Help

Conversational Intelligence

Airgentic Help

Conversational Intelligence Configuration

The Conversational Intelligence screen is where you configure the automated topic clustering and category classification that powers the Categories & Themes insights.

This guide covers how to create taxonomies (your own category structures) and configure the topic/theme clustering that automatically discovers what customers are asking about.


Accessing the Screen

Navigate to Conversational Intelligence from the main admin menu. This screen has two main tabs:

  1. Topics & Themes — Configure and run the automated clustering analysis
  2. Taxonomies — Define your own category structures for governed classification

Taxonomies

A taxonomy is a classification system you define to organise customer questions according to your business structure. Unlike themes (which are discovered automatically), taxonomies give you governed, consistent categories that align with your organisation.

Why Use Taxonomies?

  • Business alignment — Categories match your departments, products, or service areas
  • Ownership — Assign owners to categories for reporting and accountability
  • Consistency — The same categories apply across time, making trends comparable
  • Multiple views — Create different taxonomies for different stakeholders (e.g., by department vs. by enquiry type)

Creating a Taxonomy

  1. Click New in the Taxonomies card
  2. Configure the taxonomy settings:
Setting Description
Name A human-readable name shown in reports (e.g., "Schools and Departments")
Category Type What kind of categories this contains — helps organise your taxonomies
Assignment Mode Single assigns each question to exactly one category (use for routing). Multiple allows questions to have several categories (use for tagging)
Hierarchical categories Enable if categories have parent-child relationships (e.g., Department > Team)
Apply to all services When checked, the taxonomy applies to all services in your account
  1. Click Save to create the taxonomy in draft status

Category Types

Choose the type that best describes your categories:

  • Departments — Organisational units (e.g., Economics, Law, Medicine)
  • Topics — Subject areas (e.g., Billing, Returns, Shipping)
  • Products — Product or service lines (e.g., Widget A, Service B)
  • Intents — User goals (e.g., Buy, Ask, Complain)
  • Regions — Geographic areas (e.g., North, South, International)
  • Priorities — Urgency levels (e.g., Urgent, Normal, Low)
  • Custom — Define your own type

Adding Categories

After creating a taxonomy, switch to the Categories tab to add your classification labels.

  1. Click Add Category
  2. Fill in the category details:
Field Description
Label The display name (e.g., "Undergraduate Admissions")
Parent Category Optional — nest under another category to create hierarchy
Description What questions belong here — helps with training
Owner The person or team responsible for this category
Owner Email(s) Email addresses for notifications (comma-separated)
Status Active categories are used; Inactive are hidden but preserved
Sort Order Controls display order (lower numbers appear first)
  1. Click Save Category

Training Sources

Each category can have training sources that help the AI understand what belongs there. Click on a category row to see and manage its training sources.

Training source types:

Type Description
Fetch pages with URL prefix Retrieves pages from your indexed content matching a URL pattern. The system extracts relevant text to build training data.
Text Brief Freeform description of what the category covers. Useful when URL-based training isn't available.
Exemplar Question A sample question that should match this category. Add several variations for best results.
Negative Exemplar A question that should NOT match this category. Helps distinguish similar categories.

The more training data you provide, the more accurate the classification becomes.

Publishing a Taxonomy

Taxonomies start in Draft status, allowing you to refine them before use.

When ready:

  1. Ensure all categories have been added
  2. Review your training sources
  3. Click Publish

The system will:
1. Build training data from your sources (progress shown in the header)
2. Generate semantic embeddings for each category
3. Make the taxonomy available for classification

Once published, certain settings become locked to maintain classification stability. You can still edit category names, owners, and emails.

Creating a New Version

If you need to change locked settings on a published taxonomy:

  1. Click New Version to create an editable draft based on the published taxonomy
  2. Make your changes
  3. Publish the new version

The new version will replace the previous one for future classifications.

Running Classification

Once published, click Classify Conversations to:

  • Classify all unclassified questions against your taxonomy
  • Optionally re-classify all questions (check "Re-classify all questions")

Classification also runs automatically after each nightly clustering analysis.


