Customer voice

This document covers the customer experience layer in Analytics and Insights: Customer Insights for intent classification, sentiment trajectory, and frustration detection; Feedback for end-user ratings and verbatim comments; and Feedback Insights for the results of an individual feedback survey.

Customer insights

The Customer Insights page helps you understand what customers are asking about and how they feel about the experience. It combines intent classification, sentiment scoring, frustration detection, and resolution tracking into a single view to help you identify emerging topics, detect dissatisfaction early, and measure whether the agent resolves the intents it encounters.

Navigation: Project → Insights → Customer Insights

Item	Detail
Answers	What are customers asking for, and does the agent resolve it?
Counts	All analyzed sessions, including eval and synthetic sessions.
Pipelines	The Intent Classification and Sentiment Analysis pipelines must be on.
Dates	Pipeline metrics use the date each session started.

Date range selector: Use the toggle to select 7d, 30d, or 90d.

Filters: Select All environments and All channels at the top of the page to narrow the data by deployment environment and channel. Filter selections are reflected in the page URL, so you can bookmark a filtered view or share it with a colleague.

Customer sentiment summary

A banner at the top displays a contextual summary of the analyzed data, for example: “322 analyzed conversations · 144 intent-classified · 322 sentiment-scored · 13 intents detected · 26.7% frustration.” A sentiment indicator (for example, “Mixed Sentiment”) provides a quick qualitative read. The label is Positive, Mixed, or Negative, from the Avg Sentiment bands.

KPI metric cards

Metric	Description
Analyzed Conversations	Conversations that the intent or sentiment pipeline analyzed in the selected period. The card takes the larger of the two pipeline counts, not their total. The sub-label shows the breakdown. If the two counts differ a lot, one pipeline is behind.
Unique Intents	Number of distinct intent labels the classification pipeline found. A long tail of small intents can mean the intent list needs tidying.
Avg Sentiment	Average sentiment from -1 to +1. An LLM scores every message, each conversation takes the mean of its messages, and the card averages the conversations. Agent messages are included. Healthy at 0.30 or more on this page. At a Glance uses 0.50, so the same value can show green here and amber there.
Frustration Rate	Divides frustrated conversations by sentiment-scored conversations. A conversation counts as frustrated when any message was flagged as frustrated or scored -0.3 or lower. Agent messages are included. Healthy up to 10%, critical above 25%. A warning triangle appears when the rate exceeds the threshold.
Resolution Rate	Share of conversations where the customer’s main intent was fully resolved. An LLM labels each conversation’s main intent resolved, partial, or unresolved, and the card divides resolved by all labeled. Healthy at 70% or more. A warning triangle appears when the rate drops below the threshold. The sub-label shows the count of resolved conversations the system evaluated.

Resolution Rate isn’t the same as Containment Rate on At a Glance. A conversation can be resolved and still involve a person.

Intent distribution

A horizontal bar chart ranks detected intents by volume. Each bar shows the intent name, count, and percentage of total conversations. The chart lists the top seven intents and groups the rest into an “Other” category. The footer indicates the total classified intent assignments and how many low-volume intents the chart grouped. A large Other or Unknown share means the classifier needs better intent definitions.

Sentiment trajectory

A side-by-side panel displays conversation counts across three sentiment directions:

Direction	Description
Improving	The last scored message is at least 0.2 higher than the first.
Stable	Sentiment moved less than 0.2 either way.
Declining	The last scored message is at least 0.2 lower than the first. Read these conversations first.

The footer indicates the total conversations with sentiment data contributing to the trajectory analysis.

Trends over time

A section below the distribution charts plots five daily series. Read counts on the left axis and percentages on the right axis.

Series	Description
Confidence	Daily average classification confidence, right axis.
Intent Classified	Conversations classified each day, left axis. When it matches Sentiment Scored, the two lines overlap.
Resolution Rate	Daily resolution rate, right axis.
Sentiment Index	Daily sentiment on a 0 to 100 scale: (average sentiment + 1) ÷ 2 × 100. A value of 50 is neutral.
Sentiment Scored	Conversations scored for sentiment each day, left axis.

Intent performance

The Intent Performance table shows how well the agent handles each intent.

Column	Description
Total Sessions	Conversations classified with the intent.
Confidence	Average classification confidence for the intent. Green above 85%, plain from 50%, amber from 25%, red below 25%.
Evaluated	Conversations with this intent that the judge labeled resolved, partial, or unresolved.
Resolution Rate	Resolved divided by evaluated. Green at 70% or more, amber from 50%, red below 50%.
Partial Rate	Partial divided by evaluated. A partial result means the agent acknowledged the request with partial action, such as a promised callback or an escalation, but didn’t fully resolve it.

Low-confidence intents, such as Unknown, are the ones the classifier couldn’t place. Read them first.

---

Feedback insights

The Feedback Insights page reports the results of a single feedback survey. Use it to read the headline score, see how the score moved, find the segments driving it, and open the individual responses behind it.

Feedback Insights reads the responses collected against a feedback template. Responses reach the page from agents that reference the template and from external systems that submit through the public feedback API. Until you create a template, the survey list offers only the built-in message_feedback capture, and the page prompts you to set up a survey before it reports anything.

