AI Agent
Guide to setup and train AI to respond and manage conversations on your behalf on WhatsApp
Overview
If you receive repetitive questions from clients and would like to manage day-to-day conversations automatically, Periskope’s AI Responder can handle it for you. Train this AI to answer important questions on your behalf or notify you of important tasks and TODOs by raising Tickets and Tasks intelligently.
AI Auto-Responder Demo
How to Enable Periskope AI Responder
Go to the AI Settings to enable AI Responder for your org
To enable the Periskope AI responder, go to the Settings page > AI Settings > Toggle Enable AI Responder
Enable AI Responder in the chat where you would like AI to respond
To start responding with AI, you need to turn on the AI responder in the sidepanel of the chat where you want to respond with AI
You must have AI Flagging enabled to turn on AI Responder
How to Train Periskope AI
Go to the AI Knowledge Base
To get started with training your Periskope AI, head over to Settings > AI Knowledge Base
FAQs vs Documents
You have two ways to train the AI: by setting up FAQs or uploading a Training Document
- FAQs: Better for handling specific questions and instructions to respond. (e.g. have AI ask follow up questions and debug an issue)
- Documents: Better for giving AI large contexts (e.g. for finding information about company services and answering a relevant question)
Structuring Your Knowledge Base
How to Structure FAQs
When creating FAQs, follow these best practices for optimal AI responses:
-
Include Multiple Question Patterns
- Add various ways users might ask the same question
- Example for a pricing question:
Adding multiple questions to an FAQ for better identification
-
Add Instructions for controlled AI Responses
You can guide the AI’s response behavior using special instruction blocks:
-
Basic Instructions
-
Conditional Instructions
-
Follow-up Instructions
Add follow up instruction for multi-turn conversations if you wish the AI to follow up with different responses or escalate the issue by raising a ticket.
Yes, the AI responder can raise a ticket automatically to escalate an issue to a human! 😄
Adding multiple questions to an FAQ for better identification
How to Structure Documents
When uploading training documents, organize information for better AI comprehension:
-
Topic-Based Paragraphs
- Group related information into single, focused paragraphs
- Keep each paragraph focused on one main topic
- Example:
-
Hierarchical Information
- Start with general information
- Follow with specific details
- End with examples or use cases
How does AI know when to respond
For any flagged message sent by an external member, Periskope AI checks if:
- A relevant answer exists in the knowledge base for the questions asked
- No internal member has responded to the query
How to stop AI from responding
When the AI Responder is ready to respond or listening to the conversation you will see the AI button in the chat box along with a tag that says “AI is thinking…”. This means that the AI is either responding to a message or listening to the conversation and will respond to any future messages.
AI Activation State
If you wish to stop AI from responding to the chat for the current conversation, simply hover on the AI button and click it to take over the conversation. The AI will activate again when a new message gets flagged in the conversation.
The AI is also interjected when an internal member responds to the conversation where the AI is in Thinking state.
Take over AI
AI Control Panel
The AI Control Panel allows you to fine-tune how your AI Agent behaves across all conversations. Access this panel through Settings > AI Settings.
AI Control Panel Settings
🎯 Activation Settings
Keep AI Agent on by default on all chats
- When enabled, all new chats and existing chats will have AI Agent turned on by default
- You can still manually turn off AI Agent for specific chats if needed
- When disabled, you’ll need to manually enable AI Agent for each chat where you want it active
⏲️ Timing Configuration
Delay Time
- Sets how long the AI should wait before taking over a conversation after a message is flagged
- Default: 30 seconds
- Range: 0-6000 seconds
- Use cases:
- Shorter delays (10-30s): For high-priority support channels where quick responses are critical
- Longer delays (60-300s): When you want to give human agents more time to respond first
Snooze Timing
- Determines how long the AI should remain inactive after a human agent has interjected in the conversation
- Default: 30 seconds
- Range: 0-6000 seconds
- The AI will not respond to new messages during this snooze period, allowing humans to fully handle the conversation
- After the snooze period, AI will reactivate when new messages are flagged
🌟 Personalization
AI Nickname
- Customize how the AI identifies itself in conversations
- Default: “Periskope AI”
- This name will appear when the AI introduces itself or signs messages
- Choose a name that aligns with your brand (e.g., “Support Bot”, “Assistant”, your company name + “AI”)
Company Overview
- Provide context about your business that helps the AI understand your company and respond appropriately
- Include information about:
- Your products/services
- Company values and tone
- Key integrations or tools you use
- Typical customer concerns or questions
- This context improves AI response quality and ensures brand-consistent communication
Example Company Overview:
🤖 AI Agent Roles and Behaviour
Define how the AI should behave, its tone, and specific roles it should take when responding to customers.
AI Behaviour Configuration
- Set the AI’s communication style and approach
- Define specific roles for different types of inquiries
- Configure how the AI should handle various customer scenarios
- Examples of roles:
- Customer Support Agent: Helpful, patient, solution-focused
- Sales Representative: Persuasive, informative, goal-oriented
- Technical Support: Detailed, methodical, troubleshooting-focused
Tone and Communication Style
- Formal vs. casual communication
- Empathetic vs. direct responses
- Proactive vs. reactive approach
- Brand voice consistency
👨🏻💼 Human Agent Roles
Define when and how human agents should take over from AI, and their specific responsibilities.
Human Agent Responsibilities
- Specify which types of queries require human intervention
- Specify which team members have which responsibilities
- Define escalation criteria and handoff procedures
- Configure notifications for human agents
Handoff Scenarios
- Complex technical issues requiring expert knowledge
- Sensitive customer complaints or disputes
- Sales inquiries requiring personalized pricing
- Account-specific questions needing access verification
Example Configuration:
🏷️ Ticket Creation Rules
Define the criteria and conditions for when tickets should be automatically created by the AI.
Automatic Ticket Creation
- Define conditions for when AI should create tickets instead of responding
- Define how the Ticket should be raised or what it should say.
Example Rule:
📝 Private Note Creation Rules
Define when the AI should create private notes for internal use rather than responding to customers.
Private Note Scenarios
- Internal observations about customer behavior or needs
- Important context that should be saved for future reference
- Flags for potential upselling or cross-selling opportunities
- Notes about customer preferences or special requirements
Example Configuration:
Maintain Flag Status after AI Response
- When enabled, AI will not unflag messages after responding
- Useful for keeping messages flagged for human review even after AI has responded
- Helps maintain oversight and quality control
Advanced Parameters
✅ Confidence Score
- Sets the minimum confidence level (0-1) for AI responses
- Higher values mean more conservative AI responses with greater certainty
- Lower values allow AI to respond to more queries but with potentially less accuracy
🌡️ Temperature
- Controls randomness in AI responses (0-2)
- Lower values make responses more focused and deterministic
- Higher values make responses more creative and varied
☢️ Top P (Nucleus Sampling)
- Controls diversity via nucleus sampling (0-1)
- Lower values focus on more likely responses
- Higher values allow for more diverse response options
📉 Top K
- Limits the number of highest probability tokens to consider for each step
- Lower values make responses more focused
- Higher values allow more response variety
Start with default settings and adjust based on your AI’s performance. Monitor conversations to see if the AI is too aggressive (reduce confidence score) or too conservative (increase temperature slightly).