Ask plain-English questions about pipeline, bookings, rep performance, and churn · Powered by Claude
Grounding: The server-side system prompt contains a flattened text summary of the synthetic GTM dataset (deals by rep × specialty × source × quarter, the marketing funnel, churn cohorts). Claude reasons over the summary directly — no tool use, no SQL generation, no vector retrieval. The Smart Insights tiles above use the same seeded dataset client-side so the panel and the assistant stay aligned.
In production: this would be a tool-using agent over a governed metric registry (the kind documented in NorthStar). The assistant would issue queries against live warehouse tables; the metric definitions in the registry would constrain what aggregations and joins it can perform, so two analysts asking the same question always get the same answer. Text prompt-stuffing is a demo simplification, not the recommended pattern.
Limits: Synthetic data only — questions outside the dataset (e.g. "what's our YoY ARR?") will get a polite "I don't have that data" response. Rate-limited to 10 requests per IP per minute. The system prompt lives server-side and never accepts client-supplied overrides.