Natural-Language Data Queries with LLMs
Asking questions in plain language makes BI easier, but a model can invent numbers. We explain why turning the question into a validated report definition is safer.

Typing 'in which region did margin fall last quarter?' into a box and getting an answer is appealing. Large language models (LLMs) make that possible; but when the subject is numbers, the problem is this: the model writes a fluent sentence, which does not mean it calculated the number correctly.
Two approaches
- Having the model write SQL: gives a quick prototype. But the same question can produce a different query on different days, a table or filter can be chosen wrongly, and it is hard to prove that the resulting number matches the dashboard.
- Having the model fill in a report definition: the model does not write free-form SQL; it only fills in a structured definition made of predefined measures and dimensions. A validator checks the definition, then a fixed compiler produces the SQL.
Why the second one
If the dashboard and the natural-language layer go through the same compiler, producing two different numbers becomes structurally impossible. Arithmetic is not left to the model; the calculation is done in deterministic code, and the model only interprets the result. If a measure that is not in the catalog is requested, the system raises an error instead of making something up.
Where Floemdata stands
This is the design approach of Floemdata Analytics: today the reports are produced from a defined catalog of measures and dimensions through a single compiler; the dashboard's company summary view is computed outside that path with its own query, and that the two paths give the same result has not been measured yet (not verified). The natural-language layer (an AI assistant) is on the roadmap; it is not a product feature yet (not verified).
Before you try natural-language BI
- Write down the 10 questions asked most often and match each with a report whose correct answer you know.
- Let the system ask or refuse when it is unsure; it should not guess.
- Show next to every answer which measure and filter were used.
- Keep data access limited per person; data the user cannot see must not leak into an answer.
In short: natural language is an interface, not a calculation engine. Turn the question into a validated definition and leave the number to code.