Skip to main content
Give your AI a clear task before asking it to introduce your service or understand customer needs. This video uses a merchant outreach assistant to demonstrate agent creation, voice selection, product knowledge, and conversation testing. The video is about 3 minutes 50 seconds, with Mandarin narration and Traditional Chinese captions. Follow the English instructions below for the same workflow.

Video chapters

1. Create an outbound agent

Open AI agents, click the plus button in the upper-right corner, choose a blank agent, and set its type to outbound. Give it a name that describes its purpose, such as Merchant outreach assistant. Decide what the call should accomplish first. For example, learn whether the merchant needs help answering calls, then offer a demonstration if they are interested. A clear goal helps you write the opening and follow-up questions.

2. Choose a voice and write the opening

Select Change voice, choose the language and accent, and preview voices that fit your brand. After bringing a voice back to the agent settings, save the agent to apply it. Start with a speaking rate close to 1× and adjust after listening. Your opening should identify the assistant, explain the purpose, and ask whether the customer has time.
If the conversation needs a customer response, keep the option to hang up automatically after the opening turned off.

3. Add and assign knowledge

Open Knowledge base in the sidebar and add product or service information. You can upload a file, provide a URL, or paste text. The video demonstrates text: enter a title and service description, submit, wait for processing, and review the content. Return to the agent’s knowledge section, select the new material, and save to complete the assignment.

4. Write the prompt and use AI editing

Open the system prompt editor and describe the assistant’s role, goal, tone, and answer boundaries. Use this structure as a starting point:
With AI editing, choose whether to organize the existing script or rewrite from the current settings. Add instructions for a particular tone or goal before generating. Review the facts and flow afterward. Undo or generate another version if needed, then save the version you want to use.

5. Configure iPhone call screening

For the screening workflow shown in the video, enable voicemail detection in advanced settings, set the action to hang up directly, and save. Then enter a short identity and reason for calling and save the screening reply. When the phone asks who is calling and why, the AI uses this short reply first. It waits for the person to speak before playing the full opening. This feature requires a supported outbound voice and a text opening. Recorded openings and one-time notifications are not supported; follow the eligibility shown in the settings.

6. Test the conversation in Playground

Save the agent, then select Test AI agent to open Playground. Allow browser microphone access on first use and check your input device. You can select contact data as conversation context. This test takes place in the browser and does not call the selected contact. Start the test call, wait for the AI to answer, and begin speaking. Check the following during the test:
  • The voice, speaking rate, and opening match your expectations.
  • The assistant waits for an answer after each question.
  • Product answers agree with the assigned knowledge.
  • Busy customers, refusals, and unknown questions follow your handling rules.
Save after each change and start a new test. Before making outbound calls, also check the AI agent line and availability requirements.

Next: let your agent retrieve live data

For product stock or price questions, continue with Connect an API tool to let the assistant retrieve results from your system.