Business people at Merck were spending 20–30 minutes navigating BI dashboards to answer questions that should take two. We replaced that with a conversational analytics layer built on Cortex Analyst and Semantic Views – integrated in a custom-built enterprise chatbot on a FastAPI and Next.js stack. We will cover the full architecture: how Semantic Views encode business definitions for accurate SQL generation, how governed access works across global omnichannel data, and how Cortex Analyst fits into a production system with custom LLM routing and other external APIs and full-observability with zero-trust. We’ll share what moved the needle – including how we quantified a 10x reduction in time-to-answer and what we’d build differently next time.
Before replacing a dashboard with a chatbot
For a team considering a similar change, these are useful questions to take into scoping:
- Which recurring business question takes too long to answer, and which decision depends on it?
- Do people disagree about the metric, or struggle to find a metric whose definition is already agreed?
- Which datasets, regions and permissions must each answer respect?
- How will a reviewer check the answer against its source and an agreed reference result?
The distinction matters: faster access cannot resolve conflicting business definitions. A bounded set of questions and controlled metrics gives you something concrete to evaluate before expanding the interface.
For the data-access approach, see talk to your data and semantic layers. If you need to decide where an intervention would help, start with a company audit.