Noor: five care organisations build one AI intake assistant together

Anyone in Groningen who needs social support ends up in a chain of organisations that each have their own intake route, opening hours and criteria. For someone in a vulnerable situation, finding the right door is already a task in itself, and the question rarely comes during office hours. Five large care organisations in the region therefore decided to build one digital front door together: Noor. Below you can read what Noor does, how reliability is built in, how five organisations run one assistant together and what other organisations can learn from this approach.

Table of contents

1. Why applying for social support is complicated
2. What Noor does
3. The route is fixed, the AI holds the conversation
4. Five organisations, one knowledge base
5. What the test phase has to show
6. What other organisations can learn from this
7. How we can help
8. Frequently asked questions

Why applying for social support is complicated

Social support is, in practice, chain work: shelter, guidance, addiction care, work and daytime activities sit with different organisations, each with its own intake. Which organisation fits depends on the situation: age, housing situation, whatever else is going on. That information is scattered across websites and leaflets, and the people it concerns are rarely in a position to calmly put that puzzle together.

The result: people apply to the wrong organisation, drop out, or only come into view once the situation has worsened. And the professionals in the chain spend part of their time referring people who should have knocked on a different door.

What Noor does

Noor is an AI intake assistant, reachable as a chat, that does three things which otherwise wait until a desk opens:

• Clarify the question. Noor asks questions in plain language and in the language of the person seeking help, and builds up a picture of the situation step by step.
• Prepare the application. The relevant information is organised, so the intake at the organisation afterwards is faster and more complete.
• Refer to the right organisation. Based on the situation, the person seeking help ends up at the chain organisation that fits, with concrete next steps, right down to a call script or an application email.

For information questions (opening hours, conditions, what to bring), Noor draws on a shared knowledge base and cites the source. If the answer is not in there, Noor says so honestly, instead of inventing something plausible. For this audience, that is no detail: a made-up opening time means someone standing in front of a closed door.

The route is fixed, the AI holds the conversation

The most important design choice behind Noor: the language model determines the tone and wording of the conversation, but the reliability sits in layers around it that always behave the same way.

• The referral follows a decision tree that the care organisations maintain themselves in a dashboard. Which situation leads to which organisation is a choice made by the professionals in the chain, laid down in rules that work the same way for every person seeking help. The AI is not allowed to deviate from that; every piece of advice Noor gives is checked against those rules before it reaches the person seeking help.
• Safety comes before everything. Signals of acute need, such as being without shelter tonight, violence or suicidal thoughts, are recognised separately. In those situations, the person seeking help is shown the safe route straight away, right up to 112 and Veilig Thuis, whatever else happens in the conversation. Such a safety signal also stays active for the rest of the conversation: an unnecessarily cautious response is a small inconvenience, a missed crisis is irreversible.
• Every step is recorded. For every conversation, it can be traced which information was used and how a referral came about. That makes the system auditable for the consortium, and adjustable when practice calls for it.

This setup is also the lesson we draw from the project: with AI for vulnerable groups, you do not want to ask the model for good behaviour, you want to enforce it with checks around it.

Five organisations, one knowledge base

Noor is not the product of one organisation. Five large care organisations in the Groningen region signed a cooperation agreement in the spring of 2026 and run the assistant together. The knowledge Noor draws on comes from the professionals of those organisations, and a shared management group keeps the content up to date and safeguards the quality of the customer journeys.

That sounds organisational, and that is exactly the point: most of the work on a good AI assistant is not technology. It is five organisations agreeing on routes, wording and responsibilities, and keeping that maintained. The technology then makes those agreements executable every day, day and night.

What the test phase has to show

Noor is currently being tested with staff and experts by experience from the participating organisations. They test the assistant on the questions that really matter: is the referral correct for situations from practice, is the tone appropriate for people in a difficult period, and does the assistant do the right thing when there are signals of need. What comes out of that goes straight back into the system: the decision tree, the knowledge base and the wording are adjusted based on the test rounds before Noor becomes more widely available.

In the meantime, the test phase delivers two things that were not the goal beforehand. The first is a nice bonus: staff use Noor to see how things work at the other organisations in the chain. The shared knowledge base turns out to be an internal reference work too, and that saves a lot of phone calls. The second is perhaps the most valuable finding: the questions Noor cannot answer make visible where the gaps in the shared knowledge and services are, and where there is overlap between organisations, and whether that overlap is actually smart. That is steering information about the chain itself, which was not there without Noor.

That order is deliberate: first have it tested by the people who know the audience best, only then scale up.

What other organisations can learn from this

Three lessons from this project are more broadly useful than social support:

1. Put the decisions with people, the execution with AI. A decision tree or rule set that professionals manage themselves gives AI behaviour that is explainable and adjustable. That is exactly what regulators and the EU AI Act ask of AI in the social domain.
2. Build in honesty. An assistant that only answers based on a managed knowledge base, with source references and an honest "I do not know", is the safe choice for any organisation with a public-facing role.
3. Do not underestimate the collaboration part. Where several departments or organisations share one assistant, the real work lies in joint management of knowledge and routes. Arrange that governance from day one.
4. An AI assistant is also a measuring instrument. The questions your assistant cannot answer are steering information: they show where knowledge is missing, where services overlap and whether that overlap is smart. You would not have that picture of your own organisation or chain without the assistant. Arrange from the start that someone reads those signals and does something with them.

You can read more about what is being built in the region in our overview AI in Groningen: what is happening and where you start as a business.

How we can help

We build the technology behind Noor, from our office on the Zernike Campus. Do you want to know where your organisation stands? Take the free AI scan: in a few minutes you see which phase of AI maturity you are in and what a logical next step is.

Feel free to get in touch for a no-obligation conversation. We first discuss what fits your organisation; only then does something go on paper.

Frequently asked questions