Work · Consumer retail
An AI agent that drafts support replies from the help centre and order data
Built a narrow support-triage agent that classifies incoming tickets, pulls the relevant order and policy context, and drafts a reply for a human to approve.
Illustrative example — representative of real engagements, published as a template until client-approved write-ups replace it.
First-response time
6h → 25min
Tickets auto-drafted
~68%
Cost per drafted reply
< $0.02
- A two-person support team was drowning in repetitive questions about shipping, returns, and order status.
- Answers existed in the help centre and the order system, but stitching them together by hand was slow.
The approach
- Scoped the agent to drafting only — a human always approves before anything is sent.
- Indexed the help centre for retrieval; gave the agent a read-only tool into the order API.
- Built the workflow in n8n so the team can tweak routing rules themselves.
- Added per-run token logging and a cheaper model for the easy classifications.
The outcome
- Most tickets arrive with a ready-to-send draft and the sources it used.
- The team reviews and sends instead of researching and writing.
- Monthly LLM spend stayed under the cost of a few hours of the old manual work.
Have a project in mind?
Tell me what you are building. You get a scoping call and a fixed quote — no obligation.