The problem
Business partners ask the same finance questions repeatedly, and the answers sit in dashboards they rarely open. The obvious AI answer, sending data to an external model, was off the table because of data governance.
The approach
- A deterministic retrieval-and-compute engine with conversational memory. It finds the right measure and dimension, runs the query, and states the number.
- No external LLM. Everything runs on internal infrastructure, so no enterprise data leaves.
- An integration seam designed up front, so an approved internal LLM can be dropped in later without reworking the application.
What it taught me
Constraints are a design input. A smaller, governed tool that people can actually use beats an impressive demo that Security will never approve. It also made me the natural Finance AI liaison: I now host sessions on AI adoption and its history and help peers find realistic use cases.