100%.	Green above 85%, amber from 25%, red below 25%. Low confidence points to unclear requests or a weak intent list.
Quality and Sentiment (per agent)	LLM	Each agent’s average quality (1 to 5) and sentiment (-1 to +1).	Averages over the agent’s evaluated conversations.	Compare agents against each other.
Trend (per agent)	LLM	Daily quality for the agent as a sparkline.	One point per day; needs at least two days of data.	Green at 3.5 or more, amber from 2.0, red below.

Trends tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Multi-Metric Trends	LLM + platform	Daily containment %, escalation % and quality score.	The KPI-row calculations, one value per day.	Quality (1 to 5) shares the 0 to 100 axis, so it looks flat. Hover it to read the real value.
Sentiment & Quality Over Time	LLM	Daily average sentiment and quality.	Daily averages of the sentiment and quality pipeline scores.	Both share one axis, so small sentiment changes look smaller than they are.
Daily Outcome Distribution	LLM + platform	Resolved, escalated and other outcomes per day.	Daily outcome counts. “Other” groups partial, unresolved, abandoned and unclassified.	A growing grey “Other” band means more conversations end without a clean resolution.

ROI tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Period ROI	Platform + your settings	Savings as a multiple of the AI spend that produced them.	Period savings ÷ LLM cost of contained conversations × 100. Shows 0 when there is no contained cost.	Green above 100%, amber above 0%. 5,000% means savings were about 50 times that AI spend.
FTE Equivalent	Platform + your settings	How many full-time human agents the contained volume represents.	Contained conversations × average handle time ÷ 9,600 minutes (one working month). “Capacity value” is that FTE × the monthly cost of one person.	Always measured against one month, so a 7-day range understates it and a 90-day range overstates it.
Cost / Conversation	Platform	Average AI cost per conversation.	Total LLM cost in the range ÷ Conversations. Each call’s cost is estimated from its tokens and its model’s price. Uses the estimated AI cost from settings when no real cost is recorded.	Green while below the human cost per conversation. The cost total is not limited to production traffic; the conversation count is.
AI Cost vs Estimated Human Cost	Platform + your settings	Daily AI cost next to what people would have cost for the same resolved conversations.	AI cost per day = cost per conversation × that day’s conversations. Human cost per day = resolved conversations × human cost per conversation.	The gap between the bars is the daily saving.

Conversations tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Quality, Helpfulness, Accuracy (per conversation)	LLM	The judge’s scores for a single conversation, 1 to 5.	Taken from the conversation’s latest quality evaluation.	Quality green at 3.5 or more, red below 2.5.
Status (OK / Flagged)	LLM	Whether the conversation was flagged for low quality.	Flagged when the overall quality score is below the flag threshold, 2.5 by default.	Open flagged conversations first; they show what the agent got wrong.

Quality Monitor

Page guide: Quality Monitor.

Health and KPIs

Metric	Calculated by	Definition	How it’s calculated	How to read
Quality Health	Platform	One-line summary: conversations evaluated, weighted overall score and issue flags.	Evaluated is the largest count across the five pipelines; flags are added up across dimensions.	The banner turns amber or red as soon as any one dimension does.
Overall Quality	LLM	One score for overall answer quality, 0 to 100%.	Weighted average of five dimension scores (each 0 to 1): quality 30%, hallucination 25%, guardrails 20%, knowledge gaps 15%, context 10%. Weights rebalance if a dimension has no data.	Healthy above 70%, warning above 50%, critical otherwise.
Faithfulness Score	LLM	How well answers avoid unsupported or contradictory statements.	Average faithfulness score from the hallucination pipeline. A conversation is flagged below 0.5.	Higher is better. The “N flagged” line counts conversations to review.
Knowledge Coverage	LLM	How well the knowledge the agent retrieved supports its answers.	Per conversation: half retrieval precision plus half citation rate. Gaps are counted when the judge detects missing knowledge.	Low coverage usually means the knowledge base is missing content, not that the agent is wrong.
Safety Score	LLM	How accurately guardrails behave.	Per conversation: 1 minus the larger of the false-positive and false-negative scores. A violation is either score above 0.5, or a bypass.	Higher is better. Not the same calculation as Agent Performance’s Safety Score.
Context Preservation	LLM	How well context carries across turns and hand-offs.	Average context score. Flagged below 0.6 or when a duplicate response is detected.	Low scores mean customers had to repeat themselves.

