Conversational analytics

What is conversational analytics?

Conversational analytics means you ask a question about your business in ordinary words and you get an answer a person can use: a number, a chart and two sentences that say what the number means. You do that inside your BI tool, inside a chat assistant, or inside Teams. Nobody opens a report editor, nobody writes a query.

The picture to keep in mind is the owner of a fifty-person company on the train at seven in the morning, typing "sales last week per region" into the Power BI app on their phone, and getting a bar chart, a total, a short explanation and, if the tool does its job, a note of exactly which measure and which date range were used. That is the whole promise in one screen.

How the assistant turns your sentence into a query is covered in the text-to-SQL entry. This entry is about what you get to see, and about the question that matters more than the technology: how do you know the answer is right?

What happens between the question and the answer

Every serious product does the same five things, in this order, and each step is a place where the answer can go wrong.

1. It interprets the question against a model
   Your words get mapped onto things the model knows: a measure called Net Sales, a column called Region, a date table. The synonyms and descriptions the model builder wrote decide whether "turnover" lands on the right measure.

2. It builds a query
   DAX, SQL, KQL or a query against a semantic layer, depending on the tool. Some tools let the model write freely, others only let it pick a measure and a few filters.

3. It runs the query as you
   The result comes back under your permissions.

4. It writes the narrative
   A language model turns the result table into a few sentences and picks a chart type. This is what makes it feel like an analyst answered you.

5. It cites what it used
   Microsoft calls this "How Copilot arrived at this", Looker calls it "How was this calculated?", Amazon Quick puts the generated SQL in an Explanation panel. This is the step you should refuse to live without.

Why the semantic model decides the quality

On raw tables, the assistant has to guess. Ask "what is our margin per product group" and it will find a revenue column, find something that looks like a cost column, decide on a join, and give you a plausible number. It might use purchase price where finance uses landed cost, and you will never know, because the number looks fine.

On a semantic model with a measure called Margin, the assistant calculates nothing. It picks the measure your finance team defined, applies the filter, and returns the company's number. The definition was decided once, by a person, and every question reuses it.

That is why Microsoft's guidance for Copilot in Power BI is mostly about the model and hardly about prompting: descriptive names, synonyms, a schema that leaves out the technical columns, AI instructions with your vocabulary, and a model you explicitly mark as approved. Google says the same thing about Looker, where the LookML definitions are the "source of truth", and ThoughtSpot has a coaching step where an analyst teaches Spotter the terms of the business.

The corollary is uncomfortable. When a measure does not exist, most assistants improvise. Copilot in Power BI will write ad hoc DAX for "year-over-year growth" if your model has no such measure. Sometimes that is what you wanted. Sometimes it is a definition of growth nobody signed off on, presented with the same confidence as the real one.

Conversational analytics compared to a dashboard

The useful difference is who decides the question.

With a dashboard, the analyst decides. Someone thought in advance about the twelve questions the sales team asks every Monday, built a page for each, and checked the numbers. You can only ask those twelve questions, and the answers are dependable.

With conversational analytics, you decide, in the moment. The thirteenth question, the one nobody anticipated, gets an answer too. The price is that nobody checked that answer before you saw it.

So the two are not rivals. A dashboard is for the numbers that recur and have to be identical every time: the board pack, the weekly sales meeting, anything a regulator sees. Conversational analytics is for the follow-up question: "fine, but is that mostly one customer?" That is also why the owner on the train is the ideal user and the analyst is not. The owner wants an answer. The analyst wants to build, and for that you still need the report editor.

What the tools do today

This is the state of play in September 2026, as examples rather than a buying guide.

Copilot in Power BI answers with a visual, a text summary that lists the fields it used, and an expandable "How Copilot arrived at this". It can write ad hoc DAX and let you inspect the query. Model owners can add verified answers (a human-approved visual tied to trigger phrases, shown with a checkmark) and AI instructions. The Copilot pane inside a report is generally available; the standalone chat, the app chat and the mobile app chat were still in preview at the time of writing.

Fabric data agents are the version you configure yourself: up to five data sources, instructions, example queries, read-only, generally available, and callable from Microsoft 365 Copilot, Copilot Studio and Teams. Row-level and column-level security still apply because the agent queries as the person asking.

Looker Conversational Analytics runs Gemini over LookML. You build Explore data agents with instructions, verified queries and a glossary, and every answer has a "How was this calculated?" section with the raw field names, calculations and filters.

Amazon Quick Sight, the analytics part of what Amazon now calls Amazon Quick, works from Topics: a semantic layer over up to twelve datasets with custom instructions. The chat agent writes SQL across them and shows it in an Explanation panel.

Tableau Agent in dashboards gives an overview, per-visual insights and a Q&A box on the underlying published data source. It was in beta in the 2026.2 release and needs a Tableau+ or Tableau Cloud+ subscription.

What makes an answer trustworthy

None of these tools is trustworthy out of the box. They become trustworthy through five decisions you make, and we would not roll one out at a customer without all five.

• Verified answers for the recurring questions. Tie the twenty questions people ask most to a visual a human has checked.

• A fixed set of measures. The assistant picks from measures that exist. If you leave it free to invent calculations, write that down as a decision and tell the users.

• The query stays visible. Every answer shows which measure, fields and filters it used, and someone on your team can read that. An answer you cannot inspect is an opinion.

• A test set with known answers. Before the rollout, write fifty questions in the words your colleagues actually use, with the correct answer next to each, and run them. Run them again after every model change. A score below what you would accept from a new analyst means the model is not ready, whatever the demo looked like.

• Curated models only. Point the assistant at one or two approved models, never at every dataset in the tenant. Fewer tables means fewer wrong guesses. Power BI has an admin setting to show only approved items in the standalone Copilot for exactly this reason.

What to watch out for with conversational analytics

Ambiguous questions get confident answers. "Sales last month" can mean invoiced, ordered or shipped, and "last month" can be the calendar month or the last thirty days. A colleague would ask which. The assistant picks one. Put the definitions in the model and the instructions, and teach users to open the citation.

The time period goes wrong quietly. "Growth in March" against a date table that only knows weeks, or a fiscal year that starts in April, gives a number that is right for the wrong period. A date table with the right calendar, and a description that says so, prevents most of it.

Users cannot always tell a wrong answer from a right one. That is the whole risk in one sentence. The owner on the train sees a chart and a confident paragraph, and if the definition underneath is improvised, nothing on the screen says so.

Row-level security still applies, and the assistant must not see more than the user. Check that the tool queries as the person asking, not as a service account with access to everything. Fabric data agents and Copilot in Power BI do. Also read the small print: Microsoft says verified answers with row-level security are not fully supported during the preview and should not be relied on as a security feature.

Language. Copilot in Power BI and Fabric data agents officially support English only at the time of writing. A Dutch or French question sometimes works, and that is not the same as being supported.