Tagging is the part of chat work everyone agrees is valuable and nobody wants to do at the end of a long conversation. So the tag panel in the agent console has a Suggest button that asks a model to do the first pass. It is a small feature with a few behaviours worth knowing, because two of them decide whether your tag data ends up trustworthy or quietly wrong.
What the button does
Open a chat in the console, find the conversation tags panel, and click Suggest. The button needs an active conversation; with nothing open it tells you to open a chat first. Behind it, the transcript and your list of enabled tags go to the model, which returns up to three tag names, and those come back as suggestions on the conversation a second or two later.
It is agent-initiated, not automatic. Nothing is tagged unless somebody presses the button, which means the feature never runs up costs in the background and never surprises you with tags on conversations nobody looked at.
It can only pick tags you already have
The model is given your enabled tags by name and told to choose from them. When the answer comes back, every name is matched against that same list and anything that does not match exactly is discarded. It cannot create a tag, cannot rename one, and cannot suggest a category you have not defined.
Two things follow. The first is reassuring: your tag list cannot be polluted by a model inventing plausible-sounding categories, which is the failure mode people reasonably fear. The second is the constraint that decides the whole feature's usefulness. If your tag list is vague, overlapping or enormous, the suggestions will be vague, overlapping and arbitrary, because the model is choosing between the options you wrote. If you have no enabled tags at all, the button does nothing whatsoever.
So the work that makes AI tagging good is the same work that makes human tagging good: a small, clearly defined, mutually exclusive list, with descriptions that say when each one applies. Our guide on chat tagging and categorization is the prerequisite, not the alternative.
Suggestions arrive as pending, not applied
A suggestion is stored differently from a tag an agent chose. It appears as a pill with a dashed outline and a small keep control, marking it as proposed rather than decided. Keeping it confirms the tag and records the agent who kept it as the person who applied it, which is what you want for accountability. Dismissing it removes the row entirely.
A tag already on the conversation is never suggested again, whether an agent applied it or a previous run proposed it, so pressing Suggest twice does not double anything up. And if an administrator disables a tag while a suggestion is in flight, that suggestion is dropped rather than applied, so a retired category cannot come back through the side door.
Pending tags still count in your reports
This is the behaviour to plan around. The distinction between a suggestion and a confirmed tag lives in the conversation panel, where an agent can see the dashed outline. It does not reach the reporting: the tag usage counts on your tag screen, and the most-used tags on the AI usage screen, count every applied row regardless of whether anyone confirmed it.
The consequence is straightforward. Suggestions that nobody reviews are not neutral, they are votes. A team that presses Suggest on every conversation and reviews nothing has a tag report that reflects what a model guessed from the first part of each transcript, presented with the same authority as tags a human chose deliberately. That is worse than not tagging at all, because it looks like data.
The fix is a habit rather than a setting: whoever presses the button reviews the result before moving on. It takes seconds, since there are at most three pills to look at, and it is the difference between a useful trend and a fictional one.
It reads the beginning, not the whole chat
Only the opening stretch of the conversation is sent. The console trims what it collects, and the server trims it again to roughly three thousand characters before the model sees it, so a long conversation is categorised on how it started.
For most chats that is the right call and costs nothing, because the reason someone got in touch is nearly always in their first few messages. It matters in exactly one situation, which happens to be an important one: a conversation that begins as a small question and turns into something else. A billing query that becomes a cancellation, a how-to that becomes a bug report. Those are the chats you most want tagged correctly, and they are the ones where the suggestion will describe the opening rather than the outcome. Press Suggest by all means, then correct it.
The times it goes quiet
Nothing new to add gives you a message saying the conversation is already covered. That is the common case and it is fine.
Worth knowing: a run that fails on the provider's side, a rate limit or a slow response, produces the same quiet result rather than an error. So a Suggest that reports nothing is not proof that your tags were already right. If you press it on a substantial conversation with an empty tag list showing and get nothing back twice in a row, treat that as a signal to check rather than as an answer.
The feature also requires the managed AI plan. Accounts running AI on their own provider key get a message pointing at the upgrade instead of suggestions, which surprises people who reasonably assume their own key pays for everything. That split is explained in managed AI replies or your own key, and the rest of what the assistant does with your conversations is in the AI chatbot feature tour.
What to measure
The share of suggestions kept. If almost everything is kept, either your tag list is excellent or nobody is really looking. Ask an agent to talk you through three recent conversations and you will know which.
Tagged share of conversations. The point of the feature is coverage. If the tagged share has not moved since you enabled it, agents are not using the button, and a two minute demonstration fixes that faster than a policy does.
Your quietest tags. A tag that never gets suggested and rarely gets chosen is a tag whose definition nobody shares. Rewrite it or retire it.
Put it into practice
- Fix the tag list first. Suggestions can only be as good as the options you wrote.
- Review every suggestion. Unconfirmed pills still count in your reports.
- Expect the opening to dominate. Correct chats that changed direction.
- Do not read a quiet result as agreement. It can also mean a failed call.
- Keep the list short so the model and your agents are choosing between real alternatives.
- Check the plan requirement before promising the feature to your team.
- Spot-check monthly by re-reading five tagged conversations end to end.
Used with a reviewed tag list, the Suggest button removes most of the friction that stops tagging happening at all, and leaves the judgement where it belongs. Used unattended, it fills your reports with confident guesses about the first paragraph of every conversation. The feature is the same in both cases; the difference is entirely in whether somebody looks. The same principle applies to everything the assistant produces, which is why reviewing AI answers is a habit worth building at the same time.