Soon everyone builds dashboards. Who guards the number?

Microsoft announced that you will soon be able to have an app generated in Power BI Desktop in plain language. You start from a model, describe what you want to see, and an AI builds it. Outside Microsoft things are moving the same way: an agent writes a dashboard in a quarter of an hour that used to take days.

That is good news. I build with AI myself, and it goes faster than ever. But it changes where the risk sits. Building a dashboard used to be hard, and that difficulty stopped a lot of bad dashboards. That brake is now gone.

The danger is not that everyone can visualise. The danger is that nobody checks any more what is being visualised.

A good-looking chart makes a wrong number believable

You distrust a messy spreadsheet. You see the manual adjustments, the columns that do not quite add up, and you ask questions.

You do not distrust a sleek app with neat cards, smooth charts and the company colours. The finish gives the number an authority it has not earned. And with AI that finish is free.

That is the first pitfall, and the most important one: the quality of the design no longer says anything about the quality of the number.

Ten colleagues, ten versions of revenue

Suppose five people in a company each have their own app made. The sales manager counts revenue by invoice date. Accounting by posting date. One person leaves out returns, another does not.

Each app is right on its own. Together they contradict each other. That problem is not new, it is the same as in when three people give three answers. Only now it happens faster, and every version looks convincing.

We have spent years getting companies away from ten spreadsheets with ten truths. The risk is that we soon get ten apps in their place.

AI builds what you ask for, not what you need

Ask an AI for a dashboard with revenue per month, and you get a dashboard with revenue per month. Neat, fast and exactly what you asked.

Nobody asks any more which decision you make with it, what you compare it with, or what is normal. Those questions, from the question to ask before you build a dashboard, take time. And when building itself no longer takes time, that time suddenly feels like delay.

The result is more of the dashboard nobody opens, only made faster.

A formula that works is not a formula that is right

In a data app the DAX for each chart sits in a separate file. That is an advantage: you can read it, test it and improve it. But then someone has to actually do that.

A measure that returns a number does not necessarily return the right number. The classics: a date table that is not marked as a date table, so your measures come back empty. A year-to-date comparison that puts a full previous year against half a current year, as in why a comparison by day lies to you. A filter that affects something elsewhere in the model without anyone noticing.

An AI writes such formulas fluently, and they look plausible. Whoever does not know DAX cannot see the difference. The DAX lesson covers the pitfalls you need to know to judge a formula, even if you do not write it yourself.

The built-in test checks whether it runs, not whether it is right

Microsoft builds an automatic browser test into the data app template. It opens the app and checks whether everything loads. That is useful: you do not want to publish a page that does not load.

But a page that loads with a wrong number passes that test. Nothing in that process compares the margin with the accounts. That still has to be done by someone who knows what the margin should be.

Code nobody can maintain

A data app is a web application in React and TypeScript. Most people who build Power BI reports today do not know those languages. That is fine as long as the AI writes it.

Until something breaks. An update of a library, a field renamed in the model, a feature Microsoft changes during the preview. Who fixes it then? The colleague who had the app made cannot read the code. In the community this is one of the objections you hear most.

A Power BI report nobody understands can still be opened and looked at. An app nobody understands is a black box.

A hundred apps, and none of them official

When building costs nothing, a lot gets built. After a year a company has dozens of apps. Some are used daily, some once, and some are still used although they have been showing wrong numbers for months.

The question then is no longer "do we have a dashboard?" but "which dashboard is the right one?". If nobody can answer that, everyone decides based on whichever screen they happened to open.

The cost you do not see

Building is faster, but it is not free. In Fabric, everything an app does consumes capacity: the queries on the model, the database, the storage. And every time an AI agent builds or changes something, tokens are used. For one app that is little. For fifty apps nobody cleans up, it adds up without anyone noticing.

What you do about it

This is not a plea to leave AI alone. It is a plea to do deliberately what used to happen by itself.

• One model as the only source. As many apps as needed, but they all read from the same model with the same measures. Then an app can look different, but not calculate differently. That is also why the data model is the work.
• Definitions on paper. What revenue is, what an active customer is, is settled before anyone has an app built.
• Official and sandbox. Everyone may experiment. But what is shared to make decisions on is marked separately and checked.
• Someone who checks. Before an app is shared, someone compares the key numbers with the source. Not the code, the numbers.
• One owner per app. Who is responsible when it breaks? No owner, no publication.
• Cleaning up. Apps that have not been opened for three months go.

Where the value moves

When everyone can make a good-looking dashboard in a quarter of an hour, making a good-looking dashboard is no longer a craft. The value moves to what sits underneath: a model that is right, data from Odoo, Exact or Yuki brought together correctly, definitions everyone stands behind, and the question of which decision is made with it.

That was really always the case. You could just hide it behind the work the charts required. That no longer works.

What you can automate

The checking itself can largely be automated. You can have the key numbers of every app compared with the source daily, with an alert on any deviation. You can have the DAX files of a data app checked automatically before they are published. And you can measure the use of every app, so you see which ones nobody opens any more and can be cleaned up.

The claim

AI makes visualising cheap. That makes the number underneath the only thing that still counts, and that is exactly what gets skipped most easily. Let everyone build, but make sure everything calculates from one model and that someone guards the number.