I built Annie to conduct AI opportunity assessment conversations for Return My Time. She asks about a business, follows the answers, and gathers the context we need for a report.
In the walkthrough, I open her configuration in Retell AI, show how she connects to our website, and test her using my own business as the example. You can hear the pauses, follow-up questions, and the way she handles an early finish.
If you want to build your first business voice agent, start by deciding what a useful completed conversation should leave you with. For Annie, that means a clear account of the work, the bottleneck, and the information we still need.
Give the agent one job
Annie's job is discovery for an AI opportunity assessment. Her questions help us understand where a business could make more money, save time, or improve quality.
That gives the conversation a purpose. When I tell her I want to bring more brands into our e-commerce business, she asks which recurring part of the work gets in the way. We end up discussing how we find decision makers and manage outreach.
For your first version, write down the job, the information the caller needs to provide, and what should happen when the call finishes. Keep that description beside you while you configure the agent.
Build the instructions and conversation flow
In Retell, the global prompt holds the agent's identity and rules across the conversation. Annie's includes her role, voice, speech patterns, knowledge sources, and boundaries on what she can say.
Use that prompt to answer practical questions: Who is the agent speaking for? What is she trying to learn? What should she do when the caller asks something outside her scope?
Then map the conversation into stages. The configured flow I walk through includes a welcome, recording acknowledgement, broad discovery, targeted exploration, a summary, and a wrap-up. It also has routes for questions, objections, and a caller who wants to leave early.
The stages give you smaller pieces to inspect. You can look at the welcome separately from the discovery questions, or change the wrap-up without rewriting the whole conversation.
Retell's global settings documentation explains where the shared prompt belongs. Use the current documentation alongside the video if the dashboard has changed.
Give her a question bank with a purpose
Our assessment knowledge base contains a master question bank organized around three outcomes:
- Revenue: where could the business become more effective?
- Time: which work takes effort that could be reduced?
- Quality: where could the product, service, or customer experience improve?
I connect that knowledge base to the broad-discovery and targeted-exploration stages. Annie can choose relevant questions and follow up on the answers while working within that scope.
For your own agent, start with the material needed for its specific job. In this build, that means assessment questions. Retell supports knowledge bases at the agent and node level, so you can attach context where the conversation needs it.
Test with a conversation you can judge
For the demonstration, I use Capstone Brands as the example and talk through our outreach process. That lets me judge whether Annie is following what I say.
When I explain that follow-up can span different channels and months, she asks where we record the contacts, outreach attempts, and next follow-up dates. I say Google Sheets. She then asks who updates the sheet and does the work.
Those questions turn a broad complaint about outreach into details someone can work with: the system of record and the person responsible.
Live test: Annie asks where follow-ups are recorded and who owns them. Source: 16:40–17:16.
Listen for that connection between the caller's answer and the next question. Also pay attention to pauses, repeated questions, and places where the agent assumes more than the caller said. Keep a list of changes to make after the test.
Try ending the call early
Later in the test, I tell Annie I have another meeting and ask to wrap up. She summarizes what she heard, separates known information from missing information, and asks me to confirm her understanding.
She includes the outreach bottleneck and the tools discussed, but flags that we still lack a precise time breakdown and measured conversion results. Those unknowns should stay visible when the conversation becomes input for later work.
Early-exit test: Annie summarizes the conversation, identifies unknowns, and asks for confirmation. Source: 21:32–22:36.
Include an early finish in your own test. Check whether the agent respects the request, preserves what it learned, and gives the caller a clear next step.
Connect the call to the work that follows
The website demonstration starts from an already prepared test checkout. I explain that I used Retell's software development kit to embed Annie in returnmytime.com, then start the assessment after microphone access is ready.
The call is one part of the assessment workflow. Its transcript and audio are inputs for the AI opportunity report process I describe after the demonstration. The video shows the conversation and explains that handoff; it does not show a finished report being delivered.
Decide what your agent should hand over, who or what receives it, and how you will check that the next step happened. A completed call gives you something to inspect before relying on the rest of the workflow.
Plan for maintenance
During the recording, I encounter Retell deprecation notices and show how I use Codex, the documentation, and our existing skills to update the integration. I also review which older test agents can be removed and which two should stay.
You can describe the work in natural language, but you still need to decide what the agent is allowed to change. For a first build, keep the job small enough that you can inspect the configuration and recognize whether the test conversation worked.
Adapt the template to your business, run a conversation you understand, and revise the specific places where it gets stuck.
Get the voice agent template
Get the voice agent starter template in our free AI Operator Hub Skool community. Use it as the starting point for your own business voice agent.