> ## Documentation Index
> Fetch the complete documentation index at: https://docs.itellico.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Gather Insights

> Use LLM-as-a-judge to extract structured insights and quality metrics from conversations

export const Screenshot = ({lightSrc, darkSrc, alt, caption, maxWidth = "880px"}) => {
  return <div style={{
    margin: "1rem auto",
    maxWidth,
    width: "100%"
  }}>
      <Frame>
        <img className="block dark:hidden" src={lightSrc} alt={alt} />
        <img className="hidden dark:block" src={darkSrc} alt={alt} />
      </Frame>
      {caption ? <p style={{
    marginTop: "0.5rem",
    fontSize: "0.875rem",
    color: "inherit",
    opacity: 0.8
  }}>
          {caption}
        </p> : null}
    </div>;
};

## What Gather Insights Does

Gather Insights uses AI after each call to evaluate the transcript, extract structured information, and answer the specific questions you care about. This is your **LLM-as-a-judge** layer for quality scoring, categorization, and field extraction.

**Access:** In your agent editor → **Analytics** tab → **Post-Call Analytics** section

<Screenshot lightSrc="/images/agent-editor__analytics-tab-light.png" darkSrc="/images/agent-editor__analytics-tab-dark.png" alt="Analytics tab showing goals, gathered insights, and analytics settings" caption="Use the Analytics tab to configure goals, gathered insights, analytics language, and the analysis model." />

## Analytics Language

<Badge>Expert Mode</Badge>

Set the language for AI evaluation:

**Access:** In your agent editor → **Analytics** tab → **Settings** section → **Analytics Language**

Select the language the AI uses to evaluate goals and insights when analyzing transcripts and answering insight questions.

You can also set the **Analysis Model** in the same **Settings** section.

<Tip>
  This setting applies to both goals and insights.
</Tip>

***

## Default Insights vs Custom Insights

**Default insights** — Pre-configured by itellicoAI:

* Cannot be edited or archived
* Toggle on or off with the Active switch
* Optimized for common use cases

**Custom insights** — Created by your team:

* Full control over name, description, and type
* Can be edited, archived, and restored
* Tailored to your own workflows

<Info>
  Default insights give you a starting point. Add custom insights for your own reporting, QA, or extraction needs.
</Info>

## How To Decide What To Add

| Use this when you want...                   | Best fit            |
| ------------------------------------------- | ------------------- |
| a ready-made common quality signal          | **Default Insight** |
| a custom question tied to your own workflow | **Insight**         |
| a yes/no business outcome                   | **Goal**            |

***

## Setup

<Steps>
  <Step title="Access Post-Call Analytics">
    In your agent editor → **Analytics** tab → **Post-Call Analytics** section → **Add** → **Add Insight**
  </Step>

  <Step title="Configure the insight">
    **Name:** Internal label

    ```text theme={null}
    Example: Customer Satisfaction Score
    ```

    **Description:** The exact question AI should answer

    ```text theme={null}
    Example: On a scale of 1-5 stars, rate the customer's overall
    satisfaction based on their tone, language, and whether their
    issue was fully resolved during the call.
    ```

    <Tip>
      More specific descriptions produce more reliable AI judgments.
    </Tip>
  </Step>

  <Step title="Choose the answer type">
    **Yes/No** — binary answer

    ```text theme={null}
    Example: Was the customer's issue fully resolved?
    ```

    **Open** — free text answer

    ```text theme={null}
    Example: What was the main topic discussed?
    ```

    **Data Point** — one extracted value

    ```text theme={null}
    Example: What order number did the caller provide?
    ```

    **Single Choice** — one option from your list

    ```text theme={null}
    Example: Which product line was discussed?
    ```

    **Multiple Choice** — one or more options from your list

    ```text theme={null}
    Example: Which objections did the caller mention?
    ```

    **Rating** — 1-5 stars

    ```text theme={null}
    Example: How professional was the agent?
    ```
  </Step>

  <Step title="Save">
    Click **Add Analysis** to save.
  </Step>
</Steps>

***

## Use Insights For Structured Extraction

Use **Data Point** or **Open** insights when you want the AI to pull a specific field or summary out of the call. This is the clearest post-call equivalent of “variable extraction”.

