Lilly Tech Systems
Beginner
Designing Virtual Agents
Build AI avatar agents that handle customer inquiries naturally, resolve common issues autonomously, and know when to escalate to human support.
Agent Persona Design
Your virtual agent's persona shapes every customer interaction. A well-designed persona builds trust and sets appropriate expectations.
- Name and appearance: Choose a professional, approachable avatar that reflects your brand identity
- Tone of voice: Define whether the agent is formal, friendly, casual, or technical based on your audience
- Transparency: Always disclose that the customer is interacting with an AI, not a human
- Personality boundaries: Define what the agent should and should not discuss (no personal opinions, no competitor mentions)
Conversation Flow Architecture
Effective virtual agents follow structured conversation patterns while remaining flexible enough to handle unexpected questions:
The Support Conversation Loop
- Greeting: Welcome the customer and identify their intent
- Classification: Categorize the inquiry (billing, technical, general, complaint)
- Information gathering: Ask clarifying questions to understand the specific issue
- Resolution: Provide the answer, perform the action, or guide the customer through steps
- Confirmation: Verify that the issue is resolved to the customer's satisfaction
- Escalation (if needed): Transfer to a human agent with full context preserved
Good to know: The most critical moment in an AI avatar interaction is the escalation handoff. When the avatar transfers to a human agent, all conversation context must be preserved so the customer does not have to repeat themselves. A poor handoff destroys the trust built by the avatar.
Handling Common Scenarios
| Scenario | Agent Behavior | Escalation Trigger |
|---|---|---|
| Order status | Look up order, provide tracking info | Delivery issue or lost package |
| Password reset | Verify identity, send reset link | Account locked or security concern |
| Product question | Search knowledge base, provide answer | Complex technical issue |
| Billing dispute | Explain charges, show invoice details | Refund request over threshold |
| Complaint | Acknowledge, apologize, offer resolution | Angry customer or repeated complaint |
Escalation Design
Design your escalation system with clear triggers and smooth transitions:
- Confidence threshold: Escalate when the AI is less than 70% confident in its response
- Sentiment trigger: Escalate when customer frustration is detected
- Topic boundary: Escalate for sensitive topics (legal, discrimination, safety)
- Repeat failure: Escalate after two failed resolution attempts
- Customer request: Always honor "let me talk to a human" immediately
Pro tip: Make escalation a positive experience, not a failure. The avatar should say something like: "I want to make sure you get the best help possible. Let me connect you with a specialist who can assist with this." Frame the handoff as proactive care, not a limitation.
💡 Try It: Design an Agent Conversation
Write a complete conversation flow for a virtual agent handling a return request. Include the greeting, information gathering, resolution steps, and both a successful resolution path and an escalation path.
Test your flow by role-playing both sides of the conversation. Does it feel natural? Are there gaps?