University of Amsterdam: How We Use AI to Teach Students Legal Reasoning
AI offers enormous opportunities for education, but also a risk: if you just read the answer, you miss the thought process behind it. The challenge was to build an AI solution that actually helps students understand that thinking process.
From proof of concept to learning experience
Together with the lecturers, we developed a proof of concept in which students can use an AI agent to ask questions about real statements made by rechtspraak.nl. Not a generic chatbot, but a system that works with a carefully selected dataset of five themes, compiled by the lecturer.
This is how students don't learn alone what is a statement, but especially why that's how it was done. The AI thus becomes a tool for reasoning, not copying.
The core of the solution
We've built an AI agent that can not only search legal material, but also understands what the texts are about. The selected statements from rechtspraak.nl were therefore technically prepared and stored in Azure AI Search, so that the AI can find them by meaning rather than just by keyword.
Via Azure OpenAI the chatbot can then formulate answers that refer to the right sources and summarize the content clearly. This way, students can ask questions in a natural way and gain insight into the underlying arguments behind real statements.
Data and architecture
The basis lies in the rechtspraak.nl API, with which we retrieve the statements automatically. Through a AI Search pipeline the texts are split into smaller parts that are suitable for the AI model, then vectorized and stored in the AI Search index.
The AI agent can then search semantically and combines the statements found with the student's question. The result is a substantive answer with substantiation and source references.
The challenge: legal precision without training
We did not train the model, but iteratively refined the results with prompt engineering. Together with the instructor, we tested five models for accuracy, then seven variants of prompts and multiple search accuracy settings. This is how we found the right balance between factual precision and legal nuance.
The teacher tested the setup with her course material. When everything was correct, we put the system live for use in the lectures.
A concrete example
A student asks: “When can an employer fire an employee immediately?”
The AI searches the collected statements, recognizes the cases related to this topic and shows the relevant case law. This way, the student receives not one ready-made answer, but an overview of similar issues and their substantiation.
By comparing these statements, the student learns how judges reach their verdict and which arguments play a role in them. For example, the AI agent helps students to reason for themselves, not by giving answers, but by providing the right examples.
Technical choices and learnings
We chose Microsoft Azure, because it fits perfectly with the infrastructure of EduGen AI, the educational platform that integrates the PoC. Azure offers mature tooling for vector search, chunking, and RAG, allowing us to quickly experiment with different models and configurations.
The most important lesson? Fast iteration works better than looking for perfection. Testing five variants in parallel provides more insight than spending weeks tinkering with one prompt.
Learn to reason together
What made this project special is the collaboration between technology and content.
Albert-Jan:
“In a short time, together with a content expert, we created something amazing. Not about technology for the sake of technology, but about showing how AI can strengthen legal thinking.”
A great example of what we at Blis mean by AI-first work: technology that does not replace, but deepens. And it helps people to reason better, precisely because they work together with AI.
Want to know more about our AI-first projects? Then check out our AI Solutions and Cases page.
