Bringing 3,000 legal models into one conversation
Register Belastingadviseurs is replacing its outdated document search with an AI chat assistant that helps 5,500 members find the right model in plain language.
Register Belastingadviseurs (RB) is a professional association for tax advisors and lawyers. Its roughly 5,500 members get access to calculation models, standard agreements and legal models. We worked with RB's management, its specialist group for technical content and the ICT team.
Challenge
Members searched for documents through the search on RB's website. That search was outdated, with filters that did not help. Many documents exist in near-identical versions, and the search could not tell them apart. Calculation models often use only abbreviations, so a search for "private holding" missed relevant files. Experienced members knew the library almost by heart. Younger members often could not find documents at all. As a result, members turned to general AI tools and other databases, and use of the library kept declining.
Solution
Members describe their situation in one or two sentences. The system searches all documents in the background, suggests the best matches and asks follow-up questions when needed.
- Model cards: summarise each document's purpose, use cases and jargon, so full terms match abbreviations.
- Smart search with re-ranking: matches the question to the right documents and puts the best first.
- Chat interface: lets members search and refine in plain language.
- Fill-in function (in development): fills fields in legal models, for example from a pasted meeting transcript.
- By design: The assistent explains when a document applies, but never gives tax or legal advice.
Built around how members search
Organising documents by what they are for, not only by what they say.
- System design We summarised each document into a model card instead of splitting it into chunks, because many documents overlap heavily.
- User research We interviewed two members with different levels of experience.
- Retrieval We stored the model cards in a vector database and tested retrieval methods; a re-ranker worked best.
- Chat and integrations We built the chat interface and connected it to the RB website and AFAS, the document source.
- Fill-in function We are now building field recognition and auto-fill for legal models.
Jochem BoumanAI Engineer
Reimer KoopalAI Engineer
Roel RotteveelAI Project Manager & COOImpact
A working search and chat interface are already in place. Members are expected to find the right model in one or two tries instead of searching for ten minutes or more. Younger members get the same access as colleagues who know the library by heart. That should bring members back to RB's own documents. Lawyers are also expected to save time by filling in agreements straight from their notes.
In development The project is not live yet; testing starts once the fill-in function is finished.
More cases
At AdamI, we transform businesses by using the technology of artificial intelligence
