Smart Customer Support - A Multi Agent Ticketing System Using OpenAI Agents SDK

#ai #openai #webdev #agents

Introduction

In this project, I built an intelligent car customer support system using AI agents which converts customer support emails into structured support tickets with AI generated draft response based on internal document references which are then reviewed by human agents.

Handling hundreds of customer support emails at scale is difficult, especially when many request are repetitive. Currently, the support teams often read through hundreds of emails everyday, analyze each request and send meaningful response to each customer. This is a tedious and time consuming process, especially if the customer service agent has to search through various internal documents before responding to each customer.

My goal with this project is to automate this repetitive parts of the support workflow while still giving the humans the final say. This project is a collaboration between AI and human agents. Full code for this project can be found on Link

Project Features

📩 Email ingestion (Gmail)
🤖 AI agent analyze support emails and generates
- Priority
- Summary
- Draft response
🧑‍💼 Human-in-the-loop approval before sending response to users
- ✏️ Editable draft responses
- ✅ Approve / ❌ Reject tickets
📊 Analytics dashboard
- Tickets per day
- Tickets by priority
- Tickets by status
- Total tickets & pending review

🚀 Technologies

PostgreSQL – Data storage
FastAPI – Backend API implementation
LLM (GPT-5 Nano) – Language model powering the AI agent
Next.js – Frontend web application
Docker Compose – Service orchestration

Architecture Overview

AI Agent Orchestration Design

Instead of using a single agent with a large prompt, in this project I have used separate specialized agents with a clear defined role and strict boundaries.

1. Email Analyzer Agent

The Email Analyzer Agent is responsible for processing incoming emails.

Its job is to:

Inspect the email body
Detect whether it contains a valid customer complaint or request
Assign a priority level (LOW, MEDIUM, HIGH, CRITICAL)
Generate a one sentence summary

If the email does not contain a valid request, the agent explicitly marks priority as INVALID.

2. Car Expert Agent (RAG based Domain Expert)

The Car Expert Agent acts as a car specific knowledge agent. It does not rely on general LLM knowledge. The agent provide response by strictly referencing car documentation using a Retrieval-Augmented Generation (RAG) approach. If the requested information is not found, then the agent provide the response accordingly.

3. Orchestrator Agent (Coordinator & Decision Maker)

The Orchestrator Agent is the brain of the system. It coordinates the entire workflow by invoking other agents as tools and enforcing business rules.

Its responsibilities include:

Calling the Email Analyzer Agent
Deciding whether a ticket should be created
Generating a human-readable ticket number
Detecting whether the request is car-related
Conditionally invoking the Car Expert Agent
Producing a draft email response for human review

class OrchestratorAgentOutput(BaseModel):
    priority:str
    summary:str
    draft_response: str
    subject: str
    ticket_number: str

orchestrator_agent = Agent(name="orchestrator agent",
                           instructions=""" 
                            You are a customer care car support orchestrator.
                            Process:
                            1. Use email_analyzer to classify priority and summarize.
                            2. If priority == INVALID:
                            - Create a polite rejection response
                            3. If priority != INVALID:
                            - Determine if the request is about car operation, maintenance, or faults
                            - ONLY IF it is car-related, call car_expert
                            4. Use car_expert output strictly; do not add assumptions
                            5. Produce a draft email response and subject

                            Output MUST match OrchestratorAgentOutput. 
                           """, 
                           model="gpt-5-nano",
                           tools=[    
                            create_ticket_number,     
                            email_analyzer_agent.as_tool(
                                tool_name="email_analyzer",
                                tool_description="Perform analysis of the email and provide a brief summary and priority of the email",
                            ),         
                            car_expert_agent.as_tool(
                                tool_name="car_expert",
                                tool_description="Provides correct answers regarding the car by referring the car documentation",
                            ),
                           ],
                           output_type=OrchestratorAgentOutput
                           )

Here, the orchestrator agent use the email_analyzer_agent and car_expert_agent as tools. This approach keeps the prompts small and focused, it prevents role overlap while also allowing easy tool extension.