# Build an AI Phone Agent for Customer Support (Full Demo) Source: https://www.youtube.com/watch?v=I-qlDMpDKkA Recap page: https://rapidrecap.app/video/I-qlDMpDKkA Generated: 2025-12-01T13:06:08.129+00:00 --- ## Quick Overview Jordan Cunningham from Relevance AI demonstrates the creation and workflow of an AI Phone Agent named Harley, which acts as an outbound support agent by integrating with Pylon, using a multi-agent system approach that includes a Call Screener & Notifier Agent to perform prerequisite research before Harley makes the call, resulting in faster, more informed, and more human-sounding customer interactions. **Key Points:** - Harley is an outbound AI phone agent designed to solve the issue of limited phone support access across different subscription tiers. - The system utilizes a multi-agent workforce architecture, starting with a Call Screener & Notifier Agent that performs prerequisite research (ticket info retrieval, Pylon note posting, knowledge base search) before Harley initiates the call. - Harley uses a performance-optimized model (like GPT-4) to ensure high-quality, responsive, and human-like voice interactions, avoiding choppy back-and-forth waits. - The workflow is triggered via a Pylon webhook command: `/call [phone number] [message]`, which connects Pylon to Relevance AI. - Harley can handle various support capabilities including searching knowledge, web search, reading screenshots, and sending personalized LinkedIn DMs. - The system supports scheduled calls across international time zones and automatically posts a call summary back to the Pylon ticket for documentation. - The presentation concludes by showing the configuration settings in the platform, emphasizing ease of use and the positive customer experience resulting from pre-call preparation. ![Screenshot at 01:15: The title slide introduces the presentation topic: 'Relevance AI \| AI Phone Agents in SaaS Support' and the specific focus: 'Building Harley: Our Outbound Support Phone Agent'.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-01-15.png) **Context:** This presentation features Jordan Cunningham, an engineer from Relevance AI, detailing the development and implementation of an AI-powered outbound phone agent named Harley. The core problem addressed is the traditional SaaS limitation where phone support is often reserved for higher-tier customers. Harley is designed to provide high-quality, immediate voice support across all subscription tiers by using a multi-agent workforce architecture that leverages AI tools and pre-call research. ## Detailed Analysis Jordan Cunningham details the build and workflow of Harley, an outbound AI phone agent designed to equalize access to phone support, which is typically restricted by subscription tiers in SaaS models. Harley functions as a multi-agent workforce, not a standard single agent. The process begins when a user raises a support ticket via the in-app widget containing their issue and contact information. This triggers a webhook in the ticketing system (Pylon), which sends the initial data to a 'Conversation Screener & Notifier Agent'. This screener agent performs prerequisite steps: 1) Get Ticket Info, 2) Post Note to Pylon (as an internal note), and 3) Research Everything (using knowledge base, tools, and Google searches). The screener then passes this 'starter pack of information' to Harley, the main support agent, enabling Harley to sound informed immediately upon calling the customer, avoiding frustrating holds or delays. Harley is configured to use a performance-optimized model (like GPT-4) for superior voice quality and emotional range. The configuration settings, shown in the 'Phone mode settings', allow customization of greeting modes, phone numbers, voice models (like Eleven Multilingual v2), voice speed, background noise reduction, and call timeouts. The workflow ensures that the conversation flows smoothly, with Harley immediately addressing the issue, backed by the pre-call research. If escalation is necessary, the guidelines stipulate how Harley should handle that process, ensuring complex issues are routed to human agents efficiently. Finally, after the call, a summary is automatically posted back to the Pylon ticket, documenting the resolution. The goal is to provide an instant, high-quality resolution that feels personal, achieving support equity across all customer tiers. ### Introduction and Problem - Jordan Cunningham introduces Harley, an outbound phone agent built to solve the SaaS support tier problem where phone calls are often restricted to higher tiers - Harley is described as a multi-agent system integrating tools in a flow. ### Call Screener Workflow Details - The process before Harley calls involves three steps: 1) Get Ticket Info (retrieve info from the ticket), 2) Post Note to Pylon (notify workflow started), and 3) Research Everything (using knowledge base, tools, marketplace, Google searches) - The screener passes this starter pack of information to Harley so the call starts smoothly. ### Phone Mode Configuration - Settings include using the Eleven Multilingual v2 voice model for English, setting voice speed to 1.1, enabling background noise reduction, and setting call timeouts (120s silence, 720s max duration). ### Agent Prompting & Guardrails - Harley's prompt emphasizes core components, authentication, tool usage (e.g., Search Brainfish Agent), staying focused on the call purpose, using respectful language, and avoiding sharing sensitive PII. ### Workflow and Result - The system is triggered by a Pylon webhook command `/call [phone number] [message]`, which initiates the conversation screener, followed by Harley making the call, resulting in a smooth, informed interaction where summary documentation is automatically posted back to the Pylon ticket. ![Screenshot at 00:04: Alex White introduces Jordan Cunningham, who is one of her team members, to present the session on AI phone agents.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-00-04.png) ![Screenshot at 01:15: The introductory slide for the presentation, titled 'Relevance AI \| AI Phone Agents in SaaS Support'.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-01-15.png) ![Screenshot at 02:33: Slide detailing 'The SaaS Support Tier Problem', contrasting traditional hierarchy with Relevance AI's 4-tier approach.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-02-33.png) ![Screenshot at 03:31: Slide titled 'AI Unlocks Support Equity', emphasizing that everyone can now access phone support due to AI integration.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-03-31.png) ![Screenshot at 04:33: Slide titled 'Introducing Harley', detailing Harley's function as an outbound phone agent solving the call access problem.](https://ss.rapidrecap.app/screens/I-qlDMpDKkA/00-04-33.png)