Trend and dimension details

Metric	Calculated by	Definition	How it’s calculated	How to read
Quality Trend	LLM	Daily score for each of the five dimensions.	Daily averages on a 0 to 1 scale. Quality is divided by 5 to fit.	Look for dimensions that dip together; they often share one cause.
Quality Evaluation sub-scores	LLM	Each rubric dimension the judge scored, such as helpfulness, accuracy, empathy or task completion.	Average of each dimension’s 1 to 5 score, divided by 5.	The lowest bar is the first thing to improve.
Hallucination sub-scores	LLM	Faithfulness score, consistency index and contradictions found.	Averages of the judge’s 0 to 1 scores; contradictions are a count of conversations.	Contradictions are the most serious; read those conversations.
Knowledge Gap sub-scores	LLM	Retrieval precision, citation rate and gaps found.	Averages of the judge’s 0 to 1 scores; gaps are a count of conversations.	Low retrieval precision points at search; low citation rate points at how answers use sources.
Guardrail sub-scores	LLM	False-positive score, false-negative score and bypasses found.	Averages of the judge’s 0 to 1 scores; bypasses are a count of conversations.	Lower is better for both scores. Any bypass deserves a look.
Context sub-scores	LLM	Context score, duplicate responses and hand-offs.	Average context score; duplicates and hand-offs are counts.	Many duplicates mean the agent is looping.

Flagged Conversations

Metric	Calculated by	Definition	How it’s calculated	How to read
Health Score (per conversation)	LLM	How bad a flagged conversation is, 0 to 1.	The lowest normalised score among the dimensions that flagged it.	Critical below 0.5, warning 0.5 to 0.7.
Flag Reasons	LLM	Which dimensions flagged the conversation, plus the judge’s written reasons.	Union of the flagging dimensions and their reasons.	Several reasons on one conversation usually point to one root cause.

Agent Performance

Page guide: Agent Performance.

KPI cards

Metric	Calculated by	Definition	How it’s calculated	How to read
Quality	LLM	Average quality score across agents, 1 to 5.	Average overall quality score.	Critical at or below the flag threshold (2.5 by default), warning up to 20% above it.
Hallucination Rate	LLM	Share of evaluated conversations flagged for hallucination.	Flagged ÷ evaluated × 100.	Lower is better. Warning from 5%, critical from 10%.
Knowledge Gaps	LLM	Conversations where the agent lacked the knowledge to answer.	Count of conversations with a gap, with the rate underneath.	Warning above 3 conversations, critical above 7. Fix with knowledge content, not prompts.
Safety Score	LLM	Guardrail pass rate.	100% minus the share of conversations flagged by guardrails.	Warning below 90%, critical below 75%. Not the same calculation as Quality Monitor’s Safety Score.
Context Score	LLM	How well context is preserved, on a 0 to 5 scale.	The 0 to 1 context score × 5.	Higher is better.
Agent Health	Platform	Agents with data, their conversations, and how many are critical, warning or healthy.	Each agent’s status is its worst metric status. Conversations add up per-agent counts.	A conversation evaluated under two agents counts twice in the total.

Agent table and trend

Metric	Calculated by	Definition	How it’s calculated	How to read
Conversations (per agent)	Platform	Conversations the five quality pipelines evaluated for the agent.	Distinct sessions with any quality-pipeline result, by session start.	Evaluated volume, not all traffic.
Quality, Hallucination, Knowledge Gaps, Safety, Context (per agent)	LLM	The KPI-card metrics for one agent.	Same calculations as the cards, limited to the agent’s conversations.	The red row is the agent to fix first. With Compare on, each cell shows ±change vs the previous period.
Quality Trend	LLM	Daily quality and daily flagged conversations.	Daily average quality (left axis) and flagged count (right axis). Compare adds the previous period’s line.	Flag spikes with a quality dip mark the days to investigate.

Customer Insights

Page guide: Customer Insights.

