How I Built an AI Agent That Handled 90% of Customer Requests Without Human Intervention

In early 2024, a phone repair shop faced high volumes of customer inquiries through various channels, overwhelming staff with repetitive tasks. This led to a project to build an AI agent named Jacobo to automate customer interactions. The main goal was to alleviate the team's workload and improve customer service. Jacobo was designed using a modular architecture with a main router that delegated tasks to specialized sub-agents. These sub-agents focused on appointments, discounts, and order processing, all querying data directly from the existing Airtable database. Key decisions included tool calling for task execution and HITL handoff for complex scenarios. After a year, Jacobo handled around 90% of interactions autonomously, freeing up staff to focus on repairs. The system's success was measured by deflection rate and the system became a valuable asset. The project generated learnings on LLM integrations, observability and simplicity of code. The business's value was enhanced, resulting in a successful sale of the business.