Conversational analytics in clariBI: ask data questions in plain English

Most BI tools make you build something before you get an answer: a query, a chart, a data model. Conversational analytics skips that step. You type "What were our top 5 products by revenue last quarter?" and get back a chart and a short written summary.

In clariBI this screen is called Ask clariBI. This post explains what happens behind an answer, which questions work well, and where you should check the result before acting on it.

What happens when you ask a question

An answer goes through three steps:

1. Planning. A language model reads your question, looks at the column names and a few sample values from the sources in scope, and decides which calculations answer it: a total, a time series, a top-N list, a period comparison, a correlation.
2. Calculating. Ordinary pandas code runs those calculations over the full stored data of your sources. The language model does not add up numbers.
3. Describing. A second model call writes a short summary from the calculated results, and the chart is drawn from the same results.

The practical upshot: the chart reflects a real calculation on your data. Mistakes, when they happen, usually come from the planning step (the wrong column, the wrong period) rather than from arithmetic. The Ask clariBI docs cover the screen in detail.

Before you start

• A plan with AI. Trial, Starter, Professional or Enterprise. Free and Lite have no AI credits.
• At least one connected source that has finished its first sync. clariBI answers from synced data, so it knows what was in the source at the last sync, not what happened five minutes ago.
• Some AI credits left. Each answer costs at least 1 credit; the exact amount is shown under the reply.

Sessions and sources

Click Ask clariBI in the sidebar and start a New Session. Give it a name and pick the sources it should use. If you pick none, it uses all of them. Keeping a session to the two or three sources that matter makes answers faster and cheaper.

Follow-up questions in the same session keep the earlier context. "Break that down by month" works after "Top 5 products by revenue last quarter". A new session starts fresh.

Questions that work well

Single numbers - "What was total revenue last month?" - "What is the average order value for Q2 2026?"

Trends - "Show monthly new customers for the past 12 months." - "How has weekly ad spend changed since January?"

Comparisons between periods - "Compare revenue in August with July." - "How does this quarter compare with the same quarter last year?"

Rankings - "Top 10 customers by revenue this year." - "Which campaigns had the lowest cost per conversion in June?"

Breakdowns and distributions - "Revenue by country for last quarter." - "How many orders fall into each order size range?"

Across sources - "Compare daily ad spend from Google Ads with Stripe revenue for June." - "Is there a relationship between sessions in Google Analytics and signups?"

Questions can span several sources. clariBI joins them by date or by a column they share, then compares or correlates. Connecting Google Analytics alongside a revenue source is a good first pair to try. For more ideas, the help center has 30 example questions that get good results.

What comes back

• A short summary of what the numbers show.
• A chart or table drawn from the calculated results. The chart type follows the shape of the result: a trend becomes a line, a ranking becomes a bar chart. You can ask for another: "show that as a table".
• Suggested follow-up questions you can click.
• The time and credits the answer used.

If you ask for a forecast ("forecast revenue for next quarter"), the reply offers a card that opens a matching forecast or creates one. Forecasts use statistical methods, not the language model, and need Starter or higher.

Saving and sharing an answer

• Save as Report turns the session into a report on the Reports page, which you can share or download as PDF.
• Export Chat downloads the conversation, questions and answers, as a Markdown file.

Limits to know about

Ambiguous questions get a guess. "How are we doing?" has no single answer. "What was month-over-month revenue growth for August?" does. If a result looks off, rephrase with the column name, the aggregation and the period.

Similar columns cause most errors. With amount and total_amount in the same table, or "orders" meaning count in one question and value in another, name the one you mean.

The summary text is not checked in chat. AI dashboards and reports compare every narrated number with the chart data and drop sentences that cannot be backed. Chat answers do not go through that check yet. The chart comes from the calculation, so when a figure matters, read it from the chart or table.

"Because" is a hypothesis. The model can describe what changed. It cannot know why unless the reason is in your data.

Only connected data. clariBI does not browse the web and cannot see sources you have not connected. Everything it reads is read-only.

Our longer piece on how natural language analytics works and where it fails goes deeper on these failure modes.

Tips for better answers

• Name the metric, the grouping and the period: "revenue by product for Q4 2025", not "show me revenue".
• Use your column names when you know them.
• Ask one thing at a time, then drill down with follow-ups.
• Limit long lists: "top 10" reads better than "all".
• Check one number you already know before trusting a new source.

The post on writing better questions for AI analytics has 40 worked examples.

Questions or dashboards

Start with the question you asked someone to pull by hand last week. Connect the source it comes from, open a new session limited to that source, and ask it the way you would ask a colleague.