Plan requirement
| Subscription | Suite Growth or higher |
| Also required | Copilot |
| Access | Admin |
Poor labelling has three causes, and all three are fixed by changing the topics rather than waiting for the model.
Measure it first
- Take twenty recent tickets for one topic.
- Read each and decide whether the label is right.
- Note the wrong ones and what they should have been.
- Repeat for your two or three most-used topics.
That gives you a number instead of an impression, and impressions about this are usually formed from the one bad example somebody remembers.
One: topics that overlap
Two topics a person could not reliably tell apart from the ticket text. The labels will be split arbitrarily between them.
Merge them, or rewrite both descriptions until the boundary is a sentence you could give a new colleague.
Two: a topic that is missing
A real subject with nowhere to go lands in whatever is nearest, contaminating that topic as well. A large "other" group, or one topic that has quietly become a catch-all, both point here.
Three: descriptions that are too thin
A topic named "Technical" with no description matches almost anything. Write what belongs in it and, more importantly, what does not.
Use the agents' corrections
Where agents keep changing one topic to another, that pair is your problem, identified for free. Look at the corrections monthly.
Change one thing, then wait
Adjust one topic and give it two weeks of tickets before judging. Changing four descriptions at once leaves you unable to tell which helped.
Comments
0 comments
Article is closed for comments.