Conversational Design

A voice bot is only as good as its design. The best TTS voice and the most accurate ASR system are useless if the dialog is poorly designed. Conversational design is the discipline that transforms technology into natural, effective conversations.

Dialog Flows

Conversation Architecture

A well-designed dialog has a clear structure:

Greeting: Who am I? What can I do? (Set expectations)
Intent recognition: What does the user want? (Open or guided question)
Slot filling: What information do I need? (Follow-ups)
Confirmation: Did I understand correctly? (Avoid misunderstandings)
Action: Execute task (API call, transfer, information)
Closing: Anything else? Farewell

Happy Path vs. Edge Cases

Happy path (everything goes perfectly): 20% of design effort Edge cases (everything that can go wrong): 80% of design effort

Typical edge cases:

User says something unexpected
User doesn't respond (silence)
User gives ambiguous answer
User changes topic mid-conversation
User becomes emotional (frustration, anger)
Background noise distorts recognition

Multi-Turn Dialogs

Complex conversations across multiple turns:

Bot: "How can I help you?"
User: "I'd like to rebook my flight."
Bot: "Of course. Can you give me your booking number?"
User: "LH-4523"
Bot: "Thank you. Your flight on March 15 from Munich to Hamburg. 
      What date would you like to rebook to?"
User: "Is there anything on Wednesday?"
Bot: "On Wednesday, March 19, there are flights at 8:30, 12:15 
      and 17:45. Which do you prefer?"
User: "The one at 12"
Bot: "12:15, perfect. The rebooking costs €45. 
      Shall I proceed?"

Critical: Context retention — the bot must remember across all turns what the conversation is about.

Persona

Why a Persona Matters

The persona defines the voice bot's personality, tonality, and behavior:

Name: Give a human or catchy name? (Controversial)
Voice: Male/female/neutral? Age? Accent?
Tonality: Formal, friendly, casual, professional?
Humor: Yes or no? If yes, what kind?
Boundaries: What does the bot never say? What does it never do?

Persona Design by Industry

Industry	Recommendation
Banking/Insurance	Professional, trustworthy, calm
E-Commerce	Friendly, helpful, somewhat casual
Healthcare	Empathetic, clear, calming
Tech Support	Patient, solution-oriented, technically competent
Hospitality	Warm, inviting, enthusiastic

Fallback Strategies

When AI Can't Continue

Every voice bot fails eventually. The design of failure is crucial:

Level 1 — Comprehension problem: "Sorry, I didn't understand that. Could you rephrase?"

Level 2 — Repeated failure: "I'm sorry, I'm having difficulty with this request. Let me connect you with a colleague."

Level 3 — Topic out of scope: "Unfortunately, that's outside my capabilities. I'd be happy to connect you with a specialist."

Anti-patterns:

❌ Infinite loop: Asking the same question over and over
❌ Silence: Simply saying nothing
❌ Blame the user: "Your request wasn't clear enough"
❌ Fake understanding: Pretending to have understood

Testing and Metrics

Testing Methods

Wizard of Oz: Human simulates the bot before technology is built
User testing: Real users talk to the bot, observer analyzes
A/B testing: Test two dialog variants against each other
Stress testing: Overload bot with edge cases and difficult users
Continuous testing: Regular analysis of real conversations

KPIs for Conversational Design

Task completion rate: How often is the task successfully completed?
Average turns: How many conversation steps until resolution?
Fallback rate: How often does the bot say "I don't understand"?
Drop-off rate: At what point do users hang up?
CSAT score: How satisfied are users after the conversation?
Containment rate: How many conversations are resolved without an agent?

Iterative Improvement

Conversational design is never finished:

Analyze conversations: Where do users fail? Where do they drop off?
Add intents: Recognize and incorporate new user intentions
Improve prompts: Test formulations that are better understood
Refine persona: Adjust tonality based on user feedback
Repeat: Endless improvement cycle

Golden rule: The best voice bot is one the user doesn't recognize as a bot — or one where they don't care because the outcome is right. Design for outcomes, not for technology.