Questions and answers for the press
Last updated: September 11, 2026
Writing about vragen.ai, or about AI on the website of an organization? Below are the questions journalists ask us, with our answer and a link to the place on the site where it is substantiated. For a response, a quote or a conversation with someone on the team, email info@swis.nl and mention your deadline.
What is vragen.ai?
Vragen.ai is an AI search function for websites. A visitor types a question in their own words and gets one answer from the organization's own content, with the source attached. It sits on the existing website as a widget: a search bar, a chat window or a button.
Organizations use it to help visitors who cannot find something get an answer anyway, and to see what people are asking. What it does exactly is on the product page.
Who is behind it?
Vragen.ai is a product of SWIS, a digital agency in Leiden, the Netherlands, that has built websites and web applications for more than 25 years. The vragen.ai team is eight people working from that office; you can meet them on the about page. SWIS is certified to ISO 27001, NEN 7510 and ISO 9001.
How is an answer produced?
Search first, answer second. Vragen.ai looks up the visitor's question in the sources the organization has designated: web pages, documents and connected systems. Only what it finds there is used by the language model to write the answer. The source appears under the answer, so the reader can check it.
The technique is called Retrieval-Augmented Generation (RAG). We explain it step by step on How it works.
Can the AI make things up?
Vragen.ai may only draw on the organization's sources and has no access to the internet. If the answer is nowhere in those sources, the assistant says so and refers the visitor on, for example to a phone number or a contact form.
That is a design choice, not a guarantee that a language model never summarizes something incorrectly. That is why every answer shows its source, and why an organization tests the answers in a separate environment before anything goes live. Afterwards every step can be reviewed: which sources were used and how the answer was built. In September 2025 an independent ethical hacker tried to throw vragen.ai off balance.
What happens to visitor data?
As little as possible. Vragen.ai does not ask for a name, an email address or an account, and sets no cookies until someone starts a conversation themselves. From that moment there are three functional cookies with no personal data in them. If a visitor types in something personal anyway, the system filters it out and deletes the original after one hour. That recognition is automated and does not catch everything with certainty; the security page describes it as a residual risk, with the safeguard next to it.
Conversations are not used to train AI models. The organization sets how long the anonymized conversations are kept. All the facts a privacy officer needs, including the data processing agreement and the input for a DPIA, are on Security and data protection, with no form in between.
Where is the data stored and which AI model is used?
The data is stored in the Netherlands, with hosting partner TrueFullstaq in Almere; management happens from Leiden. The language model is a setting. The whole chain can run in the Netherlands, with large open models at Nebul in Dutch data centers. An organization can also choose a model from OpenAI, Anthropic or Mistral, or a model through Microsoft Azure; that stays within the European Economic Area and under a data processing agreement.
So who is in the chain depends on the customer's choice and is listed in the processing register. What that chain looks like is in the article on language models in the Netherlands.
How does vragen.ai relate to the EU AI Act?
Vragen.ai informs and points the way; it makes no decisions about people. A system that informs falls under the AI Act's limited-risk category. What the Act then asks for is mainly transparency: a visitor must know they are talking to AI (Article 50, in force since August 2, 2026), and a human must remain reachable. The latter is built in, because the assistant refers visitors on as soon as it does not have the answer.
For that notice, vragen.ai supplies the controls. The organization sets per integration whether and where the disclaimer appears, at the input field and with the answer, and at what moment it shows. By default it is on with the answer. On our own site you see the notice with every widget.
Responsibility therefore rests with the organization that uses vragen.ai: under the AI Act it is the deployer, and it decides on the notice and on registration in the Dutch Algorithm Register. We supply the settings and the information needed for that. The frameworks are laid out on the page for government, the transparency obligation in its own article and the setting itself in the documentation.
How do you help an organization get started?
In four steps, usually without any technical work. We start with the website address: vragen.ai reads the pages and builds a first knowledge base from them. The organization's editors test in their own environment with real visitor questions, adjust tone and sources, and rate the answers. Only when they are satisfied does the widget go live: two lines of code on the website.
Our team helps with the content during that phase. Afterwards the organization sees in the admin environment what visitors ask and where answers can improve. The guide is public in the documentation.
Who already works with vragen.ai?
Among others VNG (the Association of Dutch Municipalities), ArboNed, NHG, QSTA, the Netherlands Youth Institute and VluchtelingenWerk (the Dutch Council for Refugees). The European Data Protection Board uses the API and the semantic search function. The stories of the people who manage it are among the articles.
What does it cost?
Prices are public on the pricing page: a monthly subscription, with what each plan includes. For an article we would rather you cite that page, so the amount is correct at the moment you publish.
Is the widget accessible?
The widget meets the WCAG accessibility guidelines, so a visitor using a screen reader or only a keyboard can ask the same question. How we test and maintain that is in the article on accessibility.
How can I check this myself?
Ask the widget on Try it: it runs on our own content. The documentation is public, as are the certificates, the model data processing agreement and the DPIA input on the security page. If you see something that does not match what we say here, we want to hear it.
Who do I contact?
Email info@swis.nl. Mention your deadline and the topic and we will make sure the right colleague responds. Logos and screenshots are available on request.