Case study · Chat & Support

An AI support agent for an online store that looks up real orders and escalates the hard cases

It reads each customer message, works out what they need, answers from live order data, starts returns, and sends complaints and billing disputes to a person.

The problem

Store support is a mix of very different messages in one inbox. Where is my order. I want a refund. Is this backpack waterproof. An angry customer threatening a chargeback. Each needs a different answer, and most need someone to look something up first.

A person has to read each one, work out the intent, find the order, write a reply and decide whether to escalate. Customers wait hours for answers that could have been instant, and the urgent cases wait in the same queue as a tracking question.

How it works

An AI step reads each message and sorts it into one of four intents. The agent then takes a different action for each.

Order status. It looks up the customer's real order and replies with the order number, status, total and tracking.
Product question. A separate AI step answers the customer's specific question.
Return. It confirms the request, tells the customer the instructions are on the way and alerts the returns team.
Anything else. Complaints, billing disputes and anything emotional go straight to a person with a full alert.

Every interaction is also posted to the right team channel in Slack.

n8n workflow: webhook trigger, analyze customer message, parse AI response, route by intent to order search, return, product question or escalation paths, each posting to Slack, then send response — The workflow. One message comes in and the intent router picks the path.

The same workflow after a return request, with the return path highlighted — A return request taking its path through the router.

Where is my order

A customer asked about their order. The agent looked up the record and replied with the order number, status and total, pulled live from the order database.

Slack message titled Order lookup showing order ORD-1006 with status Processing and total $35.20, with the customer's name and email blurred — The order lookup posted to the team. Customer details are blurred.

n8n message editor showing the Slack message template built from the order search fields — The message template. Each field comes from the order search.

Airtable order management dashboard listing orders with status, order number, tracking number and total, with customer names and emails blurred — The order table the agent reads from. Names and emails are blurred.

A return request

A customer wanted to return a jacket that did not fit. The agent recognized a return, confirmed it to the customer and alerted the returns team.

n8n output of the intent router showing the reply, intent return, and the customer's message about returning a jacket, with name and email blurred — The router's output for the return. The intent is return.

Slack message titled Return request with the customer's message about returning a jacket, with the customer's name and email blurred — The alert to the returns team.

An angry customer

A customer had been charged twice and wanted a manager. The agent did not try to handle it with a script. It escalated to a person straight away and acknowledged the customer.

Slack message titled Escalation needed with the customer's message about a double charge and intent other, with the email blurred — The escalation. A person takes it from here.

What a person still handles

Complaints, billing disputes and anything emotional.
The return itself. The agent confirms the request, the returns team handles it.
Anything that falls outside the four intents.

Built with

n8n
OpenAI
Airtable
Slack

n8n, OpenAI for intent classification and product answers, Airtable as the order database, and Slack for routing and escalations. The order database can be swapped for Shopify, WooCommerce or a custom backend without changing the agent's logic.

Your build starts with the arithmetic.

Bring a week of real examples to the free audit. We come back with the three worth doing first, what each is worth, and a fixed price. If a cheaper tool already covers it, we say so.

Book the free audit →