Tags are what make it possible to say “billing questions doubled this month”. They're also the first thing that stops being maintained: manual tagging survives a week of good intentions and then quietly lapses.
Three mechanisms apply them without anyone remembering to, and they're worth telling apart.
In your project's AI assist settings, you give a list of tags, and incoming conversations are tagged against it as they arrive — regardless of channel or workflow.
This is the one to set up first. It's the only approach that covers everything, including conversations that never touch a workflow at all.
The AI ticket tagging step does the same job at a specific point in a flow. Useful when a particular path deserves its own vocabulary — a returns flow tagging by reason, say — rather than the general list.
The plain Add tag step applies a tag you name, every time, with no interpretation involved.
This is the right one whenever the tag follows from the path rather than from the content. If someone chose “Report a bug” from your menu, tagging it “bug” needs no cleverness — and a deterministic tag is one you can trust in a report.
Whichever mechanism you use, the quality depends on the list. Keep it short and keep the meanings separate: “billing” and “payment” as two tags guarantees inconsistency, because no rule and no model can reliably tell you which one a given conversation is.
Ten distinct tags applied reliably are worth far more than forty precise ones applied at random. And a tag nobody has looked at in a report is a tag you can delete.
After a week, read fifty tagged conversations and ask whether you'd have tagged them the same way. That's the only test that matters, and it's how you find the overlapping pair that's quietly making your reporting meaningless.
