In this project, I built a small AI prototype that classifies IT service desk tickets by category and predicts ticket priority from the text of a support request. The purpose was simple: demonstrate how natural language processing and machine learning can support faster, more consistent ticket triage while still keeping human judgment in the process.
Many IT support teams receive tickets for access issues, application errors, database problems, performance degradation, deployment failures, and general requests. The first routing decision is important because a misclassified ticket can create delays, unnecessary handoffs, and missed service-level expectations.
This example uses simulated service desk tickets so the project can be shared publicly without exposing sensitive company data. Real tickets often include names, email addresses, internal system names, server details, customer information, and operational context that should not be published or used for model training without proper controls.
The Business Problem
Ticket triage is often repetitive, but it is also high impact. A support analyst reads the ticket, identifies the category, estimates urgency, and routes it to the right team. When ticket descriptions are unclear or when multiple teams support the same application, routing can become inconsistent.
The idea behind this prototype is not to automatically close tickets or replace analysts. Instead, the model acts as a recommendation layer. It can suggest a likely category and priority so the analyst has a faster starting point.
Project Approach
I used a supervised machine learning approach. Each simulated ticket has a short text description, a category label, and a priority label. The model learns patterns from the ticket text and predicts labels for new incoming tickets.
The workflow included:
- Creating a simulated dataset of IT service desk tickets
- Cleaning ticket descriptions by lowercasing text and removing unnecessary punctuation
- Converting text into numerical features with TF-IDF
- Training baseline, Logistic Regression, and Linear SVM models
- Evaluating results with accuracy, precision, recall, F1-score, classification reports, and confusion matrices
- Testing the best model on new example ticket descriptions
Why TF-IDF?
TF-IDF is useful for short text classification because it highlights words or phrases that are important in one document while reducing the weight of words that appear everywhere. In ticket data, terms like “password reset,” “database timeout,” “rollback,” or “slow response” can carry strong signals about the type of issue.
This is a practical starting point before moving to more advanced techniques. For many business workflows, a clear baseline and interpretable model can be more useful than jumping directly into a complex architecture.
Models Tested
I compared three approaches:
- Majority-class baseline: always predicts the most common label
- Logistic Regression with TF-IDF: a strong and interpretable text classification baseline
- Linear SVM with TF-IDF: often effective for high-dimensional text features
The baseline was important because it showed whether the machine learning models were actually adding value. Without a baseline, a model score can look impressive while still being weak compared with a simple rule.
Results
| Task | Model | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|---|
| Category | Baseline Majority | 0.178 | 0.032 | 0.178 | 0.054 |
| Category | Logistic Regression + TF-IDF | 0.989 | 0.990 | 0.989 | 0.989 |
| Category | Linear SVM + TF-IDF | 1.000 | 1.000 | 1.000 | 1.000 |
| Priority | Baseline Majority | 0.456 | 0.208 | 0.456 | 0.285 |
| Priority | Logistic Regression + TF-IDF | 0.878 | 0.904 | 0.878 | 0.877 |
| Priority | Linear SVM + TF-IDF | 0.967 | 0.968 | 0.967 | 0.966 |
The Linear SVM performed best overall. It achieved perfect category classification on the simulated test set and strong priority prediction performance.
These results should be interpreted carefully. The dataset is simulated, so the language patterns are cleaner than real-world ticket descriptions. In production, tickets can be vague, incomplete, misspelled, duplicated, or mixed across multiple issues. A real implementation would need anonymized historical data, stronger validation, business review, and monitoring after deployment.
Example Predictions
| New Ticket Description | Predicted Category | Predicted Priority |
|---|---|---|
| Users cannot login to HR Portal because MFA code is not working and multiple employees are impacted. | Access Issue | Medium |
| Production Billing DB has connection pool exhausted and payment processing stopped. | Database Issue | High |
| After the release, the API service deployment failed and rollback may be needed. | Deployment Issue | Low |
| User is requesting a new monthly sales report export for next week. | General Request | Medium |
| E-commerce site has slow response during peak hours for many users. | Performance Issue | High |
What I Learned
This project reinforced a few practical lessons:
- Start with a business problem, not a model.
- Use a baseline so you can measure whether AI is actually helping.
- For short operational text, TF-IDF plus a traditional classifier can be a strong first version.
- Priority prediction is harder than category prediction because urgency depends on business impact, not just technical keywords.
- Responsible AI controls matter, especially when support data may contain sensitive information.
Responsible AI Considerations
A real-world version of this system should include human review before routing changes are applied. The model should be monitored for errors, especially cases where a high-priority ticket is incorrectly lowered to medium or low. That kind of mistake can delay response and affect service-level commitments.
Privacy is also important. Before using real ticket data, organizations should de-identify personal information, mask sensitive system details where appropriate, control access to training data, and document how predictions are used.
Possible Next Steps
- Test the approach with anonymized historical ticket data
- Add metadata such as application, assignment group, business unit, and incident source
- Track high-priority recall as a key operational metric
- Create a simple dashboard for support managers
- Add analyst feedback so the model can improve over time
Conclusion
This project shows how AI can be applied to a practical IT operations workflow. Even a small text classification model can demonstrate the value of faster triage, better consistency, and improved visibility into common support patterns.
The most important takeaway is that AI should support the workflow, not blindly automate it. Used carefully, a model like this can help analysts move faster while keeping people accountable for the final decision.
AI assistance disclosure: This public write-up and repository cleanup were prepared with AI assistance and reviewed by Ravi Baghel.
References
- Bird, S., Klein, E., & Loper, E. (2009). Natural Language Processing with Python. O’Reilly Media.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://doi.org/10.6028/NIST.AI.100-1
- Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, E. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825-2830.
- Taulli, T. (2019). Artificial Intelligence Basics: A Non-Technical Introduction. Apress. https://doi.org/10.1007/978-1-4842-5028-0
- https://github.com/ravibaghel/ai-service-desk-ticket-classification
