---
title: Configure Insights AI
excerpt: Enable Insights AI and configure call subjects and audit grids for automatic AI-powered call evaluation.
order: 2
---

## Activation

Insights AI is enabled per CRM integration from the **CRM Connector** > **Settings** menu:

1. Enable **Call Data Enrichment** at the general level and check your integration
2. Enable **Transcription** and **AI Analysis** under your integration's feature toggles

Call data enrichment must be active before AI analysis can be used.

> AI services incur additional charges. See our [Pricing](/en/pricing) page for more information.

## AI Analysis Outputs

Each processed call produces the following results, available in the call log and as mappable CRM fields:

| Output | CRM field(s) | Description |
|---|---|---|
| Transcription | `transcription` | Full raw text of the conversation |
| Short summary | `summary` | Brief summary of the call |
| Detailed summary | `summary_detailed` | Longer structured summary |
| Sentiment label | `sentiments_label`, `sentiments_label_raw` | Overall sentiment: Positive, Neutral, or Negative |
| Sentiment score | `sentiments_score_average`, `sentiments_nets_score` | Numeric sentiment scores (average and net) |
| Sentiment evolution | `sentiments_evolution`, `sentiments_evolution_raw` | How sentiment changed over the call: Improving, Stable, or Deteriorating |
| Sentiment timeline | `sentiments_array` | Per-segment sentiment breakdown (array) |
| Reason for call | `reason_for_call` | AI-detected primary reason the customer called |
| Reason for frustration | `reason_for_frustration` | Detected source of frustration, if any |
| Topics | `topics_string`, `topics_array` | Free-form topics identified in the call |
| Subjects | `subjects_string`, `subjects_array` | Structured subjects matched to your subject tree |
| First Call Resolution | `kpi_fcr` | Whether the issue was resolved on first contact (boolean) |
| Audit grid | `audit_grid` | Full audit evaluation result (JSON) |
| Audit score | `audit_score` | Overall quality score from the audit grid |
| Recording URLs | `recording_urls`, `recording_urls_string` | Links to audio recordings |
| Transcript URLs | `transcript_urls`, `transcript_urls_string` | Links to transcript files |

## Speech KPIs

In addition to call-level AI analysis, CX-Engine computes the following speech metrics for each call:

| KPI | Description |
|---|---|
| `agent_questions_ratio` | Ratio of questions asked by the agent relative to total agent utterances |
| `customer_questions_ratio` | Ratio of questions asked by the customer relative to total customer utterances |
| `agent_rephrasings_count` | Number of times the agent rephrased the customer's request |
| `speech_ratio_agent` | Share of total speaking time attributed to the agent (0–1) |
| `speech_ratio_customer` | Share of total speaking time attributed to the customer (0–1) |
| `speech_rate_agent_wpm` | Agent speech rate in words per minute |
| `speech_rate_customer_wpm` | Customer speech rate in words per minute |
| `speech_rate_agent_wpr` | Agent speech rate in words per recording segment |
| `speech_rate_customer_wpr` | Customer speech rate in words per recording segment |
| `speech_longest_monologue_agent_duration` | Duration (seconds) of the agent's longest uninterrupted speech segment |
| `speech_longest_monologue_customer_duration` | Duration (seconds) of the customer's longest uninterrupted speech segment |
| `fcr` | First Call Resolution — whether the issue was resolved without a follow-up call (boolean) |

## Call Subjects

**Call subjects** are a hierarchical taxonomy you define to categorize calls. The AI identifies the topics discussed in each call and maps them to your configured subject tree.

Subjects are organized as a tree:

- Each subject has a **name** and an optional **description**
- Subjects can be nested (parent → child) to any depth
- The full path is displayed as a hierarchical label, e.g. `Support > Technical > Database`

Matched subjects are stored in the `subjects_array` and `subjects_string` CRM fields.

### Creating Subjects

1. Navigate to the **Call Subjects** section in the Insights AI settings
2. Click **New Subject**
3. Enter a name and optional description
4. To create a sub-subject, select an existing subject as the parent

## Audit Grid

An **audit grid** is a scoring framework used to evaluate the quality of each call. The AI evaluates every call against your grid and produces a per-criterion score and an overall weighted quality score.

### Structure

An audit grid contains:

| Element | Description |
|---|---|
| **Label** | Name of the audit grid |
| **Language** | Language used for AI evaluation prompts |
| **Categories** | Groups of related evaluation criteria |

Each **category** contains one or more **criteria**:

| Field | Description |
|---|---|
| **Label** | Name of the criterion |
| **Weighting** | Importance of the criterion (1–5). Higher weight means greater impact on the final score. |
| **Instruction** | Prompt sent to the AI describing what to evaluate for this criterion |
| **Checklist** | List of specific elements the AI looks for when forming its score |
| **Rating scale** | Description of what each score from 1 to 5 means for this criterion |
| **Exclusion conditions** | Conditions under which the criterion is skipped (e.g., if the situation did not arise during the call) |

The overall **audit score** is the weighted average of all criteria scores across all categories. Excluded criteria do not affect the score.

### Creating an Audit Grid

1. Navigate to the **Audit Grids** section in the Insights AI settings
2. Click **New Audit Grid**
3. Enter a name and select the evaluation language
4. Add categories and define your criteria within each category
5. For each criterion, set a weighting, write the evaluation instruction, list the checklist items, describe the rating scale levels, and optionally add exclusion conditions

> A default template grid is available to help you get started. You can duplicate it and adapt it to your needs.

Audit results are stored in the `audit_grid` field (full JSON with per-criterion scores and explanations) and in `audit_score` (overall numeric score) on each call.