KPI cards

Metric	Calculated by	Definition	How it’s calculated	How to read
Analyzed Conversations	Platform	Conversations analysed for intent or sentiment.	The larger of the intent and sentiment pipeline counts, not their combined total.	If intent and sentiment counts differ a lot, one pipeline is behind.
Unique Intents	LLM	How many distinct customer intents were found.	Distinct intent labels from the classification pipeline.	A long tail of tiny intents may mean the intent list needs tidying.
Avg Sentiment	LLM	Average sentiment, -1 to +1.	Same calculation as At a Glance.	Healthy at 0.30 or more on this page (At a Glance needs 0.50), so the same value can be green here and amber there.
Frustration Rate	LLM	Share of conversations where frustration showed up.	Frustrated ÷ sentiment-scored conversations. A conversation counts if any message was flagged frustrated or scored -0.3 or lower.	Lower is better. Healthy up to 10%, critical above 25%. Agent messages are included.
Resolution Rate	LLM	Share of conversations where the customer’s main intent was fully resolved.	An LLM labels the main intent resolved, partial or unresolved. Resolved ÷ all labeled.	Healthy at 70% or more. Not the same as Containment Rate: a resolved conversation may still have involved a person.

Intents, trajectory and trends

Metric	Calculated by	Definition	How it’s calculated	How to read
Intent Distribution	LLM	What customers asked about.	Conversations per intent; top 7 shown, the rest grouped as Other.	A large Other or Unknown share means the classifier needs better intent definitions.
Sentiment Trajectory	LLM	Whether customers felt better or worse by the end of the conversation.	Last scored message minus first: +0.2 or more is improving, -0.2 or less is declining.	Declining conversations are the ones to read.
Trends Over Time	LLM	Daily classification confidence, resolution rate, sentiment index and analysed volume.	Sentiment index = (average sentiment + 1) ÷ 2 × 100, so 50 is neutral.	Read percentages on the right axis and counts on the left.

Intent Performance

Metric	Calculated by	Definition	How it’s calculated	How to read
Total Sessions (per intent)	LLM	Conversations classified with the intent.	Count of conversations per intent label.	Shows where volume is.
Confidence (per intent)	LLM	How sure the classifier was for this intent.	Average confidence, 0 to 100%.	Green above 85%, red below 25%.
Resolution and Partial Rate (per intent)	LLM	How often each intent was fully or partly resolved.	Resolved ÷ evaluated, and partial ÷ evaluated.	Resolution green at 70% or more. High partial means the agent acknowledges but cannot finish (callback promised, escalation).

Feedback Insights

Page guide: Feedback Insights.

Scores

Metric	Calculated by	Definition	How it’s calculated	How to read
Survey score	Customer ratings	The headline score for one survey, read with the chosen method.	CSAT: share of 4 or 5 stars (8 to 10 on a 0 to 10 scale). Top-box and Bottom-box: share at the top or bottom of the scale. Mean: average rating. NPS: promoters (9 to 10) minus detractors (0 to 6).	Always one survey at a time. Changing the method changes the figure, never the data.
Margin (±)	Platform	The 95% margin of error around the score.	Standard statistical interval for the method; NPS uses the two-proportion version.	When the margin is bigger than the score, the page says “margin exceeds the figure”: do not act on it.
vs prior period	Platform	Change against the previous period of the same length.	Shown as a real move only when the change is bigger than the combined margin of both periods; otherwise “Within noise”.	Trust a coloured arrow; treat “Within noise” as no change.
Arrived externally	Platform	Share of responses submitted outside the conversation.	Externally submitted responses ÷ all scored responses.	High values mean many answers came through a link rather than the in-chat survey.

Response funnel

Metric	Calculated by	Definition	How it’s calculated	How to read
Presented, Scored, Abandoned	Platform	How many people saw the survey, answered it, or left it.	Abandoned = presented - scored.	Surveys have no answer deadline yet, so some abandoned ones can still turn into answers.
Response rate	Platform	Share of people shown the survey who answered.	Scored ÷ presented.	Low response rates make the score less representative.
Comment rate	Platform	Share of answers that included a written comment.	Commented ÷ scored. “Not applicable” when the survey asks for no comment.	Comments explain the scores; read them for low ratings.

