Asking your numbers questions in plain language

A dashboard answers the questions someone thought of in advance. Revenue per month, margin per product group, absence per department. That works, until you have a question that is not on it.

Then you have two options: start filtering yourself until you find it, or ask someone to build a new report. Both are too slow for a question you have today.

Asking questions in plain language solves that. You type what you want to know, and you get an answer from your own data.

How I use it myself

I work this way with my own data models every day. I built tooling for it, so I can work with the model directly through AI: asking questions, checking calculations, digging into something without first building a visual.

The most important thing it delivers is control. I can check a number in three places. In the source itself, the CRM or the ERP, through the API that returns the data as JSON. In the PostgreSQL database where the data lands after it is pulled, with an SQL query. And in the Power BI model, through an MCP connection that lets AI calculate directly with the measures the report uses.

If those three give the same number, I know the pulling is right, the processing is right and the calculation in the report is right. If they differ, I see straight away in which layer it goes wrong. By hand that is an afternoon of work. Through AI it is a few questions.

What struck me most along the way: it only goes well when the model is built properly. A question about revenue works when there is one clear calculation for revenue. If there are three side by side, the AI picks one. Not always yours.

What it does well

The quick question. Which customer ordered nothing last month? What did transport cost in the second quarter? Which branch dropped the most in July? Questions you would otherwise leave alone because they do not seem worth the effort.

And the follow-up question. A dashboard shows that something dropped. In plain language you ask straight on: where, with whom, since when.

Where it goes wrong

The answer sounds just as certain when it is wrong as when it is right. That is the biggest trap.

An AI sitting directly on your raw sources has to decide for itself what an active customer is, which invoices count and whether a credit note comes off revenue. It picks something defensible, and you do not see that a choice was made.

That is why this only works on a clean data model, with settled definitions. Then the choice sits in the model and not in the language model's guess.

How to do it safely

Three habits.

Have it show how it got to the answer. Which table, which filter, which calculation. If it cannot, do not trust the answer.

Check a new answer against a number you already know, the first few times. If last month's revenue matches your report, you can keep asking. And if you can check across layers, as described above, do it.

And do not give it more access than the user has. A salesperson asking questions through AI should have the same limits as in their report. See who gets to see what in your report.

What it does not replace

The fixed dashboard. The numbers you look at every week, you want to see the same way every week, not phrased differently each time. Plain-language questions are for everything next to that.

And the thinking. As the article on AI puts it: it is precise and fast, but the good question still has to come from you.

The claim

Asking questions in plain language does not make your data smarter. It makes your definitions visible, and with them your gaps.