Conversational BI

Conversational BI (conversational business intelligence) is a way to explore data by asking questions in natural language, just as you would write to a person: "How much did we sell in Greater Buenos Aires last month?" or "Show me the margin by branch." Instead of building a report or asking someone in IT for it, you type (or dictate) the question and the system returns the figure, table or chart that answers it, along with a readable explanation.

Under the hood, a language model interprets your intent, translates it into a query on your data (see text-to-SQL) and runs it against a semantic layer that knows what "sale," "margin" or "active customer" means at your company. That layer is what keeps two identical questions from returning different numbers. Conversational BI is part of what Metrix organizes: the analytics that structure your data so anyone on the team can query it.

Unlike the classic dashboard (fixed panels someone designed in advance), here the question is in charge: you explore freely, dig deeper on the fly and do not depend on the exact report you need already existing.

Why it appeared

For years, access to data had a bottleneck: the business had the questions and the technical team had the tools. A sales manager wanted to know why revenue dropped in the Cuyo region, opened a ticket, waited two days and got a report that almost answered the question, but not quite. That friction kills a data culture: people stop asking. Conversational BI attacks exactly that bottleneck by putting data queries within reach of anyone who knows how to frame a business question, with no intermediaries and no SQL.

How it works, step by step

The path of a question follows a fairly consistent sequence across platforms:

1. You type the question in natural language ("What is the average order value by channel this quarter?").
2. The system interprets it with a language model, which detects the metric (average order value), the dimension (channel) and the time filter (this quarter).
3. It resolves those terms against the semantic layer: the business dictionary that defines how "average order value" is calculated and which tables represent each thing.
4. It generates the query (the text-to-SQL process) and runs it on the data warehouse or the connected source.
5. It returns the result in the most readable format: a figure, a table, a chart and, often, a summary in words.

The least visible and most important piece is the third one. Without a solid semantic layer, the model "guesses" which table to use and ends up inventing relationships that do not exist (a real risk of hallucination in these systems). With it, the answers are consistent and auditable.

Conversational BI vs the traditional dashboard

They are not enemies: they complement each other. The dashboard is ideal for the metrics you look at every day (always the same ones); conversational BI shines when a new question comes up that no panel anticipated. The difference is clear here:

In practice, a mature company uses dashboards to monitor and conversational BI to investigate when a number on the dashboard goes out of the normal range.

A concrete example (Argentina)

A consumer goods distributor with warehouses in Buenos Aires, Rosario and Mendoza had been looking at a weekly sales panel. One week, the regional manager notices a drop in Mendoza and, instead of requesting a report, types: "Compare this week's sales in Mendoza against the same period last year, broken down by category." Within seconds he finds that the drop is concentrated in beverages and that it coincides with a stockout of a key brand. He follows up: "Which beverage SKUs had a stockout in Mendoza this week?" and puts together the replenishment order that same afternoon. Without conversational BI, that chain of three questions would have been three tickets and several days.

Common mistakes when adopting it

• Skipping the semantic layer. Connecting the tool directly to raw tables and expecting magic: without business definitions, the answers are inconsistent and dangerous for decision-making.
• Trusting it blindly. Conversational BI is an assistant, not an oracle. It should show the query it generated and the source, so you can verify (data governance before speed).
• Dirty data at the base. If your data quality is poor, natural language only speeds up the path to a wrong conclusion.
• Thinking of it as a full replacement for analysis. It handles the bulk of operational questions, but in-depth analysis by a data scientist still has its place.

Why it matters for a business

The value is not technological, it is cultural: it democratizes access to data. When asking costs nothing, people ask more, decide with evidence and stop operating on intuition. That is the promise of self-service analytics, of which conversational BI is the most natural expression. The condition for it to work is always the same: an orderly database and a single definition of each metric, so that the fast answer is also the right answer.