Natural language insights from an operating company's data warehouse

A multi-location operating company had its performance data in one warehouse, but getting an answer meant a request to the reporting team, a wait of about a week, and a static report that covered only the question as first asked. Follow-up questions restarted the cycle, and the engineers who ran the warehouse spent their time answering questions instead of improving it. New Clarity built a conversational system that turns a manager's question into a database query using the company's own definitions of revenue, margin and performance, returns the figure, a table or a chart from the same result, carries context across follow-up questions, and restricts every user to the data their role allows. Time to an answer fell from about a week to minutes.

Leaders asked questions while working through a decision, and by the time the answer came back the decision had often already been made. One-off analysis meant exporting to spreadsheets and combining datasets by hand. A business intelligence tool was in place, but it answered predefined questions, and every new question was a request to the reporting team. The company needed managers to explore the data themselves without opening the warehouse to everyone.

The system generates a database query from a plain-English question using the warehouse schema, the company's own definitions of its key measures, and a library of roughly fifty real business questions paired with validated queries and expected answers, so a question about revenue or margin is answered the way the company defines those terms rather than the way a model assumes. It keeps context across a conversation, so a manager can move from a headline figure to the breakdown behind it without restating the question. Results come back as a direct figure, a table or a chart generated from the same query, and can be exported for further work.

• Question to query. The query is generated against the warehouse schema with the company's metric definitions supplied alongside it, and the library of validated question-and-query pairs serves as worked examples for each new question.
• Conversation memory. Follow-up questions resolve against the previous query.
• One result, three views. Figure, table and chart come from the same result set, and any can be exported.
• Access by role. Each role is tied to the datasets it may see, applied when the query runs, not in the interface.
• Traceability. Every answer returns the query that produced it, so a figure can be checked against its source.
• Improvement loop. Real questions, corrections and feedback are recorded, and accuracy improved mainly by refining how the workflow interprets the company's questions rather than by changing the model.

Roles restrict what each user can query, so wide adoption did not mean wide access. Every answer carries its query, so a figure is checked against its source rather than taken on trust. The workflow runs inside the company's own environment.

Managers explore performance data themselves, most manual queries and spreadsheet analysis are gone, and the reporting team is no longer interrupted by routine requests. The people who gained most were the operating managers who had frequent questions and no direct access to the data.