62% ticket deflection. Response time under 2 minutes.

The challenge

With 2,400 support tickets arriving each month and only three agents handling them, this online retailer was running a permanent support backlog. Average first response time was four hours — unacceptable for customers wanting instant answers about order status, returns, or product questions.

The support team had tried FAQ pages and a basic chatbot, but neither had meaningful deflection rates. Customers ignored the FAQ and the chatbot couldn't handle the variation in how people phrased the same question. Tickets kept arriving at the same rate, and the team was falling further behind during peak periods.

The operations manager was under pressure to either hire two more agents — a significant cost for a business at their scale — or find a way to automate. They came to Gilligan Tech asking if AI could actually deflect tickets at the level marketing vendors claimed, without generating a flood of bad automated responses that would make things worse.

The solution

Gilligan Tech built a two-layer automation system on Azure OpenAI. The first layer uses GPT-4o mini for rapid ticket triage — classifying every incoming ticket by intent (order status, return request, product question, complaint, edge case) in under two seconds. This classification determines the routing path.

The second layer handles resolution for the high-confidence, high-frequency categories. Order status queries are resolved by connecting GPT-4o to the retailer's order management API — the model fetches the current status, translates it into a human-readable response, and sends it. Return requests trigger an automated return authorisation workflow. Common product questions are answered using a small RAG pipeline over the product catalogue.

Edge cases, complaints, and any ticket where the model's confidence falls below a threshold are routed immediately to the human queue with a pre-generated context summary — so agents spend less time reading ticket history and more time actually resolving issues.

Implementation

1. Ticket analysis: Gilligan Tech audited 3 months of historical tickets to identify the top intent categories and their volume share. Order status (34%), return requests (18%), and product questions (14%) represented 66% of total volume — the automation target.
2. Triage layer: GPT-4o mini was deployed as a fast classification layer. Structured output mode returns a JSON object with intent category, confidence score, and order ID (if present). Processing time: under 2 seconds per ticket.
3. Resolution layer: GPT-4o was connected to the order management API via function calling. For order status queries, the model fetches live data and generates a natural-language response. Return requests trigger a pre-approved workflow with a confirmation number.
4. Human escalation: Any ticket below 0.85 confidence, or in a category not covered by automation, is routed to the agent queue with a GPT-4o-generated context summary (customer history, order details, inferred issue).
5. Quality monitoring: A daily report flags any automated response that received a negative customer reaction (explicit complaint or re-open). These are reviewed and used to refine the triage thresholds and response prompts.

Results

Technology