A ticket taxonomy agents will actually use
How I design a support ticket taxonomy that agents use and AI can classify: two levels, plain names, a clear owner per tag, and a monthly prune.
Automation inherits your taxonomy. If the tags are vague, overlapping or too many, the AI will be vague too, and the dashboards will show noise. Fixing the taxonomy is usually the cheapest win before any workflow.
What usually goes wrong
A taxonomy nobody uses
- Dozens of tags added over the years
- Names written by the reporting team
- Tags that overlap
- No owner on most tags
One that works
- A short list, pruned every month
- Names in the agents' own words
- One clear home for every message
- Each tag maps to one team queue
The shape
01Intentdelivery, billing, account, product, feedback
02Detailnot_collected, late, damaged, refund, and so on
03Prioritynormal, high, VIP
04Ownerthe team queue that acts on it
How it connects
The same list feeds three things, which is why it has to be right:
Routingthe owner field decides the HubSpot queue and Slack channel
ClassificationClaude picks from exactly this list, nothing else
Reportingweekly volumes by intent and detail, same names everywhere
The monthly prune
Once a month, look at tags with almost no use and tags that collect "other". Merge the first, split the second. Ten minutes a month keeps the classifier and the reports honest.