The problem: a team of 3 and 800 tickets a month
This e-commerce business had a classic growth problem: sales had tripled in 18 months, but the customer service team hadn't grown — still 3 people handling a volume of requests that had gone from 250 to 800 a month. Response times had stretched from 4 to 36 hours. CSAT had dropped from 4.8 to 3.9 out of 5. Two of the three support agents were considering leaving.
The solution wasn't hiring 5 more people (cost: +£120K/year). It was understanding which requests didn't actually need human judgement and automating them — leaving the team to handle only the situations that genuinely required empathy, negotiation or complex decisions.
Step 1: Classifying tickets by type and complexity
Before touching any tool, we analysed 3 months of historical tickets. The result surprised even the client:
Order status and tracking
"Where's my parcel?" "When will it arrive?" — requests with a 100% automatable answer. The courier's tracking system always has the answer. It just needed an integration.
Standard returns and refunds
Return requests within the warranty period, with the product intact. The policy was clear. Every agent applied the same answer. Fully automatable.
Product FAQs and compatibility
Questions about sizes, available colours, product compatibility. All answers existed in the catalogue but weren't easy for the user to find.
Cases requiring human intervention
Complex complaints, exception situations, VIP customers, requests needing manual verification against internal systems. These 160 tickets a month were the team's real work.
80% of tickets had a standardisable answer. The human team was spending 80% of its time on work that required no judgement — and 20% of its time on what actually mattered.
Step 2: Architecture of the automation system
We built the system in three tiers, each with a clear escalation logic to the next.
Tier 1 — Instant automatic response (0-5 seconds)
An AI agent connected to the CRM and order-tracking system responds automatically to the most frequent categories. For order-status tickets: the system reads directly from the shipping provider and generates a personalised response with a real-time tracking link. For standard returns: it checks the purchase date, confirms it falls within policy, generates the return code and sends instructions.
Tier 2 — AI with a knowledge base (5-30 seconds)
For product questions and FAQs, we built a structured knowledge base with all the technical documentation, sizing guides and up-to-date policies. The AI model (Claude API) retrieves the relevant information and generates accurate answers in the brand's tone. If the confidence score drops below 0.75, the ticket is escalated to tier 3.
Tier 3 — Human agent with full context
When a ticket reaches the human team, the agent sees: the customer's full history, previous automated response attempts, the reason for escalation, and a suggested reply draft from the AI. Handling time dropped from 12 minutes to 4 minutes per escalated ticket.
The tools used (and why)
| Tool | Role | Why we chose it |
|---|---|---|
| Gorgias | Main help desk | Native Shopify integration, advanced automation rules, robust API |
| Claude API | AI response generation | Better context and tone understanding than the alternatives we tested |
| Make (formerly Integromat) | Workflow orchestration | Connects Gorgias, CRM, shipping provider and knowledge base with no code |
| Notion API | Knowledge base | Easy for the team to update, flexible structure, reliable API |
The results after 90 days
The system went live after 6 weeks of setup and testing. The numbers 90 days after full activation:
- First response time: from 36 hours to under 2 minutes for automatable tickets
- Percentage of tickets handled automatically: 74% (vs. a theoretical 80% — some borderline cases are escalated for safety)
- CSAT: back up from 3.9 to 4.7 out of 5 — higher than before the crisis started
- Cost per ticket: from £18.40 to £5.20
- The team of 3 now handles the ticket volume with no overtime, freeing up time for proactive projects
CSAT improved despite the automation because response speed matters as much as (or more than) personalisation for most customers. A correct automated reply in 2 minutes beats a human reply in 36 hours.
Want to automate your customer service?
We'll show you what percentage of your tickets is automatable and how to build the system without losing quality.
Let's talk about your project