What 4,500 questions taught Werken voor Nederland about its own site

Werken voor Nederland draws more than 12 million visitors a year and handles some 300,000 applications. Even so, the team ran into a familiar problem. The information was usually on the site, but visitors could not always find it.

An ordinary search field works well mainly when you know which term you need. Someone wanting to know whether a college degree is enough for a particular traineeship does not always think in the words the website uses.

That question is where we started, and we set up vragen.ai together with O&P Rijk. We took care of the technology and the configuration. O&P Rijk brought the content, the boundaries and their knowledge of the audience. Step by step, that produced something that fits the organization and its visitors alike.

Only your own content as the source

vragen.ai works with RAG, or retrieval augmented generation. The agent looks for relevant information in content the organization manages itself and writes an answer on that basis. The sources it used are shown alongside.

The wider internet is not used as a source. How that works exactly is on How it works and on the page about AI site search.

“We only wanted our own, reliable content as the source.”

With information about terms of employment, say, you want to be able to trust the source. That is why vragen.ai only uses the sources designated in advance.

It also means current content stays important. The crawler checks the site again regularly, so a change to the content is reflected in new answers within an hour or two.

Meanwhile the ordinary search field simply stayed. vragen.ai sits next to it, so the visitor chooses how to find information.

More than 4,500 questions in ten weeks

After half a year of testing in a walled-off environment, the agent went live on 15 June. More than 4,500 questions have been asked since.

• 4,500+ questions since going live on 15 June 2026. Source: Werf& (in Dutch)
• 75 to 80% of candidates are happy with the answer they received. Source: Werf& (in Dutch)
• 70 to 80% of questions are about vacancies. 5 to 10 percent are about the speculative application. Source: Werf& (in Dutch)

So one in ten to twenty candidates asks about the option of a speculative application. The answer differs per participating organization. With a question like that, a conversation helps more than a list of search results.

The questions without an answer point your editors somewhere

If vragen.ai may only answer from your own content, that draws a clear line as well. When the answer is not in the available sources, the agent says so.

And that is exactly where new value appears.

During the tests at Werken voor Nederland, questions came in about internal mobility and career coaching, for example. There was no public content about either yet. So the agent could not answer, but in doing so it made visible what visitors need.

According to Schoevaars, his team had never had such direct insight into what visitors are missing in the existing content. Before, that information came from experience, interviews and data, among other things. Now visitors ask the questions themselves.

That is what we built Analysis for inside vragen.ai. Questions are grouped into themes automatically and can be followed in a dashboard you configure yourself. Anyone who wants to analyze further can also export the inbox.

So vragen.ai does more than make existing information easier to find. The questions also show where content can be improved or added.

What came before going live

Werken voor Nederland is the first RAG application of the Dutch central government to answer questions in public. A careful process came before that.

For half a year the application ran in a walled-off environment. Various stakeholders were involved, and together we went through more than 900 test questions. Among other things we looked at reliability, tone and safety.

The system was also tested from the outside. In September 2025 an independent ethical hacker tried to break it, under the Cyber Security Pen Test certification of the Dutch CCV. No critical risks came out of that. The two improvement points that were found we fixed straight away.

We wrote about that process earlier in How vragen.ai performs under pressure from an ethical hacker. At the time we could not yet name the client.

The feature for anonymizing personal data came out of this process too. Anyone asking questions about a new job soon shares personal information. Think of a name, a home town or the situation at home. If a candidate enters details like that, they are removed automatically before the conversation is stored. More about how that works and where its limits lie is on our security page.

The team also tested the model on VLAM, the platform of secure local AI models that SSC-ICT developed for the Dutch government. There, European open-source language models run on its own servers and the data stays inside the Netherlands.

We wrote earlier about why we think that development matters in A sovereign AI platform. More about our own infrastructure and the data we store is on Hosting and data storage.

What other careers sites can learn from this

Three careers sites now run on vragen.ai, and we see a similar pattern there.

Visitors ask about more than job titles or vacancies. They want to know how long their commute would be, whether a particular diploma is enough, and what it is like to work at an organization. The answers are sometimes scattered across several pages. And sometimes they are not on the site at all yet.

That is exactly where a conversation helps. The visitor does not have to work out first where the information sits or which search term goes with it. They just ask.

More about this application is on vragen.ai for careers sites.

One important question we cannot answer yet: does vragen.ai lead to more applications?

At Werken voor Nederland that cannot be measured directly at the moment, because applications go to the participating organizations themselves. An evaluation follows at the end of October.

What is visible already is that candidates keep asking. They use the conversation to find out more about a vacancy and about the organization. That gives candidates better access to the information they need, and at the same time gives the organization a clearer view of what matters to them.

Want to see what your visitors would ask?

Not every project has to start with half a year of testing in a sandbox. With vragen.ai you can build a demo environment based on your own careers site. That shows you the same day which questions the agent can answer, which sources it uses for them, and where information is still missing.

Want to know more? Have a look at vragen.ai for careers sites, or read more about our pricing.