Someone Will Get Really RICH Doing This (n8n + Retell AI)

Key Concepts

• GPT-4o Real-time Model: OpenAI's advanced language model optimized for near-zero latency voice interactions, enabling highly natural conversations.
• Voice Agent/Voice Bot: An AI-powered system designed to conduct automated conversations via voice, mimicking human interaction.
• NADN: A popular no-code platform for building AI automations and workflows.
• Retail AI: A no-code platform specifically for creating and deploying voice agents.
• Latency: The delay between a speaker's utterance and the listener's response. "Zero latency" implies virtually no noticeable delay.
• Dynamic Variables: Placeholders in a prompt that are populated with specific data (e.g., customer name, last visit date) from an external source like a CRM.
• Web Hook: An automated message sent from one application to another when a specific event occurs, often used for real-time data transfer.
• CRM (Customer Relationship Management): A system used by businesses to manage and analyze customer interactions and data throughout the customer lifecycle.
• Cal.com: A popular online scheduling and calendar booking platform.
• Summarization Chain: A process within an automation workflow that uses an AI model to condense and extract key information from a text, such as a call transcript.

Overview of the Real-time Voice Agent

The video details the creation of a "craziest voice agent" using OpenAI's GPT-4o real-time model, emphasizing its "zero delays" and natural human-like interaction. This agent is designed for customer upsells and re-engagement, specifically demonstrated for a wellness spa. The core technology stack involves NADN for workflow automation, Retail AI for the voice agent platform, and integrations with Google Sheets (as a CRM) and Cal.com for appointment booking. The primary advantage highlighted is the agent's ability to conduct conversations so naturally that it's indistinguishable from a human, despite the higher cost of the GPT-4o real-time model.

Live Demonstration: Radiance Wellness Spa Call

The presenter demonstrates the voice agent by initiating a call to himself, acting as a customer named John Doe.

1. Initiation: The presenter clicks "execute workflow" in NADN, triggering a call.
2. Agent Introduction: The voice agent, named Katie from "Radiance Wellness Spa," greets John Doe, confirms it's a good time, and mentions his last visit in August for a facial.
3. Offer Presentation: Katie offers a complimentary mini facial or face massage as a thank you, inviting John to book a 30-minute session.
4. Availability Check: John asks for availability on "Monday the 13th." The agent instantly checks and lists multiple slots: 9-10 AM, 11:30 AM-noon, 12:30-2 PM, and 3-5 PM.
5. Appointment Booking: John chooses "1 PM." Katie confirms the time and requests his full name ("John Doe") and email ("[email protected]").
6. Confirmation: Katie confirms the booking for Monday at 1 PM, expresses anticipation for his visit, and concludes the call.
7. Post-Call Automation: Immediately after the call, a webhook is triggered, sending the call summary and booking details to a Google Sheet, which acts as the CRM. The summary accurately reflects the conversation and the booked appointment.

Technology Stack and Workflow

The voice agent system is built using a combination of no-code platforms and APIs:

• NADN (No-code Automation Platform): Serves as the orchestrator for the entire workflow.
  • Initiation: The workflow starts by accessing a Google Sheet containing customer data (acting as a CRM).
  • API Request (HTTP Request Node): NADN sends an HTTP request to the Retail AI API to initiate a phone call. This request includes dynamic variables like the customer's name, phone number, email, last visit date, and service details, all pulled from the Google Sheet.
  • Web Hook Listener: After the call concludes, Retail AI sends a webhook back to NADN.
  • Summarization Chain: NADN processes the call transcript received via the webhook, summarizing the conversation.
  • Google Sheet Update: The summarized call details and appointment information are then updated in the Google Sheet, providing a record for the business.
• Retail AI (No-code Voice Agent Platform): Hosts and manages the AI voice agent.
  • Agent Creation: Users create a "single prompt agent" within Retail AI.
  • LLM Selection: The GPT-4o real-time model is selected for its low latency and natural conversation. Other models like GPT-3.5 Mini are available but have higher latency.
  • Voice Selection: Voice can be chosen from OpenAI, Eleven Labs, or Cartesia, depending on the selected LLM.
  • Universal Prompt: A detailed prompt defines the agent's role (Katie, AI assistant for Radiance Wellness Spa), task (calling returning clients for a complimentary offer), tone (warm, calm, conversational), and a step-by-step process for interaction.
  • Dynamic Variables Integration: The prompt utilizes dynamic variables (e.g., {{name}}, {{email}}, {{last_visited_on}}, {{details}}) that are populated by data sent from NADN.
  • Functions: Retail AI integrates with external services via functions:
    • checkAvailability(cal.com): Checks available slots for booking.
    • bookAppointment(cal.com): Books the confirmed appointment.
    • transferCall(): Transfers the call to a human agent if requested.
    • endCall(): Terminates the conversation.
  • Web Hook Configuration: Retail AI is configured to send a webhook to the NADN workflow's specific URL upon call completion, transmitting the call transcript and other relevant data.
• Google Sheets: Used as a simple CRM to store customer data and record call summaries and booked appointments.
• Cal.com: Integrated for managing appointment availability and booking. API keys and event IDs from Cal.com are used within Retail AI functions.
• Twilio: Used to acquire phone numbers for the voice agent to make calls.

