Short answer
AI triage is using AI to read each incoming request, decide what it is about and how urgent it is, and send it to the right place. In customer service, that is the right team, an automated answer or a person, before anyone replies.- Also called
- Automated triage, auto triage, intelligent triage, AI ticket triage
- Decides
- What the request is, how urgent it is, where it goes
- Used in
- Customer service, healthcare, IT and security alerts, bug tracking
- Goes wrong when
- It forces a label onto a message that fits none
The word comes from emergency medicine, where triage means sorting patients by how urgently they need care. AI triage applies the same idea to a queue of requests: tickets, chats, emails, alerts. A model reads each one as it arrives and makes the first decision a team would otherwise make by hand.
How AI triage works
Every triage step makes the same four decisions, in order.
- Read. The model reads the whole message, the subject line and any attachments, however the sender worded it and in whatever language.
- Classify. It labels what the request is about, such as order status, a return or a payment problem, and scores how sure it is of each label.
- Prioritize. It flags urgency. An order about to ship to the wrong address matters more today than a question about next season's sizes.
- Route. It sends the request to a team, a person, an automated answer or a review queue.
Triage decides where a request goes. It does not write the reply, which is why a fast but wrong triage step quietly breaks everything after it.
Manual triage vs automated triage
Rule-based triage sits between the two: keyword rules such as "refund goes to billing" are fast, and misfire as soon as a customer writes "I don't want a refund, I want a replacement".
| Manual triage | AI triage | |
|---|---|---|
| Who sorts | A person reading the queue | A model, on every message as it arrives |
| Speed | Minutes to hours | Seconds |
| Labels | Vary by person and shift | Same rules on every message |
| Unfamiliar requests | Spotted by an experienced teammate | Need a fallback: no label, or a suggested new one |
| Best for | Small queues and judgement calls | High volume and repeated request types |
AI triage in a Shopify store's queue
Five messages from one morning show what the sorting step has to decide.
- "Where is order 4831? It said Tuesday." Order status. Routine, answered from tracking.
- "I put the wrong address, please fix it before it ships." Address change. Urgent, because the window closes when the order is fulfilled.
- "The vase arrived broken, photos attached." Damaged item. Goes to the team or a set damage process.
- "Do you ship to Canada?" Policy question. Answered from the shipping page.
- "You charged me twice." Payment problem. Goes to whoever handles billing.
Each lands somewhere different, and the address change should jump the queue.
In Aide, each destination can be a procedure written in plain words: the broken vase can match one that collects the order number, adds an internal note and assigns the conversation to the group that handles damage. The double charge goes to whoever handles billing through a routing rule, and Aide can make the address change itself for a shopper who proves the order is theirs.
Other kinds of AI triage
The term travels across fields, and the idea is the same in each.
- Healthcare. Symptom checkers and nurse-line tools suggest how urgently a patient needs care, to support clinicians rather than replace them.
- IT and security. Alert triage groups and ranks monitoring or security alerts so engineers see the real incident first.
- Software teams. Bug triage sorts incoming bug reports by severity and owner.
- LLM triage. A large language model does the sorting by reading free text, instead of matching keywords.
What good AI triage looks like
Good triage admits what it does not know.
- It can return no label rather than force a weak match.
- It sorts into the team's own categories, not a generic list.
- Uncertain or sensitive requests go to a person, with the label and the reason attached.
- Its labels are visible, so the team can correct them.
The approach this page rejects is triage as a deflection filter, sorting requests mainly to keep them away from people. Triage exists to get each request to whoever, or whatever, can resolve it.
With Aide, topic chips sit on every customer message and a topic shows on every inbox row, so a wrong label is visible where the team already works. The team corrects the sorting by adding real messages as examples to a topic, or moving an example to the right one, and Aide classifies against those examples from then on. See AI ticket tagging.
Where Aide fits
Aide, the AI platform for customer service, triages every customer message in chat, in email and inside Zendesk, Gorgias or Front. It scores each message against the team's own topics and picks up to three, best first, or none when nothing fits well, and suggests a new topic for messages no topic covers. Topics are written back to the helpdesk as tags or a ticket field, and each topic's AI level decides whether Aide answers, drafts for a person or leaves the request to the team. See what your topics tell you.