The biggest mistake businesses make with voice AI isn't technical—it's linguistic. They paste a website FAQ or a rigid sales script into the AI's prompt box and wonder why it sounds robotic, awkward, or confusing.
Writing for voice is fundamentally different from writing for text. When we read, we can scan back, skip ahead, and digest complex sentences. When we listen, we process information linearly and in real-time. If your AI speaks in paragraphs, your customers will tune out. assemblyai
This guide covers the art of Conversation Design: how to prompt your Telentir agent to sound natural, persuasive, and distinctly human.
The 3 Golden Rules of Voice Copywriting
1. The "One Breath" Rule
If a human can't say the sentence comfortably in one breath, it's too long for your AI. Long blocks of text increase latency (processing time) and cognitive load for the listener.
- Bad (Text-heavy): "Thank you for calling Telentir. We are the leading provider of AI solutions for businesses looking to scale their operations efficiently. Our hours of operation are Monday through Friday from 9 am to 5 pm EST, but you can always reach us via email at [email protected] for urgent matters."
- Good (Voice-first): "Thanks for calling Telentir. We're currently closed, but our hours are 9 to 5, Monday through Friday. Are you calling with an urgent support issue?"
Fix: Break every long sentence into two short ones. Use conjunctions (and, but, so) sparingly.
2. Front-Load the Important Info
In voice, the listener can't "bold" keywords. You have to place the most critical information at the start or end of the sentence, not buried in the middle.
- Weak: "If you want to schedule an appointment with one of our specialists, please go ahead and tell me what time works best for you."
- Strong: "What time works best for your appointment?"
Fix: Identify the goal of the sentence (e.g., getting the time) and cut everything else.
3. End Every Turn with a Question
Silence is awkward on a phone call. If your AI stops talking without asking a question, the caller won't know it's their turn to speak. This causes "dead air" friction.
- Bad: "I can definitely help with that billing issue. I'll just need your account number." (Silence...)
- Good: "I can help with that. What's your account number?"
Fix: Never let the AI make a statement without immediately guiding the user on what to do next.
Designing the "Persona": System Prompts That Work
Your AI needs a character. "Be helpful" is too vague. In Telentir's prompt builder, be specific about how it should speak.
The "Helpful Assistant" Persona (Support)
"You are Alex, a patient and concise support agent. You use short sentences. You apologize only when necessary but focus on solutions. You sound empathetic but professional. You never use jargon."
The "High-Energy SDR" Persona (Sales)
"You are Sarah, a sales development rep. You are energetic, quick, and casual. You use contractions ('I'm' instead of 'I am'). You sound like a busy colleague, not a robot. Your goal is to find the pain point quickly."
Pro Tip: Add instructions for "Filler Words."
"Use natural fillers like 'Got it,' 'Understood,' or 'Let me check' to acknowledge input before processing the next step." This trick masks latency and makes the conversation feel fluid. riseuplabs
Handling "The Interrupt" (Barge-In)
Humans interrupt each other constantly. "Barge-in" capability allows the user to cut off the AI. But your script needs to handle the recovery naturally.
Scripting for recovery: Don't just repeat the previous question robotically. Use a "re-entry" phrase.
- User: (Interrupts) "Wait, no, I actually meant next Tuesday."
- AI (Bad): "What time works best for your appointment?"
- AI (Good): "Oh, next Tuesday—got it. What time on Tuesday works?"
Telentir Feature: Our LLM engine automatically handles context shifting, but you can improve it by adding a system instruction: "If the user interrupts or corrects themselves, acknowledge the change explicitly before moving on."
The "Fallibility" Trick: Making AI Relatable
Paradoxically, making your AI admit it's not perfect makes it more trusted. If the AI doesn't understand something, don't use a generic "I didn't catch that." Be specific.
- Generic: "I'm sorry, I didn't understand."
- Specific: "Sorry, the line broke up a bit. Did you say Tuesday or Thursday?"
This is called Grounding. It shows the user what the AI heard and asks for clarification on the specific failure point.
Formatting Numbers, Dates, and Emails
LLMs are great at text, but Text-to-Speech (TTS) engines can be quirky. Use these formatting tricks in your script logic to ensure clarity:
- Phone Numbers: "555-0199" should be scripted as "5 5 5, 0 1, 9 9" (spaces create small pauses).
- Dates: "Jan 5" should be "January fifth" (write it out phonetically if the TTS struggles).
- Emails: "[email protected]" often works, but if it fails, instruct the agent to say "john at gmail dot com."
Checklist: Is Your Script Ready?
Before you launch your Telentir agent, read your script out loud (or have the AI read it) and check against this list:
- Breath Test: Can you say every sentence without gasping for air?
- Question Test: Does every AI turn end with a clear prompt for the user?
- Jargon Test: Did you remove words like "authenticate," "validate," or "query"? (Use "check," "confirm," "ask").
- Contraction Test: Are you using "I am" or "I'm"? (Always use contractions for natural speech).
- Empathy Test: Does the script say "Thank you" or "Got it" after the user provides data?
Great conversation design is invisible. When you get it right, customers forget they're talking to software. They just get their problem solved. Open the Telentir Workflow Builder and try rewriting your greeting using the "One Breath" rule today. You'll hear the difference immediately.