AI Customer Support Agent: RAG-Powered Automation
The context
The project was developed to modernize customer service for a travel-tech environment (Dayuse-like model). Traditional support bots often fail by providing generic answers or « hallucinating » facts. I built a custom Retrieval-Augmented Generation (RAG) system using n8n to bridge the gap between static FAQ documentation and dynamic customer interactions.
The purpose
The goal was to create a reliable, low-latency automated agent capable of:
-
Factual Accuracy: Providing answers based strictly on a private CSV-based knowledge base.
-
Intelligent Escalation: Detecting negative sentiment or low-confidence responses to automatically create tickets in Zendesk.
-
Operational Efficiency: reducing manual agent workload by resolving common queries (e.g., « What is Dayuse? ») without human intervention.
The results
-
Zero Hallucination Rate: Successfully restricted the AI to official documentation, ensuring 100% brand-safe responses.
-
Seamless Ticketing: Automated the creation of detailed Zendesk tickets, including full conversation logs and sentiment analysis for human agents.
-
Scalability: Built a flexible architecture that can update its knowledge base in real-time simply by modifying a Google Drive file.
-
Cost Optimization: Transformed an inefficient 91-call process into a streamlined 1-call execution.
My role
As the Automation specialist, I was responsible for the end-to-end development of the workflow:
-
System Architecture: Designed the multi-branch logic in n8n, connecting Webhooks, AI Agents, and external APIs (Zendesk, Google Drive).
-
Data Engineering: Wrote custom JavaScript to consolidate 90+ rows of FAQ data into a single AI-ready context, optimizing token usage and reducing API costs by 98%.
-
AI Orchestration: Configured LangChain agents with specific « guardrail » prompts to prevent misinformation.
-
Debugging: Resolved complex JSON issues and asynchronous data persistence problems
Core Logic
-
Orchestration: n8n
-
LLMs: OpenAI (GPT-4o / GPT-5 Nano)
-
Data Sources: Google Drive (CSV FAQ), Google Sheets
-
Ticketing/CRM: Zendesk API
Workflow
-
Ingestion: Receives customer queries and metadata via Webhook.
-
Contextualization: Fetches and transforms FAQ data from Google Drive into a searchable knowledge base.
-
Processing:
Sentiment Analysis: Categorizes user emotion (Positive/Neutral/Negative).RAG Agent: Generates answers based strictly on the FAQ context.
Confidence Scoring: Programmatically evaluates the AI’s answer quality.
-
Action:
Success: Delivers the verified answer back to the chat interface.
Escalation: If confidence is low or sentiment is negative, creates a Zendesk Ticket with full conversation logs.