A national New Zealand retail group was spending 3,200+ staff hours a month answering repetitive customer queries, with a four-hour average response time. Avtrix built a retrieval-grounded AI support agent integrated with Zendesk. It now resolves 78% of queries automatically, has cut average response time to 8 minutes, and saves the business around NZ$180,000 a year in support costs.
The Client and the Challenge
The client is a multi-store retail group with an online store and a central customer-service team. Most incoming queries were the same dozen questions in different words: order status, returns and exchanges, stock availability, store hours, loyalty points and delivery timeframes. The team was handling them by email and web chat during business hours only.
The numbers had become a problem. Over 3,200 support hours a month were going into repetitive replies; the average time to first response was four hours; evening and weekend enquiries waited until the next working day; and customer-satisfaction scores were sliding. Hiring more agents would have scaled the cost without fixing the delay.
What We Built
Avtrix designed and built a custom AI support agent that sits inside the client’s existing Zendesk workflow rather than replacing it.
- Grounded in the client’s own knowledge. Using retrieval-augmented generation (RAG), the agent answers only from approved sources — the returns policy, product catalogue, delivery rules and store information — so answers are accurate and can be traced to a source.
- Connected to live systems. Order-status and stock questions are answered from real data via the order-management and inventory APIs, not from a static FAQ.
- Human hand-off built in. Anything outside its confidence threshold — complaints, refunds above a limit, anything emotional — is routed to a person with the full conversation and a suggested reply attached.
- 24/7 on every channel. Web chat and email are handled around the clock, with the same tone and rules the client’s team already used.
OpenAI GPT-4 LangChain Python AWS Zendesk integration RAG
How We Delivered It
Discovery
A week analysing six months of tickets to find the query types worth automating and the ones that must stay with people.
Scope & fixed price
A written scope covering the query categories in and out, the hand-off rules, the success metrics and one fixed price.
Build & test on real tickets
The agent was tested against thousands of historical tickets, with the support team grading answers, before it spoke to a single customer.
Staged launch
Live on web chat first, email second, with the team reviewing a sample of every day’s conversations for the first month.
Results Delivered
78% of queries resolved automatically
Nearly four in five conversations are completed by the agent without a person touching them, with customers able to ask for a human at any point.
Response time from 4 hours to 8 minutes
Including evenings and weekends, which previously waited until the next business day.
NZ$180,000 annual cost saving
Support hours were redeployed to complex cases, proactive outreach and store support rather than cut outright.
Full audit trail
Every answer is logged with its source, so the team can see why the agent said what it said and correct the knowledge base rather than the model.
What This Means for Your Business
The pattern applies to any NZ business with a high volume of repeat enquiries — retail, e-commerce, property management, tourism, healthcare admin. The building blocks are the same: your own content, your live systems, clear hand-off rules and a staged launch. If your team is answering the same questions every day, an AI customer support agent is usually the highest-return first AI project, and it is scoped and delivered at a fixed price.
Frequently Asked Questions
Related Services
Client details are anonymised for confidentiality. Figures are as reported by the client after go-live. Delivered by Avtrix Software Solutions Limited, Taupō — serving businesses across New Zealand.