Step-by-Step Building Process

1. Obtain Free Resources:
   • Access the AI Workshop Light Community (free) via the link in the description.
   • Navigate to "Classroom" -> "YouTube Resources."
   • Download the "NADN blueprint" for the GPT real-time upsell voice AI agent.
   • Copy the "voice prompt" text.
2. Set up NADN:
   • Sign up for a free NADN account.
   • Import the downloaded NADN blueprint file into your blank NADN canvas (via "three dots" -> "Import from file").
   • The blueprint will load the entire workflow, including nodes for HTTP requests, webhooks, summarization, and Google Sheet updates.
3. Configure Retail AI:
   • Create a free Retail AI account (includes $10 free credit).
   • Create a new "Voice Agent" -> "Single Prompt Agent" -> "Start from blank."
   • Language Model: Select "GPT-4o real-time" for optimal performance (or other models for cost-effectiveness).
   • Voice: Choose a voice provider (OpenAI, Eleven Labs, Cartesia) and a specific voice.
   • Universal Prompt: Paste the copied prompt from the AI Workshop community into the main prompt area.
     • Customize the role, task, offer details, and step-by-step process to fit your specific industry or campaign.
     • Ensure dynamic variables ({{name}}, etc.) are correctly referenced.
   • Functions:
     • Click "Functions" -> "Add."
     • Add checkAvailability and bookAppointment functions, linking them to your Cal.com account using your Cal.com API key and the specific event ID for the service being offered.
     • Add transferCall (specify a transfer number) and endCall functions.
   • Web Hook Settings:
     • Copy the webhook URL from the NADN workflow's webhook node.
     • Paste this URL into Retail AI's "Web hook settings" (ensure it's a POST request).
4. Connect NADN and Retail AI:
   • In NADN's HTTP Request node, configure the Retail AI API endpoint and authentication (API key from Retail AI).
   • Ensure the JSON payload sent from NADN to Retail AI correctly maps customer data from Google Sheets to the dynamic variables used in the Retail AI prompt.
5. Google Sheet Setup:
   • Prepare a Google Sheet with customer data (Name, Phone Number, Email, Last Visited On, Details of last service).
   • The NADN workflow will read from and write to this sheet.

Cost and Performance

• GPT-4o Real-time:
  • Cost: $0.50 per minute.
  • Latency: "Very, very low" (effectively zero), leading to extremely natural, human-like conversations without awkward pauses.
• GPT-3.5 Mini (Alternative):
  • Cost: $0.024 per minute.
  • Latency: Higher, around 1220 to 1550 milliseconds, which can be noticeable.
• The presenter emphasizes that while GPT-4o real-time is expensive, its "zero latency" and naturalness justify the cost for applications where human-like interaction is crucial, as it makes the bot indistinguishable from a person.

Practical Applications and Customization

The demonstrated voice agent, while using a wellness spa as an example, is highly adaptable:

• Industry Versatility: Can be applied to "any industry" for upsells, re-engagement, customer service, or bringing existing customers back for offers.
• Prompt Customization: Users can easily modify the "universal prompt" in Retail AI to change the agent's role, task, tone, and specific offers to suit different business needs.
• CRM Integration: While Google Sheets is used in the demo, the system can integrate with any CRM via NADN's capabilities.

Conclusion

The video presents a powerful and detailed guide to building a real-time AI voice agent using GPT-4o, NADN, and Retail AI. The key takeaway is the transformative potential of near-zero latency AI for customer interaction, enabling businesses to automate sales and re-engagement with unparalleled naturalness. Despite the higher cost of the GPT-4o real-time model, its ability to eliminate awkward pauses and mimic human conversation makes it a valuable tool for enhancing customer experience and operational efficiency across various industries. The provided free resources and step-by-step instructions make this advanced automation accessible for practical implementation.