$3M+
annual cost saving
~$10
saving per ticket classified
Classification speed reduced from minutes to seconds
Subjective variation in classification eliminated
A deep neural network was trained to automatically classify IT service desk tickets by team, priority, and category, delivering over $3M in annual savings. The model processes free-text ticket descriptions in seconds with confidence-based routing that defers uncertain cases to human agents, while a continuous learning loop improves accuracy over time.
The Problem
IT service desks receive thousands of tickets daily. Each ticket must be manually triaged: classified by team, priority, and category before it can be routed and actioned. This manual classification is subjective (two agents may classify the same ticket differently), inconsistent (accuracy degrades with fatigue and workload), slow (each ticket requires an agent to read, interpret, and assign multiple fields), and expensive (every minute spent on triage is a minute not spent on resolution). The result is misrouted tickets, delayed resolutions, and significant labour costs tied to a repetitive, low-value task.
The Solution
A deep neural network was trained on a large corpus of historically classified tickets to automatically classify incoming tickets across three dimensions simultaneously: which resolver group should handle the ticket (team), urgency and impact assessment (priority), and service area or problem type (category). The model takes the free-text description as input and outputs predictions for all fields in a single inference pass, replacing minutes of manual reading with seconds of automated classification.
The critical design choice was confidence-based routing. The model outputs a confidence score for each prediction. When confidence falls below a defined threshold, the ticket is escalated to a human agent rather than being auto-classified. This ensures quality is never sacrificed for speed. The system knows when it does not know, and defers accordingly.
A continuous learning loop ties the system together. Tickets that fall below the confidence threshold are classified manually by agents, and those decisions feed back into the training pipeline as new labelled data. The model is periodically retrained, improving accuracy on edge cases, expanding coverage as new ticket types emerge, and adapting to changing language patterns and organisational structure. The result is a system that gets better over time without proportional investment.
Results and Impact
| Metric | Value |
|---|---|
| Cost saving per ticket | ~$10 |
| Annual saving | $3M+ |
| Classification speed | Seconds (down from minutes) |
| Consistency | Eliminated subjective variation |
| Coverage | High-confidence auto-classification on majority of tickets; remainder routed for human review |
Key Takeaways
-
Confidence thresholds protect quality. The system knows when it does not know, and defers to humans accordingly. This made adoption straightforward because stakeholders trusted the output would not degrade service quality.
-
The feedback loop is what makes the ROI compound. Manual classifications on low-confidence tickets become training data, creating a virtuous cycle of continuous improvement. Accuracy rises over time without proportional investment.
-
Free text is a rich signal for classification. Deep neural networks can extract classification intent from unstructured ticket descriptions with high accuracy, making automation viable even where structured input fields are sparse or inconsistently completed.