A reality check on building RAG pipelines in ERP contexts, including what works, what doesn't, and what we might do differently next time.
RAG is often sold as the future of enterprise AI: combine LLMs with vector search, and suddenly users can “talk” to their data. It sounds great, but what happens when you actually try to implement it in a structured ERP system like Business Central?
In this session, I'll share what I've learned from developing a RAG-based prototype inside Business Central. The initial idea was simple: let users select tables and fields that should be semantically searchable, embed their values into a vector index, and connect them to a conversational AI assistant. In practice, the results were mixed. Some things worked well (especially with Copilot assistance in development). Other things, like whether this even needed vectors at all, are still up for debate.
You'll learn:
- How we built a configurable RAG prototype on top of Business Central
- Why semantically matching JSON-structured ERP fields sometimes feels like the wrong abstraction
- What to consider before jumping into embeddings and vector search
- What I would change if I started again today
This session is for technical builders who want more than just AI hype and are curious what happens when you move from cool demos to production constraints, user expectations, and actual data complexity in ERP.