Problem
A data vendor lets prospects trial its datasets before buying. S&P Global Market Intelligence has run roughly 300 to 400 such trials. The exploratory data analysis (EDA) step is where they stall: the client must write SQL or Python, learn a normalized schema (some datasets have 10 to 60 tables) and decode the vendor's terminology. Business users wait for scarce engineers, and the first real insight takes days, which delays the purchase decision.
Example from the article: a client asks for IBM's quarterly financials for 2025, and plain Genie returns a list of over 1,000 matching entities instead of one company.
Solution in brief
Each client gets a private Databricks workspace with a Genie room per dataset, so the first contact with the data is a plain-language question. Rooms are generated from an "AI bundle" per dataset (readme, SQL examples, YAML semantic layer, context file). A registered company-search function resolves a name to one key that Genie reuses for the whole conversation.
Questions this case answers
- How to let non-technical users explore an unfamiliar multi-table dataset in natural language and resolve ambiguous entity names before querying
- How to shorten the trial evaluation phase of a data product so that prospects reach their first insight in hours instead of days
- How to let non-technical users explore a large, unfamiliar dataset in natural language without writing SQL
- How to make text-to-SQL resolve ambiguous entity names such as company names
Tools and methods
- Databricks Genie rooms — natural-language querying with curated answers, tables and charts
- AI bundle per dataset — readme, SQL queries, dataset YAML (semantic layer), context file
- Registered entity-resolution function (
company_search) — maps a company name to the closest match and its key - Flattened (denormalized) tables — fewer joins for text-to-SQL
- Databricks notebooks — technical users validate and extend the same analysis
- Unity Catalog — governance and access
Google or a one-time email code. No password.
