From 140 unowned triggers to a Zendesk where AI classifies and suggests replies

Challenge

The shop sells fashion and accessories in Spain and Portugal. In peak season it handles around 4,500 tickets a month, mostly by email and chat: where is my order, size exchanges, returns and courier problems.

An agency had set up Zendesk three years earlier. The person who maintained it in-house left and nobody took over. The result:

• 140 active triggers, several of them contradictory. Some tickets received two different automatic replies.
• A “contact reason” field with 47 options, filled in by hand and wrong in half the cases.
• Six agents spent the first hour of every shift sharing out tickets.
• First reply took 9 hours on average on working days.

The head of customer service wanted AI to reply faster. But putting AI on top of that mess would have multiplied the errors.

Solution

We started with Zendesk administration on the Extended plan (€1,000/month + VAT, 20 h/month).

Month 1: clean-up. We inventoried every trigger and automation: what each one did and when it last fired. We went from 140 to 38 active triggers, documented and with a clear execution order. Contact reasons went from 47 options to 12, the ones that actually drive a decision. We rebuilt the views by team and priority. I explain the method in how to clean up Zendesk triggers before adding AI.

Month 2: AI classification. When a ticket arrives, Claude reads the message, sets the contact reason and priority, detects the language and pulls out the order number if there is one. With that number it queries Shopify and adds the order and shipping status to the ticket. The now-tidy triggers route it to the right group. This is the workflow described in AI ticket triage in Zendesk.

Month 3: suggested replies. For the five most common contact reasons, Claude drafts a reply as an internal note, using the order data and following the brand’s macros and tone of voice. The agent reviews it, edits if needed and sends it. No message reaches a customer without a person approving it.

Result

Six months after the start:

• 38 active triggers instead of 140, and no more duplicate automatic replies.
• Time to first reply down 58%, from 9 hours to under 4.
• 45 fewer hours a month spent on triage and drafting. Agents no longer share out tickets in the morning.
• Automatic classification agrees with human review in 93% of cases. Agents accept 64% of suggested replies without changes.
• Explore reports by contact reason finally reflect what is actually happening.

“The most useful part was not the AI. It was having someone finally understand our Zendesk and leave it in order.” Head of customer service (anonymised)

What I would do differently

I would have measured suggested-reply quality per contact reason from day one. Returns replies worked almost immediately; courier incident replies needed two more weeks of tuning the instructions.