The problem
Even with dashboards, plenty of questions fall outside what anyone pre-built. Those went through the manual request process, and the round trip averaged about 25 hours.
What I built
An internal chatbot that answers business questions against company data directly, so the person asking does not need to know where the data lives or how to query it.
The hard problem here was never fluency. It was grounding. An analytics chatbot that is confidently wrong is strictly worse than no chatbot at all, because it produces numbers that look authoritative, travel into decks, and get acted on. A dashboard that is down is obviously down; a chatbot that is subtly wrong is invisible.
So most of the work went into two things:
- Retrieval. Making sure the model is looking at the right slice of data before it says anything.
- Constraint. Limiting what the model is allowed to assert, and making it decline rather than guess when the data does not support an answer.
Where it landed
Time-to-insight dropped from 25 hours to 3 minutes.
The number I actually watch, though, is how often it declines to answer. A system like this earns trust by being boring and reliable, not by being impressive, and the refusals are what make the answers worth something.