Problem
BlaBlaCar product managers need more data for experiments, market expansion and feature hypotheses, but Data Analyst capacity does not grow at the same pace. Requests wait in a queue for weeks, and hypotheses wait with them. A generic LLM can write valid SQL, yet it does not know the company's tables, and it does not know unwritten business rules. A query that runs can still count the wrong thing.
Example from the article: a trip must be counted using a complex geographic zoning rule. With only a schema, the LLM wrote code that ran but applied flawed business logic.
Solution in brief
The team gave the LLM two kinds of context in three steps: a JSON catalog of the BigQuery schema, then the last three months of analysts' queries as few-shot examples. They added three "zones of autonomy" that say when a PM may use a result alone and when an analyst must review it, plus SQL training for PMs.
Questions this case answers
- How to let non-technical staff query a warehouse through an LLM when the model does not know the schema or the unwritten business rules
- How to set trust rules for LLM-generated SQL so that high-stakes decisions still get expert review
- How to let product managers get data answers without waiting weeks in the analyst queue?
- How to make an LLM write SQL that matches our real tables and business rules, not just valid syntax?
Tools and methods
- Text-to-SQL assistant (LLM chat) — PMs describe a question in plain language and get a query plus an explanation.
- JSON schema catalog from BigQuery — tells the model which tables and columns exist.
- Few-shot examples from analysts' queries — teach the company's business logic.
- Three zones of autonomy — safe, risky, dead; decide who must check a result.
- SQL training for PMs — lets PMs read queries and spot hallucinations.
Google or a one-time email code. No password.
