AI Ticket Categorisation and Routing in 2026

In a traditional service desk, an L1 agent reads each incoming ticket, classifies it into a category (Hardware, Software, Access, Network, Other), assigns a priority and severity, and routes it to the correct queue. The triage step takes 2 to 8 minutes per ticket depending on complexity. For an organisation handling 100,000 annual tickets, that is 3,300 to 13,000 hours of triage labour per year, equivalent to 1.5 to 6 full-time agents whose work product is metadata.

AI categorisation eliminates most of that labour. The AI reads the ticket on intake, applies the category, sets severity based on language signals, applies priority based on user impact, and routes to the queue. The human agent receives a pre-classified ticket and accepts or corrects the classification with a single click. Triage time drops to under 30 seconds in mature deployments. The displaced labour redirects to higher-value work or absorbs ticket-volume growth.

Beyond labour saving, AI categorisation improves quality. Human triage in busy queues produces a substantial rate of misclassification, often 10 to 20 percent on fine-grained categories. Misclassified tickets get routed to the wrong queue, sit longer, and require re-routing. AI categorisation with a 90 percent accuracy rate often outperforms tired human triage. The combination of faster and more accurate first-time routing improves mean time to resolution by 15 to 30 percent in published cases.

AI categorisation accuracy is almost entirely a function of training data quality and quantity. The minimum useful corpus is approximately 5,000 historical tickets with human-applied categories, resolutions, and routing destinations. Below this, the classifier struggles with long-tail categories. Above 50,000 tickets, additional volume yields diminishing returns; the marginal accuracy improvement from going from 100,000 to 500,000 tickets is typically less than 2 percentage points.

Quality matters more than quantity. A 10,000-ticket corpus with consistent human categorisation, complete resolution notes, and clean category metadata will outperform a 100,000-ticket corpus where 30 percent of tickets are mis-categorised and resolutions are blank. The first step in any AI categorisation deployment is a training-data audit: how consistent is the existing categorisation, what percentage of tickets have complete resolution notes, what is the agreement rate when two agents categorise the same ticket.

If the existing data fails the audit, the right path is data remediation before model training, not model training on bad data. Most enterprises need 80 to 200 hours of analyst time to label a curated training set, retire mis-categorised historical data, and define a clean category taxonomy. This work is unglamorous and unavoidable. Vendors that promise to skip it are setting up the deployment for accuracy disappointment in months three through nine.

The category taxonomy itself deserves attention. Most enterprises operate with category hierarchies that grew organically over years and contain redundancy, ambiguity, and unused branches. A pre-deployment taxonomy review (rationalising to 30 to 80 leaf categories, retiring unused branches, merging overlapping categories) typically improves classifier accuracy by 5 to 10 percentage points and improves human-agent satisfaction with the AI categorisation simultaneously.

Predicting severity and priority is harder than predicting category. Category is largely a function of what the ticket is about (Outlook, network, VPN). Severity is a function of business context (is the user a finance executive on the day of a board meeting, is the system a revenue-impacting application, is the issue affecting a single user or a team). The AI classifier has limited visibility into business context unless that context is explicitly modelled in the training data.

Practical severity prediction in 2026 reaches 65 to 80 percent accuracy. The most successful pattern is a hybrid: the AI predicts a baseline severity from language signals (downtime keywords, impact statements, user role) and the human triager confirms or escalates. This catches the obvious P1 cases (outage, critical user, revenue-impacting) at AI speed while preserving human judgement for ambiguous cases. Pure-AI severity is brittle; pure-human severity is slow; the hybrid pattern outperforms both.

The implementation pattern for severity worth knowing: include the user's role metadata, the affected system's criticality tier, and the time of day in the classifier features. A ticket from a CFO at 4pm about the board-meeting webcast is materially more urgent than the same words from a contractor about the optional weekly social meeting. The AI needs this context as structured input, not as language to infer.

Classifier accuracy degrades over time without active maintenance. The drivers are vocabulary drift (new applications, new acronyms, new processes), organisational drift (team reorganisations, queue restructuring), and behavioural drift (users describing the same issue with different language). Mature deployments maintain accuracy through a structured feedback loop with three components.

First, instrument the corrections. Every time a human agent re-routes or re-categorises a ticket the AI handled, that correction should be captured as a training signal. The vendor platform should expose this metric (correction rate per category, correction patterns by team) so the AI ITSM admin can see drift early. A correction rate above 15 percent on a category is a signal that the category needs taxonomy review, training-data refresh, or both.

Second, schedule a quarterly retraining cycle. The retraining incorporates the recent corrections, retires categories that have been deprecated, and adds new categories for genuinely new ticket types. A quarterly cadence is sufficient for most enterprises; faster cadences add operational overhead without proportional accuracy gain. Vendors typically include retraining in the platform service; in-house builds need to schedule and operationalise it.

Third, run a monthly review of low-confidence routing decisions. The AI assigns a confidence score to each classification; decisions below threshold (typically 0.7) should already be escalating to human triage. The monthly review looks at the distribution of low-confidence decisions to identify patterns (new system, new vocabulary, ambiguous category) and adjusts the taxonomy or training data to address them. This is roughly 4 to 8 hours of analyst time per month and pays back as steady accuracy improvement quarter over quarter.