Navigation: Project → Insights → Feedback Insights

Item	Detail
Answers	What do customers say about the experience, and is it changing?
Survey	Every figure is for one survey at a time. The score method changes the figure, never the data.
Movement	A change counts only when it is bigger than the combined margin of both periods.
Tabs	Overview, Trend & distribution, Segments, Responses. Each tab has a Primary / Follow-up switch.

Date range selector: Use the toggle to set the reporting window, for example, Last 7 days.

Filters: Select All environments and All channels at the top of the page to narrow the data by deployment environment and channel.

Export: Select Export CSV to download the current results for offline analysis.

Survey and score method

Two controls at the top of the page define every figure below them.

Control	Description
Survey	Selects the survey to report on. The list contains every feedback template in the project, plus the built-in message_feedback capture. Every figure on the page belongs to the survey you select.
Scored as	Selects how the page reads the recorded ratings. Changing the score method changes the figure, never the underlying data.
Primary / Follow-up	Switches between the survey’s main question and its follow-up question. Each question has its own score.

The score methods available depend on the survey’s rating scale.

Method	Calculation	Offered for
CSAT	Share of 4 or 5 stars, or 8 to 10 on a 0 to 10 scale, among all rated responses.	Star and 0 to 10 ratings
Top-box %	Share of responses at the top of the scale.	Star ratings
Bottom-box %	Share of responses at the bottom of the scale.	Star ratings
Mean	Average rating.	Star ratings
NPS	Promoters (9 or 10) minus detractors (0 to 6), as a share of all rated responses, from -100 to +100.	0 to 10 ratings
Positive %, Ratio	Thumbs up and down, read as a positive share or as a ratio.	Thumbs ratings

Overview

The Overview tab reports the headline score for the selected survey and the figures that qualify it.

Figure	Description
Score	The score for the selected survey and method.
Margin	The 95% margin of error around the score. NPS uses the two-proportion version.
Responses	Rated responses in the period. This is the denominator for every score on the page.
Movement	The change against the previous period of the same length. It shows as a real move only when the change is bigger than the combined margin of both periods. Otherwise it reads Within noise.
Arrived externally	The share of responses submitted outside the conversation, through the public feedback API rather than the in-chat survey.

When the margin is bigger than the score itself, the page warns that the margin exceeds the figure. Don’t act on that score.

Below the score, the response funnel shows how many of the people who met the survey went on to answer it.

Stage	Description
Presented	Surveys rendered in a conversation, whether or not anyone answered.
Scored	Responses where the user submitted a rating. This is the denominator for every score on the page.
Commented	Responses that included comment text. Reads as not applicable when the survey doesn’t ask for comments.
Abandoned	Surveys presented and not answered inside the selected window.

The funnel also reports two rates. The response rate is scored divided by presented. It needs a presented count, and reads as not available until the survey records one. The comment rate is commented divided by scored. It reads as not applicable when the survey doesn’t ask for comments. Abandoned equals presented minus scored.

Surveys have no answer window, so a late response can still turn an abandoned survey into a scored answer.

Read the margin before the number

The page styles a change as movement only when the change clears the combined margin of both periods.

An index built from two proportions carries roughly twice the margin of a plain percentage at the same sample size. NPS is such an index: it subtracts the share of detractors from the share of promoters. At the same number of responses, the same visible gap on an NPS figure means less than it would on a plain percentage.

Trend and distribution

The Trend & distribution tab shows how the score moved across the selected period, and how the responses spread across the rating scale. Use the granularity control on the Trend panel to change the interval, for example, from daily to a wider grouping.

By default, the Trend panel groups scores hourly for ranges up to 2 days, daily up to 31 days, weekly up to about 6 months, and monthly beyond that. The line skips periods with no responses. Bars under the line show the response count for each period. Short bars mean the point rests on very few answers.

The Distribution panel counts responses at each rating. Detractor ratings (1 to 2) show in red, the neutral rating (3) in amber, and satisfied ratings (4 to 5) in green. Satisfied ratings count toward CSAT.

Segments

The Segments tab breaks the score down by the dimensions the platform records with each response.

Segment	Compares scores across
Agent	The agents that presented the survey.
Rendered language	The language the survey rendered in.
Version	The project version in use when the response came in.

Select Movers first to order segments by the size of their change. The page orders segments by significance, so a segment with few responses can’t headline the list on a dramatic-looking number alone.

Each segment shows its score, margin, response count, and movement. Small segments have wide margins, so act only on segments marked as a real move.

Responses

The Responses tab lists the individual responses behind the figures on the other tabs. Four controls let you narrow the list.

Filter	Description
All agents	Filter by the agent that presented the survey.
All ratings	Filter by a specific rating value.
All languages	Filter by the language the survey rendered in.
Search comments	Match text within submitted comments.

Each row shows the rating, the channel where the survey was answered (web, api, voice, or web_debug), the language, the agent version, and the comment. Under Follow-up, the comment column shows the follow-up answer.

If no responses match, the tab reports that it found no feedback for the selected filters. Widen the date range or clear a filter.

Follow-up questions and NPS

Select Follow-up to read the follow-up question’s score, margin, and movement. It works the same way as the primary question.

For a 0 to 10 survey, select NPS under Scored as. NPS subtracts the share of detractors (0 to 6) from the share of promoters (9 or 10), so it runs from -100 to +100. Its margin is wider than a plain percentage at the same number of responses, so check the margin before you report a change.