Signalwell · A natural-language analytics copilot
Added a natural-language query layer on top of a marketing analytics warehouse, so non-technical marketers can ask questions in plain English and get a chart back, not a support ticket for the data team.
01/ The challenge
Every "how did the April campaign perform by channel" question went into a queue for the one analyst who knew the warehouse schema. Turnaround was one to three days, by which point the marketing team had usually moved on.
02/ The approach
We built a query planner that translates plain-English questions into validated SQL against a locked-down read replica, with an automatic chart-type picker and a fallback to "ask the analyst" when confidence is low. Every generated query is logged and reviewable.
03/ The outcome
- Questions self-served
- 74%
- Avg answer time
- <10s
- Analyst queue volume
- -73%
74% of questions that used to go to the analyst queue are now answered directly, in under 10 seconds. The analyst's queue dropped from roughly 40 tickets a week to around 11, almost all genuinely novel questions.
“I used to wait days for a simple channel breakdown. Now I just ask and the chart is already there by the time I would have finished writing the ticket.”