Quality and Review

The Reports tab shows how well your taxonomy is performing:

Quality Metrics

  • Coverage — Percentage of questions assigned to at least one category
  • Confidence distribution — How confident the AI is in its assignments

Assignment Summary

Shows how many questions have been assigned to each category, broken down by:

  • Suggested — Automatically classified, pending review
  • Below Threshold — Best match, but confidence didn't meet the threshold
  • Confirmed — Human-verified as correct
  • Rejected — Human-corrected to a different category

Category Performance

Track each category over time:

  • Assignments — Total questions assigned
  • Usage — Questions in the most recent analysis
  • Trend — Change compared to previous analysis
  • Anomaly — Statistical spikes or drops

Click Review to see individual questions and provide feedback.

Near Matches

The Near Matches tab shows questions where two or more categories scored nearly identically (within 4%). These are worth reviewing to:

  • Confirm the correct assignment
  • Add as training examples to improve future accuracy
  • Identify categories that may need clearer differentiation

Unassigned

The Unassigned tab shows questions that didn't meet the confidence threshold for any category. These may represent:

  • Gaps in your taxonomy
  • Questions too generic to classify
  • New topics you haven't accounted for

Use this to identify missing categories or add training examples.


Topics and Themes

The Topics and Themes section configures the automated clustering that discovers what customers are asking about.

Enabling Clustering

Use the Enable Topic & Theme Clustering toggle to turn clustering on or off for this service.

  • Enabled — Questions are automatically clustered into topics and themes on a nightly schedule
  • Disabled — Clustering is paused; no new analysis runs will occur

Disabling clustering does not delete existing themes — it simply stops new analysis runs. You can re-enable it at any time.

How Clustering Works

Each month, the system:

  1. Collects all questions from conversations
  2. Uses AI to identify semantically similar questions
  3. Groups them into clusters — collections of very similar questions
  4. Maps clusters to persistent themes for tracking over time

This happens automatically on a nightly schedule, but you can also trigger manual analyses.

Analysis Status

The status panel shows:

  • When the last analysis ran
  • How many questions were analysed
  • The number of clusters and themes identified
  • Whether an analysis is currently running

Running a New Analysis

Click New Analysis to trigger clustering for the selected month. This is useful when:

  • You've just set up the service and want initial themes
  • You've made significant changes to your content
  • You want to refresh the current month's preliminary data

Analysis Month

The dropdown lets you select which month to view or analyse. Each month has its own clustering run, allowing you to track how themes evolve over time.

Resetting Themes

If you need to start fresh, click Reset Themes to:

  • Delete all existing themes and clustering history
  • Clear the theme mappings

After resetting, run a new analysis to create fresh themes based on your current data.

Warning: This action cannot be undone.

Advanced Clustering Settings

For most use cases, the default settings work well. However, advanced users can adjust clustering behaviour:

Setting Description Default
Min Cluster Size Smallest number of similar questions to form a cluster 5
Min Samples How many similar neighbours a question needs 3
Cluster Merge Distance Distance threshold for merging similar clusters (0 = no merging) 0
Cluster Selection EOM finds stable clusters; Leaf keeps smallest clusters EOM
Distance Metric How similarity is measured (Euclidean or Cosine) Euclidean
Dimensionality Reduction (UMAP) Reduces embedding dimensions before clustering Enabled

Warning: Changing these from defaults will likely impact clustering quality unless you have a specific reason.


Nightly Schedule

Both taxonomies and themes update automatically:

Time (UTC) Process
03:30 Clustering — Groups questions into clusters and maps to themes
04:00 Classification — Assigns questions to published taxonomy categories

This means insights are typically updated by early morning each day.


Best Practices

Designing Taxonomies

  1. Start with your org structure — Categories should reflect how you naturally think about your business
  2. Keep it manageable — Start with 10-20 categories; you can always add more
  3. Use hierarchy sparingly — Deep hierarchies can be harder to manage
  4. Write clear descriptions — Help the AI understand what belongs in each category
  5. Add training examples — A few exemplar questions significantly improve accuracy

Improving Classification Accuracy

  1. Review near-matches — These borderline cases reveal where categories overlap
  2. Add negative exemplars — Help distinguish similar categories
  3. Check unassigned questions — May reveal gaps in your taxonomy
  4. Provide feedback — Use thumbs up/down on classified questions to train the system

Monitoring Themes

  1. Watch for emerging themes — New clusters may indicate new issues or opportunities
  2. Track theme trends — Declining themes may indicate successful interventions
  3. Review theme descriptions — AI-generated labels usually capture the intent well
  4. Use themes alongside taxonomies — Themes discover what you didn't anticipate; taxonomies track what you planned for


← Back to Insights overview

You have unsaved changes