Ticket flow: types, frequency, cost
AI for customer support · Lesson 1 / 20
Export first, tooling second
Any conversation about AI in support starts with a ticket export, not with picking a model. Until you know how many tickets arrive in a week, what they consist of and what closing one costs you, you are buying a solution for a problem you have never measured.
Five types of tickets
- Status questions. Where is my order, did my payment go through, when do I get my refund. The answer lives in a database, not in an agent's head. In e-commerce and fintech this is 25–40% of the flow.
- How-to questions. How to change a plan, where to download an invoice, how to connect a domain. The answer exists in the docs; the customer did not find it or did not understand it.
- Something is broken. A real defect: an error, an outage, a wrong calculation. Needs diagnosis and often an engineer.
- Disputes. Refunds, compensation, denials, complaints. The decision costs money and requires authority.
- Not support at all. Sales, partnerships, job applications, spam. These are routed, not answered.
What one ticket costs
The rough calculation is enough: monthly support payroll including taxes and tool subscriptions, divided by the number of closed tickets. Most teams land somewhere between one and five dollars per ticket. Then you break that number down by topic: a status question costs four minutes of work, a refund dispute costs three days of correspondence plus a manager's time.
Ninety-day export, minimum set of columns: id | created at | channel | topic as written | first response, min | resolution time, h | message count | escalated yes/no | rating What you need to pull out of it: 1. Top 20 topics by frequency and their share of total volume 2. Median resolution time inside each type 3. Share of tickets whose answer already exists in the knowledge base
Insight. In most teams the top ten topics carry 60–75% of the volume, and a tail of a hundred rare topics carries the remaining quarter. Automating the tail costs more than handling it by hand.
Common mistake. Tracking average resolution time across the whole flow. A status question closes in four minutes, a refund dispute in three days. The average describes neither scenario and barely moves when you actually improve something.
Pro tip. Add a repeat column to your export: did this same customer write about this same topic in the previous seven days? The repeat rate is the most honest signal that your previous answer did not close the problem.
Cheat sheet
- Export ninety days of tickets before any AI discussion.
- Split the flow into five types: they need different remedies.
- Cost per ticket = support payroll / closed tickets.
- A median inside a type beats an average across the flow.
1. What share of the flow do the top ten topics usually carry?
2. Why is average resolution time across the whole flow misleading?
3. What does the repeat column in the export reveal?