Skip to main content

Agent Training

Configure how your AI assistant handles sensitive topics and special situations.

Overview​

The Agent Training page lets you define Sensitive Topics — categories of user messages that require special handling. When the AI detects a sensitive topic, it either refuses to answer or hands off to your support team.

Navigate to Agent Training in the admin sidebar to manage sensitive topics.

Sensitive Topics​

Define topics that require special handling. Each sensitive topic triggers one of two reactions:

ReactionDescription
Refuse to AnswerAI politely declines to discuss the topic
Handoff to SupportAI offers to connect the user with human support

Built-in Topics​

Deway includes two default topics that are always active and cannot be modified:

  1. Malicious Content — Blocks requests for illegal activities, harassment, prompt injection attempts, and other harmful content. Reaction: Refuse to Answer.
  2. Human Handoff Request — Detects when users explicitly ask to talk to a human, representative, or live agent. Reaction: Handoff to Support.

Creating Custom Topics​

  1. Navigate to Agent Training
  2. Click Add Sensitive Topic
  3. Fill in the topic details
  4. Click Add Topic
  5. Click Save Changes at the top of the page

Each custom topic has these fields:

FieldRequiredDescription
TitleYesShort name (e.g., "Pricing & Billing")
DescriptionYesWhat triggers this topic — be specific so the AI can detect it accurately
ReactionYesRefuse to Answer or Handoff to Support
Custom Agent ResponseNoOverride the default message the AI sends when the topic is detected
ExamplesNoUp to 10 training examples to improve detection accuracy

Custom Agent Response​

By default, the AI uses a standard response based on the reaction type. You can override this with a custom message.

Example:

I'm still learning and can't directly handle payment questions. Let me connect you with our billing team who can help.

If left empty, the AI uses its default response for the selected reaction.

Training Examples​

Examples teach the AI what should and shouldn't trigger the topic. Each example is marked as:

CaseDescription
Positive (should trigger)Messages that SHOULD be caught by this topic
Negative (should not trigger)Similar messages that should be handled normally

Example for a "Pricing & Billing" topic:

ExampleCase
"Can I get a discount?"Positive
"I want a refund"Positive
"How much does the Pro plan cost?"Negative

Negative examples are important — they prevent false positives on messages that are similar but shouldn't trigger special handling.

Editing and Deleting Topics​

  • Click the edit icon on any custom topic to modify it
  • Click the delete icon to remove it
  • Built-in topics cannot be edited or deleted

Copy to Another Project​

If your tenant has multiple projects, you can copy all training configuration from the current project to another:

  1. Click Copy to another Project in the header
  2. Select the target project
  3. Confirm the overwrite

This replaces all training configuration in the target project.

Tips​

Test After Changes​

After saving changes:

  1. Open the chat widget on your site
  2. Send messages that should trigger your topics
  3. Verify the AI responds with the correct reaction
  4. Test edge cases with messages that are similar but shouldn't trigger

Use the Chats Page​

Review real conversations in the Chats view to:

  • Spot topics that need special handling
  • Find false positives (topics triggering when they shouldn't)
  • Identify gaps where users need support but aren't being routed

Review Regularly​

  • Review sensitive topics quarterly
  • Add new topics as your product evolves
  • Adjust examples based on real user messages

Next Steps​