Live Human Voice AI Agent Conversation Design

Flow of how I build out and test voice AI agent scripts.

AI Voice Script Build-Out and Testing

Live Human AI provides services to companies that include outbound calling, inbound support, AI text, email, live transfers, and appointment setting.



This project explains the process of building and testing the script for outbound calls.

Details

Role: Conversation designer and tester

Tools: ChatGPT, cell phone, Google Docs, Live Human app

Team: Product manager, developers, and conversation designer

Timeline: Ongoing

Why the need for a conversation designer?

Cold calling is a challenging area for businesses- people often don’t give cold calls the time of day. 


My job is to design the script to sound personable and not robotic, and to keep the call flow going that leads to booking an appointment with a human to continue the sale funnel.

Success Metrics 
 
01  The script flows so the end user isn’t left confused during or after the call.

02  The user books a call with the client.
   
03  Ideally, the user can’t tell they are talking to a robot.

Who are we desiging for? 

Business owners from various fields who want to generate more buisness, but want to save the "man-power" of their employees calling potential customers. 

Process

PM Call with the Client

Our product manager will hold a call with the client to learn about the company and their goals for the outbound call.

Starting the Script

After the PM call with the client, I’ll take the script and upload it to ChatGPT.

I’ll give the Chat the guidelines and parameters for creating the script.

The Chat will then provide the script:

PSA: You Cannot Trust the First Draft ChatGPT Provides

Asking Chat to create the script saves a lot of time, but it’s very important that the script is refined before handing it over to the customer. 



Chat will provide you a script, but doesn’t always follow the directions correctly.

Refining the Script

From here, I refine and update the script so that it flows naturally.

Ready for Deployment

Once the script is at the point that it flows naturally, I’ll copy the script and add it to a Google Doc. From there, I will post the script within the Live Human app and set the script to live.

Testing the Script

At this point, we want to see how the call flows over a phone call since this is an outbound cold call.



I’ll test each branch with its “Yes”, “No”, “Maybe” paths to make sure the script flows like it should, and highlight that part in blue once confirmed correct. If orange, I’ll go back to refine.



I’m checking to see if the bot doesn’t follow the correct paths or if it doesn’t end a response with a question. If a response doesn’t end with a question this can easily kill the conversation.

You might end up with different versions if the client has different target users for their product or service, like nonprofits and law offices, for example.

You can test out different voices that are more fitting for the AI’s personality the client wants to portray during the call.

Handover to the Client

After testing the call, we will hand it over to the client to test and sign off on before releasing it for their users.

Key Points

-Importance to not assume ChatGPT will write the script correctly.


-Importance to test the script once it’s loaded in Live Human’s platform that will be used for cold calling to confirm the bot doesn’t hallucinate and make sure the conversation is continuous.

-Need to test the script to make sure the bot doesn’t sound robotic or is too slow.