Trend, distribution and segments

Metric	Calculated by	Definition	How it’s calculated	How to read
Trend	Customer ratings	The score over time with response counts.	Hourly up to 2 days, daily up to 31 days, weekly up to about 6 months, then monthly.	Short bars mean the point rests on very few answers.
Distribution	Customer ratings	How answers spread across the rating scale.	Count of responses at each rating.	Red is low, amber neutral, green high. A split between both ends means mixed experiences.
Segment score	Customer ratings	The score split by agent, language or version.	Each segment’s score and margin, ordered by significance so tiny segments cannot top the list.	Act only on segments marked as a real move.

Voice Analytics

Page guide: Voice Analytics.

Dashboard tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Total Calls	Platform	Ended voice calls in the window.	Count of calls that ended. Test calls are included.	”vs prev” is the relative change against the previous period of the same length.
Avg MOS	Telephony	Network audio quality, 1 to 5.	Mean Opinion Score from the inbound network quality reports (RTCP) of each call, averaged over calls where it was measured.	Higher is better. Around 4 and above sounds good. Low values mean choppy or delayed audio.
ASR Quality	Platform estimate	How well speech recognition seemed to work, 0 to 100.	100 × (1 - weighted problem score). Repetitions and corrections 25% each, clarity 20%, hesitation and low confidence 15% each. Speech-to-speech calls excluded.	80 or more is excellent, 60 to 79 good, below 60 needs attention. A heuristic, not word accuracy.
E2E Latency	Platform	How long callers wait for the agent to start replying.	Per turn: from the caller’s final transcript reaching the runtime to the first synthesized audio. Averaged per call, then across calls.	Lower is better; the target is under 2,000 ms. Does not include the provider’s end-of-speech detection time.
Barge-In Rate	Platform	How often callers interrupted the agent.	Per call: interruptions ÷ turns, then averaged.	Lower is better. A rise often means answers are too long.
DTMF Fallback	Platform	How often callers used the keypad instead of speaking.	Per call: keypad turns ÷ turns, then averaged.	Lower is better, unless the flow asks for keypad input on purpose.
TTS Quality	Platform estimate	Estimated quality of the agent’s synthesized voice.	Each turn starts at 4.5 and loses points for slow first audio, slow connection, few audio chunks, errors or barge-in.	Higher is better. The score can’t exceed 4.5.
Silence	Platform	Share of the call where no one was speaking.	Per call: silence ÷ call length. Includes the agent’s response waits and gaps over 500 ms.	Lower is better. High silence with normal latency points to long pauses elsewhere in the flow.
Hourly Breakdown	Platform	Calls, MOS, latency, barge-in and ASR score for each hour.	Hourly averages with “measured / total” coverage under each value.	Check coverage before trusting an hour’s average.

Outbound Calls tab

Metric	Calculated by	Definition	How it’s calculated	How to read
Total dials, Answered, In progress	Telephony	Outbound call attempts, how many connected, and how many are still running.	Answered = connected to a session or completed. In progress = not yet in a final state.	Answer rate = Answered ÷ Total dials. Phase and Outcome filters do not change these cards.
No answer, Busy, Failed, Canceled	Telephony	Calls that did not connect, by reason.	Count of calls with each final outcome reported by the provider.	A rise in Failed usually means a number or provider problem.
Machine	Telephony	Calls answered by a machine or voicemail.	Answering-machine detection classified the answer as a machine.	High values mean calls reach voicemail; consider timing or a voicemail message.

Analytics Explorer

Page guide: Analytics Explorer.

