Skip to content

AI agency in Tilburg

We build AI that classifies and summarises incoming text, testing it in short sprints with your team. That’s how WUA structures respondents’ open-ended answers.

We automated WUA’s respondent research with a topic classification model

“Close communication and detailed sprint reviews ensured the model met our needs exactly, resulting in a successful launch with our clients.” Yannick Ludwig, Head of Product, WUA

Read the case study

How much time do you spend reading and forwarding responses, emails and notifications?

Text comes in through forms, emails and surveys, and someone has to read and forward it. AI sorts it by topic, summarises it and sends it to the right person so your team can respond faster.

Schedule a conversation

Classify, summarise and respond to text, with human oversight

Discuss your challenge

Open-ended responses by topic

For WUA, we built a model that classifies respondents’ open-ended answers by topic.

Employees’ questions

For PostNL, we developed Dex, a self-learning chatbot that serves as a central point for employees’ questions.

Tone and sentiment

For the Respect Foundation, we built a social media monitoring tool with LLM sentiment analysis.

A human decides

If there’s uncertainty or the topic is sensitive, the text goes to a member of staff, along with the reason.

From sample texts to a model your team trusts, in short sprints

  1. Defining topics

    Work with your team to define the categories using real texts from your own inbox.

  2. Sprint reviews

    Compare the results with your team’s judgement in each sprint, until the model aligns with it.

  3. In your workflow

    Integrate the model into your email, ticketing or survey system, with checks for new topics.

Do we have enough text to build a model?

Find out in a conversation

A good fit if

You receive a lot of text every week, your team spends a lot of time reading and forwarding it, and you have a rough idea of which topics keep coming up.

Not a good fit if

You receive only a handful of messages a week. In that case, a good inbox rule or a form with selectable options is enough to start with.

What the scope depends on

How many channels and languages are involved, how detailed the categories need to be, which systems we connect and who maintains the categories afterwards.

Will our text go to an AI provider?

We decide that with your IT team and privacy officer beforehand. A model can run in your own environment, and personal data can be removed before processing.

Keywords, a trained model or a language model: which suits your texts?

Keyword-based rules

Choose this if a few words reveal the topic. Rules are quick to create, but miss anything phrased differently.

A trained model

Choose this if the topics are fixed and you have plenty of examples. Our advice: start here for large volumes. It is predictable and inexpensive to use.

A language model

Choose this if you also want to summarise or answer questions, or if the topics change often.

We’ve been building text-based AI for organisations across the Netherlands since 2016

Data Science Lab was founded in 2016 and is a Microsoft and Databricks partner. We’ve worked with WUA, PostNL and the Respect Foundation, and share what we learn at our Techdays.

Book an introductory meeting

The text you’re reviewing by hand now is what our first conversation is about

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your point of contact. In our first conversation, we’ll discuss what text comes in and what needs to be done with it.

Which stream of text would you like to start with?

Tell us where the text comes in and what your team does with it. 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

An AI company like Data Science Lab builds applications that take over or speed up a task, then maintains them. This page focuses mainly on text: categorising and summarising open-ended responses, emails, reports and customer feedback.

You can find all our applications under AI applications.

No. We work from Westerdoksdijk in Amsterdam. We work online with organisations in Tilburg and the rest of Brabant, and visit for workshops and sprint reviews.

None of the case studies here took place in Tilburg; they show what we build.

It uses a model to assign a text to one or more topics. For WUA, we used NLP, embeddings and machine learning to make open-ended responses immediately available to view by topic.

Read more about the technology at text analysis with NLP.

Yes. The model determines the topic and urgency, then routes the email to the right queue. If there is any doubt, it goes to a staff member.

Read how this works at email classification.

By comparing its results with the choices your own team made on texts they have already categorised. For WUA, we did this in sprint reviews until the model met their needs.

After launch, we continue to check samples.

You will see them in texts the model cannot categorise accurately. We collect those texts and work with your team to decide whether to add a topic. The model then learns to recognise it.

That way, the categories continue to reflect what actually comes in.

For text from customers or citizens, you often need a DPIA and agreements on where the model runs and what data is retained. Categorising and forwarding is usually not a high-risk application, but we assess this case by case.

Read more about the AI Act.

We agree on that in advance: us, your team, or us initially and your team later. This includes regular checks that the categories remain accurate and someone responsible for adding new topics.

Read how we keep models running in production at MLOps.