# We Built an AI Support Agent That Resolves 80% of Tickets — AssemblyAI

## Executive summary

Matt Lawler of AssemblyAI details the creation of 'Joey,' an advanced AI support agent designed to solve the scaling bottleneck inherent in the Forward Deployed Engineer (FDE) model. Faced with 1,000 API signups daily and limited human resources, the team replaced an ineffective off-the-shelf bot (10% resolution) with Joey. Joey, built on the Claude Agent SDK, achieves an 80% end-to-end resolution rate for approximately $700/month. The architecture leverages local markdown documentation, external embeddings, agentic file search, and a large `CLAUDE.md` system prompt. The agent's capabilities were further enhanced by integrating AssemblyAI's Voice Agent API, allowing for real-time, low-latency voice interactions.

## Key takeaways

- Scaling the FDE Model: The FDE model, while effective for building trust, does not scale to handle high volume (e.g., 1,000 API signups/day). Automation is necessary to maintain a high-quality customer experience at scale.
- Agentic Architecture Improvement: Joey was built to overcome limitations of vendor bots by giving the agent full control over its infrastructure, including file system access, tool calling, and the ability to write/debug code.
- Performance Gains: The agent increased the end-to-end resolution rate from 10% to 80% in the first week of deployment, costing around $700/month.
- Voice Integration: The agent was given a voice interface using the Voice Agent API, which strings together speech-to-text, LLM, and text-to-speech over a single websocket connection for real-time, low-latency interaction.

## Technical details

- Agent Architecture: Joey is built on the Claude Agent SDK and utilizes four major components: 1) Local markdown documentation (ensuring the agent has access to all current docs, even if the main doc site is down); 2) Embeddings (from Voyage) for efficient retrieval; 3) Agentic file search for deep resource stringing; and 4) A large `CLAUDE.md` file containing guardrails and operational advice.
- Deployment and CI/CD: The system was deployed on Railway, enabling rapid iteration. Fixes can be written as a PR and deployed live within approximately 30 seconds, allowing for real-time bug fixes during a live session.
- Voice Agent Implementation: The voice functionality uses the Voice Agent API, which manages the entire pipeline (speech-to-text $\rightarrow$ LLM $\rightarrow$ text-to-speech) over a single websocket connection, handling pauses and interruptions.

## Practical implications

- FDEs should proactively automate their own roles by building the product their customers are building (dogfooding).
- The best way to understand customer needs is to build the solution yourself, gaining better empathy for technical challenges like latency and turn-taking.
- AI agents can handle 100% of inbound tickets, reserving human intervention only for complex issues (e.g., rate changes, legal agreements).

## Topics

AI Agents, Voice AI, LLMs, Automation, Technical Scaling, FDE, AssemblyAI, Voice Agent API docs, Introducing the Voice Agent API

Source: https://www.youtube.com/watch?v=pyvRID_CZZU