Auto-labeling: topic, priority, confidence
Every new conversation is stamped with topic, priority, and a confidence score. Filters and routing rules use these.
Auto-labeling: topic, priority, confidence
Every new conversation gets three pieces of metadata stamped onto it the moment it lands:
- Topic. What kind of ticket this is.
- Priority. How urgent it is.
- Confidence. How sure the AI is about the label, 0 to 100%.
You use those to filter the inbox, drive routing rules, set SLA, and gate AI behavior.
Topics
The default topic taxonomy:
- bug. A claim that the product is broken.
- billing. Subscription, invoice, refund, payment method.
- how-to. "How do I do X" type questions.
- feedback. Feature requests, opinions on the product.
- abuse. Threats, harassment, spam from customers.
- spam. Non-customer noise (cold outreach, automated mail).
- other. None of the above.
These are not configurable today. They are deliberate: a small, sharp set covers ~95% of inbound and keeps cross-team meaning consistent.
Priorities
- low. Cosmetic, nice-to-have, no time pressure.
- normal. Standard support work.
- high. Real impact, customer is blocked or unhappy.
- urgent. Outage-level, paying customer down.
Priority is calibrated against your past tickets. The classifier looks at how your team has historically prioritized similar messages.
Confidence
Confidence is a 0 to 100 score on how sure the AI is about its labels. Two layers:
- Topic confidence. How sure about the topic.
- Priority confidence. How sure about the priority.
The lower of the two is the displayed score.
Confidence floors
You set a minimum confidence floor. Below it, labels are not auto-applied. The conversation is left unlabeled and the AI flags it for human review.
Defaults:
- Auto-label floor: 60%.
- Auto-route based on label floor: 75%.
You can change both on AI → Auto-labeling.
Read Confidence thresholds for the broader picture, including how confidence interacts with Auto-send vs draft vs suggest.
Where labels show up
- The conversation header in the inbox.
- The filter sidebar (filter by topic, priority).
- Routing rules. See Routing rules.
- SLA policies. See SLA policies.
- AI receipts. See AI receipts.
- Custom field auto-fill (see below).
Custom-field routing
The classifier also fills custom fields on the conversation when you map a topic or priority to a custom field. For example, map "billing" topic to a team field with value billing-team, and routing rules can branch on the field directly. The mapping lives on the Auto-labeling page.
How labels drive behavior
Three common patterns:
- Bug bypass. Tickets labeled "bug" skip AI drafting and go to engineering. See Guardrails and bypass labels.
- Urgent routing. "Urgent" priority pages on-call.
- Topic-based assignment. "Billing" goes to the billing-trained agents.
Editing a label
Agents can edit a label at any time. The edit is captured and feeds back into AI calibration. If your team consistently corrects "feedback" to "feature-request" via tags, the AI learns the pattern.
Manual edits also bypass confidence: a human-set label is treated as 100% confident.
Disabling auto-labeling
You can disable it entirely on the Auto-labeling page. Topic and priority become manual-only fields. Receipts still record what the AI would have suggested, but they do not get stamped.
This is rare. Most teams find labels useful even when AI replies are off.
Labels and AI replies are separate
Auto-labeling and AI drafting are independent. You can:
- Auto-label without AI replies. Useful for triage-only setups.
- AI replies without auto-labeling. Possible but unusual: bypass labels are a key guardrail.
Custom topics
Not supported today. The seven-topic taxonomy is fixed. We add to it carefully, since every change reshapes routing rules and reports across all customers.
If you need finer-grained categorization, use tags on top of the topic. Tags are free-form and per-workspace.
Recommended setup
- Auto-labeling: on.
- Auto-label floor: 60%.
- Auto-route floor: 75%.
- Bypass labels: bug, abuse, spam (default).
- Tag layer for finer categorization where you need it.
Related
Was this article helpful?