Classifying Every Support Ticket the Moment It Arrives

The challenge

A support ticket arrives as free text written by someone having a bad day. Before anyone can route it, prioritise it or report on it, somebody has to decide what it actually is. Done by hand that decision is inconsistent between technicians, and it is made again every time the queue is triaged.

Where we came in

Classification is the unglamorous foundation the rest of the AI work stands on — you cannot retrieve a similar ticket, measure a response time, or let an agent attempt a reply until you know what kind of ticket you are holding. We treated it as infrastructure rather than a feature, which is why it runs on every ticket rather than on request.

What we did

• Built a classification step that assigns level and category as tickets arrive, rather than at triage time
• Generated transformer embeddings of both the problem and the resolution text, stored as vectors alongside the ticket
• Used those embeddings for similarity search, so a new ticket can find the ones already solved
• Generated readable ticket titles from customer text instead of leaving whatever was typed in the subject line
• Moved our response-time metric onto the classifier's first-action determination rather than a manual field
• Ran classification as background work with retry and backoff, so a model outage delays a ticket rather than dropping it

Where it landed

• Every ticket is classified on arrival rather than at the point somebody gets to it
• Similar past tickets are findable by meaning rather than by keyword search
• The average-response-time figure published on this website is derived from the classifier's own determination