Every IT service team knows the first 30 minutes of the day: the inbox full, the ticket queue full, and before anything substantive can happen, the items have to be sorted. Which are standard requests? Which are urgent? Which go to which team? Which are duplicates? Which are not really an IT topic at all?
This sorting work is not difficult — it is exhausting. It eats up the team’s focused time without solving a problem. That is exactly where one of the clearest levers for AI in IT administration sits.
What a good classifier delivers
A productive AI classification takes over four things:
- Categorisation: which service, which application, which type of request (incident, advice, request, information).
- Prioritisation: how urgent the request is, based on keywords, requester, affected application.
- Routing: which team is responsible, ideally with a concrete agent suggestion (based on history with similar cases).
- Duplicate detection: is a similar case already being worked on, that should be merged?
All four steps can happen before the first human touches the ticket. When the service agent opens the ticket, a justified pre-classification is already there — they check, accept or correct.
What data is actually required
The honest precondition: an AI classifier is only as good as the data it learned from. The first two steps in such a project are rarely technical:
- Prepare historical tickets. Tickets from the past 12 to 24 months are cleaned, anonymised (where personal data should not enter the model) and brought into a uniform format.
- Ensure label quality. If historical categorisation itself is inconsistent — and it almost always is — the AI will learn that inconsistency. A short audit pass with the team — which categories are still genuinely used, which are synonyms — is usually the most valuable preparation.
This is often sobering. It is also often the point where the team first arrives at a clear language for its own work — which is valuable in itself.
What pilot operation looks like
A typical pilot runs in three phases:
- Shadow mode (two weeks): the AI classifies every new ticket without the suggestion being visible. In the background, suggestion versus actual processing is compared. Result: a concrete hit rate — typically between 70 and 90 per cent.
- Suggestion mode (four to six weeks): suggestions are shown on the ticket; the service agent accepts or changes them. Every correction becomes a training signal.
- Auto mode (above 95 per cent hit rate in a category): tickets the AI is confident about, falling into an approved category, are routed automatically — without a human intermediate step.
Step 3 matters only where the hit rate supports it. With uncertain suggestions, the human stays in the loop. There is no reason to prefer bad automation to good semi-automation.
What effort is actually saved
Realistic numbers from real projects:
- 60 to 80 per cent of standard requests are correctly pre-classified.
- Time per ticket in first level drops by 30 to 50 per cent — mainly in the initial sorting phase.
- Reaction time for high-priority tickets drops noticeably, because they no longer drown in the standard queue.
- Escalation rates drop because more tickets reach the right team first time.
What does not happen: staff become redundant. What happens: they have more time for the genuinely demanding cases — the ones where their experience really matters.
Pitfalls we know
Three errors we see regularly in pilots:
- Going to auto mode too early. A 85 per cent hit rate sounds good but means every seventh ticket is mis-routed. That eats the efficiency gain, because mis-routed tickets escalate.
- Black-box suggestions. If the AI only proposes a category without saying why — which keywords drove the call — staff lose trust. Explainability matters more than a few percentage points of additional accuracy.
- A static model. A classifier that is not regularly retrained ages fast. New applications, new vocabulary, new staff — everything moves. An update cycle of four to twelve weeks is mandatory.
What remains in the end
A productive ticket classifier is one of the few AI projects whose value shows clearly in the first three months. It is also an excellent entry into AI-supported IT administration — because the team learns how to handle AI suggestions, how to measure quality, how to recalibrate. What works here transfers to other IT activities: incident classification, change risk assessment, knowledge recommendation.
Sorting tickets is not what IT teams are trained for. It is exactly what AI can take off their plate.