Overview and LLM Performance

Metric	Calculated by	Definition	How it’s calculated	How to read
Sessions	Platform	Sessions in the range.	Logical sessions from the session summary; production only by default.	Matches At a Glance Conversations when the ranges match.
Messages	Platform	Messages exchanged in conversations.	Count of message events.	Messages ÷ Sessions gives conversation length.
LLM Calls	Platform	Successful calls to language models.	Count of successful LLM calls, excluding per-turn rollups.	Includes every agent and internal step that called a model.
Errors	Platform	Errors recorded in the range.	Error counts added up across every event category.	Bigger than the “Error” slice of the donut, which counts only one category.
Tokens	Platform	Input plus output tokens.	Sum over LLM calls. The Overview figure includes failed calls; LLM Performance counts successful calls only.	Small differences between the two tabs are expected.
Cost	Platform	Estimated LLM provider cost.	Sum of estimated cost for calls with a known model.	An estimate from price tables, not an invoice.
Avg Latency and P95 Latency	Platform	How long LLM calls took: the average, and the time 95% of calls beat.	Computed over successful LLM calls in the range.	P95 shows the slow tail that the average hides.
Cost by Model and Model Performance	Platform	Calls, tokens and cost per model and provider.	Sums per model.	Shows which model drives spend.
Event Volume and Events by Category	Platform	All platform events per day, and their split by type.	Counts of platform events.	Spikes usually mean bursts of traffic or test runs.

Sessions, Traces, Generations, Query

Metric	Calculated by	Definition	How it’s calculated	How to read
Total Sessions and Error Rate	Platform	Sessions matching the filters, and the share with at least one error.	Error rate = sessions with an error ÷ filtered sessions, over all sessions.	The most reliable two cards on this tab.
Active and Avg Duration	Platform	Live sessions and average length.	Active always reads 0, because the explorer lists only finished sessions. Avg Duration is computed from the 50 loaded rows.	Use Sessions under Operate for live sessions. Treat Avg Duration as a sample.
Per-session and per-call details	Platform	Spans, turns, duration, tokens, cost, latency and time to first token.	Recorded for each session (Sessions, Traces) or each LLM call (Generations).	Use these to explain an outlier found on the summary cards.
Query results	Your query	Answers to your own SQL over the analytics tables.	Read-only SQL, limited to this project and the selected range, capped at 1,000 rows.	Save a useful query to a Custom Dashboard with “Add to dashboard”.

Custom Dashboards

Page guide: Custom Dashboards.

Widgets

Metric	Calculated by	Definition	How it’s calculated	How to read
Any widget	Your query	Whatever the widget’s query measures.	A saved SQL or visual query over the analytics tables, filtered by the dashboard’s date range and filters.	Open the widget’s query (the </> icon) before trusting its title; titles are typed by hand.

Pipelines

Page guide: Pipelines.

Pipeline health

Metric	Calculated by	Definition	How it’s calculated	How to read
Health dot	Platform	Whether the pipeline ran successfully in the last 24 hours.	Success rate = completed runs ÷ all runs (running ones included).	Green above 95%, amber 50 to 95%, red below 50%, grey means no runs.
Last processed	Platform	When the pipeline last ran.	Time of the most recent run.	”Never processed” on an enabled pipeline means the dashboards it feeds get no new data.
Recent runs summary	Platform	Runs in the window by trigger, outcome and average duration.	Counts of runs started by a session ending, an event or by hand; completed, failed and running; mean duration.	Failed runs mean gaps in the quality and sentiment dashboards.

Billing & Usage

Page guide: Billing & Usage.

Usage

Metric	Calculated by	Definition	How it’s calculated	How to read
Included Sessions	Platform + your settings	Sessions counted for billing.	All sessions minus test, eval, synthetic, excluded-channel and below-threshold sessions, under the billing policy.	Lower than Insights totals by design.
Billing Units	Platform + your settings	Total billable units.	Per session: ceil(duration ÷ billing interval), on engaged or wall-clock time. LLM, tool and voice-minute add-ons are added on top.	Long sessions cost more units; the policy table at the bottom shows the interval.
LLM Calls and Interactive Turns	Platform	LLM calls and user turns counted in the billing window.	Counts after billing exclusions.	Turns ÷ sessions gives conversation length.
Voice minutes	Platform + your settings	Billable voice minutes.	Minutes from included voice sessions.	Compare with Voice Analytics call counts for the same window.
Excluded sessions	Platform + your settings	Sessions not billed, by reason.	Grouped by rule: excluded channel, below the interaction threshold, test, eval or synthetic.	Explains the gap between Insights totals and Included Sessions.
Channel Breakdown and Usage by billing scope	Platform + your settings	Sessions, units and LLM calls by channel, and units per billing scope.	Sums per channel; a supervisor and its child agents roll up into one scope.	Shows which channel or conversation drives the bill.