**Good extraction-style prompts:**

* `What is the customer's order number? Return only the order number or "not provided".`
* `What date and time did the customer request for the appointment?`
* `What is the main cancellation reason? Return a short phrase only.`
* `Summarize the exact next steps agreed on in 1-2 sentences.`

<Tip>
  If you want clean operational data, tell the AI exactly how to answer: category only, short phrase only, specific date/time, or `"not provided"` when missing.
</Tip>

## Question Types

<Tabs>
  <Tab title="Yes/No">
    **Binary answer questions** for clear metrics.

    **Best for:**

    * Compliance checks
    * Quality verification
    * Feature usage detection

    **Examples:**

    * `Was payment information collected?`
    * `Did the agent follow the script?`
    * `Was the caller transferred?`
    * `Did the customer agree to the offer?`
  </Tab>

  <Tab title="Open">
    **Text response questions** for detailed insights and field extraction.

    **Best for:**

    * Information extraction
    * Issue categorization
    * Root cause analysis
    * Conversation summaries

    **Examples:**

    * `What was the primary topic discussed?`
    * `What objections did the customer raise?`
    * `Which product was discussed?`
    * `Summarize the customer's technical issue`
  </Tab>

  <Tab title="Data Point">
    **Single-value extraction** for operational fields that should stay short.

    **Best for:**

    * Order numbers
    * Requested appointment dates
    * Product names
    * Short labels or IDs

    **Examples:**

    * `What order number did the caller provide? Return only the order number or N/A.`
    * `What appointment date did the caller request? Return YYYY-MM-DD or N/A.`
    * `Which account ID was mentioned?`
  </Tab>

  <Tab title="Single Choice">
    **One option from a list** when you want consistent categories.

    **Best for:**

    * Primary call reason
    * Lead source
    * Complaint severity
    * Product category

    **Examples:**

    * `Choose the main call reason: Billing, Support, Sales, Cancellation, Other.`
    * `Choose the highest escalation level: None, Team Lead, Manager, Legal.`
  </Tab>

  <Tab title="Multiple Choice">
    **One or more options from a list** when several categories can apply.

    **Best for:**

    * Objections mentioned
    * Products discussed
    * Compliance checks passed
    * Follow-up channels requested

    **Examples:**

    * `Which objections did the caller mention: Price, Timing, Trust, Missing Feature, Other?`
    * `Which follow-up channels were requested: Email, Phone, SMS, In-person?`
  </Tab>

  <Tab title="Rating">
    **1-5 star scale** for qualitative assessment.

    **Standard scale:**

    * **5** = Exceptional performance
    * **4** = Above average performance
    * **3** = Average or satisfactory performance
    * **2** = Below average performance
    * **1** = Minimal or poor performance

    **Best for:**

    * Quality scoring
    * Sentiment analysis
    * Performance evaluation
    * Satisfaction measurement

    **Examples:**

    * `Rate customer sentiment (1=very negative, 5=very positive)`
    * `How well did the agent handle objections?`
    * `Rate call quality (1=poor, 5=excellent)`
    * `Customer satisfaction score`
  </Tab>
</Tabs>

## Recommended Starting Set

For most teams, a strong first insight set is:

1. **Customer satisfaction** as a rating
2. **Issue resolved** as yes or no
3. **Primary call topic** as single choice or open text
4. **Customer sentiment** as a rating

Start small. Once those answers are operationally useful, add more.

