Articles Customer story

How the NJi built a safe AI guide with vragen.ai and grew its own AI knowledge along the way

The NJi built a safe AI guide with vragen.ai, and grew its in-house AI knowledge at the same time.

Lisette Vredenburg Product owner

6 min read

The NJi website with the assistant 'Vraag het NJi', showing an audience selector and an answer to a question from a young person

Why the NJi started with vragen.ai

The reason was clear: AI is developing at speed, and it goes straight to the heart of what a knowledge institute does. When knowledge is your core task, you want to stay in charge of how that knowledge is used and shared.

A demonstration of vragen.ai showed that this was possible. Thanks to a RAG solution (Retrieval-Augmented Generation) running on its own content, the NJi had no reason to worry about black box models or unclear data sources.

That built confidence, especially because vragen.ai demonstrably met their high requirements for security, privacy and data control. So the NJi could experiment with AI safely, without the risk of external models learning from their data.

Together with vragen.ai, the NJi developed its own AI search assistant: Vraag het NJi. A tool that answers questions from the institute's own knowledge, with sources cited and within clear safety boundaries. And one that helps visitors find their way through the sheer volume of information on the NJi website.

From reluctant to AI ambassador

Marleen van Riet, senior communications advisor at the NJi, was still hesitant about AI when the project started. She used it now and then, but did not consider herself an early adopter. The technology did not appeal to her straight away, and the terms and the way it worked were unfamiliar. "How does generative AI work? What is a language model?"

Working with the vragen.ai team changed that. Helping build the tool gave her insight into how to use AI responsibly, how to keep control of the content, and how much that matters inside a knowledge organization like the NJi.

"I understand much better now what AI is and how to use it responsibly. That has made me far more open to it, including to the ethical and social questions around it."

Marleen van Riet, senior communications advisor at the NJi

The experience brought more than understanding: it brought confidence. Even an external AI event had little new to offer by then, because the level of AI knowledge inside the NJi had risen considerably.

The assistant as a guide through an abundance of information

Knowledge is the NJi's core product. They want to keep interpreting, sharing and managing it in a way that fits their editorial standards, with one continuous goal: helping every audience find the right answer faster and more precisely.

That underlying knowledge has been built up carefully, rests on research and holds up under scrutiny. Which is exactly why reliability is one of the biggest concerns. If the AI assistant says something wrong, it reflects directly on the NJi's credibility. That was the largest worry internally.

So they chose a careful approach. Not a free-roaming chatbot that holds a conversation, but not a bare search engine either. Vraag het NJi deliberately sits in between. In some cases the assistant gives short answers, and above all it points precisely to reliable sources on the website. That keeps the content within the NJi's boundaries.

Insight per audience, based on real questions

The NJi shares its knowledge by audience: professionals, policymakers, parents, young people and education. Originally the focus was mainly on professionals and policy. During the pandemic it became clear that parents also had a strong need for reliable information. That audience has been part of the offering ever since. Young people are increasingly getting their own place in it too.

That has consequences for how the knowledge is opened up. On the website, visitors now see content that matches their profile. The same line runs through the AI assistant. Each audience has its own configuration: in tone, in content and in what information is relevant. Users pick their perspective, and the assistant tunes its answers to it.

At the same time the NJi wants more insight into what audiences are really looking for. To make that possible, an extra data layer was added. The vragen.ai dashboard now shows which questions come from which audience. That was not possible before.

That allows for more targeted monitoring and analysis. Each audience has its own group of staff keeping an eye on the questions, and the configuration is adjusted where needed. The picture of what is going on, and where knowledge is missing, gets sharper with every round.

Systems set up per audience for the NJi through the vragen.ai dashboard

Set up for safety and reliability

The AI assistant has been configured carefully, with attention to safety, privacy and reliability. On sensitive subjects such as depression, suicide or personal cases, the assistant follows extra guidelines. It does not engage with the substance, but points directly to appropriate help or the contact page. Personal data is anonymized automatically.

News items are excluded, because they date quickly. The AI can find specialist databases such as guidelines, instruments and statistics, but it doesn't reproduce their contents. Instead the assistant refers to the source. That keeps the context intact and avoids the risks that come with interpretation.

For every subject and source, they looked at what role the AI is allowed to play: show it or not, interpret it or not, always with control.

What vragen.ai makes visible

"Keeping our knowledge up to date is our core task. This is a flywheel that can drive that."

Marleen van Riet, senior communications advisor at the NJi

During the test phase it became clear that not all knowledge was equally current. With statistics especially, the AI sometimes fell back on older sources, simply because no more recent data was available.

That was a reason to look more closely at how current the content was. The assistant shows immediately where information lags behind or is missing. For the communications team that makes it easier to raise the alarm inside the organization. It helps them decide which subject areas need updating first, and improve content faster, based on real questions from users.

From scenarios to going live

In the beginning the emphasis was on testing. Plenty of "what if" situations were explored, with an eye on risks and exceptions. What if someone enters a personal case? What if a sensitive subject comes up? What if data turns out to be outdated?

"We got stuck in the hypothetical scenarios. It became a swamp of what-if questions. That was the moment to say to each other: it's time to go live."

Marleen van Riet, senior communications advisor at the NJi

There comes a point where you can't keep testing forever. With a clear mandate from the board, the NJi dared to take that step: go live, and learn from practice.

The NJi now has an AI search assistant under its own direction. It makes reliable NJi knowledge easier to find and delivers valuable insight per audience. At the same time it speeds up the development of AI skills inside the organization.

What the NJi would pass on to other organizations

See it as a strategic experiment, but keep hold of the reins. And think about monitoring up front: a live assistant demands something of your organization too.

"Make sure the right people from the organization are on board and testing with you. We really needed the subject-matter people to be able to judge the answers. Then don't get bogged down endlessly testing what-if scenarios. In the end, real users give you the best picture of the questions that actually get asked.

Set up your monitoring properly, with clear follow-up, management and ownership. That way you learn what is going on, how the tool performs, and where you as an organization have work to do."

Marleen van Riet, senior communications advisor at the NJi

Looking ahead: step by step

Over the coming period the NJi wants to keep building. With better audience analysis based on real questions. With structural improvements to how current their content is.

The assistant is now part of the website menu. This is still a pilot phase, in which the NJi tests with real users how the tool performs and what else is needed. An evaluation follows in January to decide how to continue.

In time the assistant might replace the search function entirely. That is not on the table yet, but the foundation is there and the process is iterative. Every step brings new insight.

Want your own AI search assistant?

Want to experiment safely with AI on your own knowledge, like the NJi, with full control over sources, answers and audiences? vragen.ai helps you with configuration, optimization and monitoring.