We’re building a bot

I held off on this longer than I probably should have.

The case for an AI-facing interface has been obvious for a while. The case against it (hallucination risk, data security, governance exposure) was obvious too. What moved me was straightforward: I've spent years building tools for power users. Direct data access, SQL literacy, self-service dashboards. That capability serves maybe 10% of the organization. The other 90% still need answers. They're also my customers.

Snowflake Cortex gives me enough control to take the risk seriously. Three things that brought me around:

Verified questions. Cortex Analyst lets me map critical questions to specific SQL queries. When someone asks something similar, it pulls from that verified set to guide the answer. That's not a complete hallucination solution, but it creates institutional checkpoints on the questions that actually matter, and it keeps the most critical queries from drifting.

The walled garden holds. Whether we're routing through the SQL engine or RAG, the data security model stays intact. In a regulated environment, that's non-negotiable. I'm not exposing anything I wouldn't expose through a standard BI tool.

Power user development doesn't reach everyone. It's not supposed to. Some people just need to ask a question and get a straight answer. That's a legitimate use case and I've been underserving it.

Our alpha is one small table. It gives me a controlled scope with low stakes, and we get to build piece-by-piece through the summer. Modest expansion for beta when we're comfortable with what we're seeing.

What that expansion looks like is still open. If Cortex Analyst runs cleanly on structured data, we keep going that direction and work through our internal schemas. If it becomes a lift, we pivot to Cortex Search and run a semi-structured or unstructured test instead. The alpha tells us which problem we're actually solving before we commit to anything.