***

## Examples

<AccordionGroup>
  <Accordion title="Customer Satisfaction" icon="face-smile">
    **Name:** Customer Satisfaction Score

    **Type:** Rating (1-5 stars)

    **Description:**

    ```text theme={null}
    Rate the customer's overall satisfaction based on their tone,
    language, and responses (1-5 stars):

    5 stars = Very satisfied, enthusiastic, all needs met
    4 stars = Satisfied, positive interaction
    3 stars = Neutral, basic needs met
    2 stars = Somewhat dissatisfied, some frustration
    1 star = Very dissatisfied, angry, unresolved issues
    ```
  </Accordion>

  <Accordion title="Issue Resolution" icon="check-circle">
    **Name:** Issue Fully Resolved

    **Type:** Yes/No

    **Description:**

    ```text theme={null}
    Was the customer's issue, question, or request fully resolved
    by the end of the call? Answer Yes only if the customer confirmed
    satisfaction or explicitly agreed the issue was resolved.
    ```
  </Accordion>

  <Accordion title="Structured Extraction" icon="brackets-curly">
    **Name:** Requested Appointment Date

    **Type:** Data Point

    **Description:**

    ```text theme={null}
    Extract the appointment date requested by the caller.
    Return only the date in YYYY-MM-DD format.
    If no date was requested, return "not provided".
    ```
  </Accordion>

  <Accordion title="Call Topic Classification" icon="tag">
    **Name:** Primary Call Topic

    **Type:** Open

    **Description:**

    ```text theme={null}
    Identify the main topic or reason for this call. Choose from:
    - Billing inquiry
    - Technical support
    - Product information
    - Order status
    - Complaint
    - Cancellation request
    - General question

    Return only the category name. If multiple topics were discussed,
    return the one that took the most conversation time.
    ```
  </Accordion>
</AccordionGroup>

***

## Insights vs Goals

Use both together for complete conversation tracking:

| Feature         | Goals                                        | Insights                                        |
| --------------- | -------------------------------------------- | ----------------------------------------------- |
| **Purpose**     | Track conversions                            | Extract insights and judge quality              |
| **Example**     | Appointment booked                           | How satisfied was the customer?                 |
| **Answer Type** | Achieved / Partially achieved / Not achieved | Yes/No, text, data point, choice, or 1-5 rating |
| **Use For**     | Success metrics                              | Quality scoring, categorization, extraction     |
| **Best For**    | Business outcomes                            | Detailed reporting                              |

**Example combination:**

* **Goal:** `Appointment booked` — was the conversion achieved?
* **Insight:** `Customer Satisfaction` — how well was it handled?
* **Insight:** `Primary Call Topic` — what did they want?

***

## Viewing Results

### Individual Conversations

1. Go to **Conversations**
2. Click a conversation
3. Open the **Gathered Insights** section
4. Review each question and AI answer

### Dashboard

Go to **Dashboard** to see insight trends and metrics.

[Learn more about Dashboard →](/manage/dashboard)

## When To Review Insight Results

Review them when you want to:

* compare quality across agents or campaigns
* understand why a goal is being missed
* route quality issues into follow-up work
* export structured call insights for reporting or downstream workflows

***

## Best Practices

<AccordionGroup>
  <Accordion title="Write Specific Questions" icon="pen">
    The AI uses your description to answer. Be detailed.

    ❌ `Was the call good?`

    ✅ `Did the agent provide accurate information, address all customer questions, and maintain a professional tone throughout the call?`
  </Accordion>

  <Accordion title="Define Rating Scales" icon="star">
    For rating questions, always define what each level means.
  </Accordion>

  <Accordion title="Keep Extraction Outputs Narrow" icon="filter">
    If you want reusable data, constrain the output:

    * category only
    * one short sentence
    * one date/time
    * `not provided` if missing
  </Accordion>

  <Accordion title="Start with Key Questions" icon="list">
    Start with 3-5 critical insight questions, not 20.
  </Accordion>
</AccordionGroup>

***

## Next Steps

<CardGroup cols={2}>
  <Card title="Conversation Goals" icon="bullseye" href="/build/analytics/conversation-goals">
    Track conversions and outcomes
  </Card>

  <Card title="Post-Call Automation" icon="envelope-open-text" href="/build/analytics/post-call-automation">
    Trigger emails and tasks from insights
  </Card>

  <Card title="Dashboard" icon="chart-line" href="/manage/dashboard">
    View aggregate insights
  </Card>

  <Card title="Prompt Guide" icon="pen-to-square" href="/build/conversation/prompt">
    Align live behavior with what you measure
  </Card>
</CardGroup>