Skip to content

Ticket classification with AI

We train a model to suggest a category, priority and handling team, and every correction improves it. We built Dex for PostNL’s service desk.

Dex helps PostNL answer questions faster and connect employees with the right colleague more quickly

Dex is a self-learning chatbot, not a ticketing model. For PostNL, we created a single point where employees can ask questions, connected to existing systems, so service desk staff can focus on more complex issues.

Read the case study

Why does a fault report go to the wrong team first?

Reports come in through the portal, by phone and by email. The first-line team selects a category from a long list, estimates the priority and forwards the ticket. Sometimes it comes back. A model suggests those choices based on the description and previous tickets.

Book a call

Each ticket gets a suggested category, priority and handling team, with an explanation

Discuss your service desk

Category and handling team

The model suggests a category and handling team based on the description and previous tickets. For Olympia, we built a classification model that matches candidates to job openings.

Duplicate reports and surges in incidents

Multiple reports of the same fault belong to a single incident. The model recognises similar reports and flags a sudden surge.

A knowledge article for the ticket

For each ticket, the model finds the most relevant knowledge article or previous solution, so the person handling it does not have to start from scratch.

Corrections help train the model

If the person handling the ticket changes the category or team, that choice becomes a new training example. You can see how often the suggestion was right for each category.

Your existing ticket history and category tree are the starting point

  1. Ticket history and category tree

    We review the ticket history and work with the service desk to remove unnecessary and duplicate categories.

  2. Testing on past tickets

    Using tickets the model has not seen before, we measure how often its suggestion is correct for each category and support team.

  3. Working within the service desk

    Suggestions appear in the ticket. At first, they are just suggestions; later, they are applied automatically where the model is confident enough.

Does our service desk receive enough tickets, and are they categorised consistently enough for a model?

Assess it in a conversation

A good fit if

Your service desk receives a high volume of requests every day, tickets regularly go to the wrong assignment group, and your ITSM system holds the ticket history, including the category and group used to resolve each one.

Not a good fit if

You receive few requests, or one team handles everything. A shorter menu in your self-service portal or a routing rule in your system will help more than a model.

What determines the scope

How complex your category tree is, how many assignment groups and languages you have, whether priority and duplicate requests are included, and which system you use, such as TOPdesk, ServiceNow, Jira Service Management or Zendesk.

Will first-line support trust the recommendation?

That’s why we show it as a suggestion first, alongside similar past tickets. We only route tickets automatically for categories where the recommendation is demonstrably accurate most of the time. Every correction remains visible.

Routing rules, your software’s AI feature or a custom model: what suits your service desk?

Routing rules in your software

Choose this if your portal already asks users to select from fixed options and each option maps to one assignment group. It’s easy to manage, but free-text descriptions fall outside these rules.

Your ITSM software’s AI feature

Choose this if your software has a feature like ServiceNow’s classification and your categories and groups are fairly standard. Even then, measure its performance on your own tickets.

A custom model alongside your software

Choose this if you use multiple software packages, want to identify duplicate tickets or need to take your own data into account. Our advice: start with the assignment group and add priority later.

We built classification models and a central service desk for Olympia and PostNL

Based in Amsterdam and a Databricks and Microsoft partner, Data Science Lab has been building data and AI applications since 2016. We built a classification model for Olympia and Dex, the self-learning chatbot, for PostNL.

Book an introductory meeting

A ticket export and your category tree are the starting point for our first conversation

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your point of contact. In our first conversation, we’ll look at your categories, your assignment groups and where tickets are currently getting stuck.

Where do most of your tickets go wrong?

Describe your service desk, the software you use and where tickets currently get stuck. We’ll reply by email, or arrange an initial conversation.

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

Ticket classification automatically assigns a category, priority and handling team to requests in your service desk. A model reads the description, looks at previous tickets and makes a suggestion, or routes the ticket if it is confident enough. This lets you classify and route tickets automatically, whether they are for IT, facilities or customer service.

Service desk AI like this is one of the AI applications we build.

E-mail classification sorts a mailbox: incoming messages get a label and are assigned to a team. Ticket classification works in an ITSM system, where a request is already a ticket with fields, a requester, a configuration item and a history. The model takes those fields into account, and the final resolution shows what the correct answer was.

If you are dealing with a shared mailbox, read about e-mail classification with AI.

Many service desks derive priority from impact and urgency: how many people or systems are affected, and how quickly the issue needs to be resolved. The model estimates both from the description and information such as the requester, department or system. Your priority matrix remains yours.

Keep it explainable: the agent can see which words and fields determined the priority.

By comparing new requests with those just received. A model converts each description into a sequence of numbers, called an embedding, so requests with the same meaning sit close together, even if they are worded differently. If many similar requests about the same system suddenly arrive, the coordinator receives an alert to link them to a single incident.

The ticket history from your system: the description, the category and handling team recorded when the ticket was resolved, its priority and, where available, the configuration item and resolution. The final values are what the model learns from. Comparing what the first-line team chose with what was changed later shows where things go wrong.

Tickets contain names and sometimes sensitive details; we mask these where the model does not need them. We first tidy up a disorganised category structure; see improving data quality.

Through the system’s API: the model reads a new ticket, writes its suggestion to a field or note, and the agent sees it on their usual screen. We check the supplier’s documentation to find out how this works for each system and version. We are not partners of these suppliers.

Ideally, the model runs on the platform you already use, such as Azure or Databricks.

By tracking how often agents change a suggestion, broken down by category and handling team. When a new system, a reorganisation or a new category is introduced, that number rises. We retrain the model using the corrections made during that period. We agree in advance who will monitor this: your own team or ours, through MLOps management.

When there are few requests, when a better portal form already guides users to the right choice, or when one team handles everything. In those cases, rules in your system are enough.

If you want employees’ questions answered directly rather than routed as tickets, a custom-built AI chatbot may be a better fit. If you want to analyse large amounts of text for topics and sentiment, read about NLP and text analysis.