From data to a living dashboard: asking the AI for a report that updates itself
A question answered once becomes a spreadsheet, and the spreadsheet ages. The next step is asking the AI to publish the answer as a page: with the queries underneath, refreshed hourly, with a link for the team. How to ask, what to demand, and what to check before sending the link to the board.
Every Monday someone repeats the same questions: how much sold, how much spent, how many leads, which campaign. The answer goes to a spreadsheet or a slide, and by Tuesday it is stale. The report that updates itself solves that — and the AI connected to the data is what can build it in minutes, because it already knows how to write the queries.
This article is the script for asking. It assumes the AI is already connected to the data and that you have already asked the small questions.
What a living report is
A published page, with its own address, made of blocks: numbers, charts, tables. Each block has a query underneath, and the platform re-runs the queries at an interval (hourly, for instance) and rewrites the page. Whoever opens the link sees the number of now, not of when the report was made. Filters on the page (by salesperson, by product) slice the same data with no new query.
How to ask
- 01Start with the questions, not the layout. “I want to follow acquisition sales per product, spend per channel and cost per real sale, last 30 days, sliced by salesperson.” The AI proposes the blocks.
- 02Demand the rule in every number. Sales = approved, acquisition, BRL currency. Spend = Meta + Google + warm-up. The rule stays written in the query, and the block hint repeats it for the reader.
- 03Ask for data from two sources only when they cross by person or by key. Adding Meta spend to Hotmart sales in the same block only makes sense if the ruler has been defined.
- 04Publish. The AI creates the report, materializes the queries and returns the link.
- 05Check before sharing: open the page, compare two numbers with the original source, and read each block’s hint.
A report without the rule written next to the number is a number nobody can dispute — and, for that reason, nobody should trust.
What to check before sending the link
- Does the total close? Sales per channel must add up to total sales; if they do not, some channel counts the same sale twice.
- Is the period the same in every block? A 30-day block next to a 7-day block looks like a comparison and is not.
- Are renewals separated from new sales? It is the most common mistake, and the one that inflates results the most.
- Does the block mixing sources explain the ruler? “Last paid touch” and “touched at some point” are different numbers and both true.
- Does whoever will open it have permission to see what is there? A report with lead names and phones cannot go by open link — it needs a password-protected link.
Edit without redoing
The report will change: a new block, a corrected rule, one more filter. The mistake is asking the AI to “generate it again” — it rewrites everything and some block changes without anyone asking. The right way is targeted editing: replace one block’s query, add a block, remove another, with the rest intact. And only the queries that changed are re-run.
A report like that is what turns the weekly numbers meeting into 15 minutes: everyone opens the same link beforehand, the numbers are current, and the time goes to the decision.
In CrazyLeads the living report is created through the MCP connector or the product chat, with queries materialized on an interval, filters on the page, targeted editing per block, a public link with a password per recipient and a record of who opened it — and the cost of each materialization measured.