Answers/Power BI

Why nobody trusts the numbers in your Power BI reports

Rob Collie·Founder and CEO, P3 Adaptive·September 2026

You know the meeting. Someone puts up a number. Someone else says that isn’t what their report shows. The next twenty minutes go to working out whose number is right, and the actual decision slides to next week.

After enough of those, people quietly stop using the reports. They export to Excel and rebuild the number by hand, because at least then they know how it was made.

I’ve spent a long time with my hands in other companies’ Power BI models. Nearly every time, the cause is how the model got built. Mostly it wasn’t built at all. It piled up.

Nobody designed it

The usual history goes something like this. Somebody sharp in finance or operations builds a model to answer one question. It works. People like it. They ask for more, and more gets added, a table here and a measure there, by whoever had time that week.

Every change made sense on the day it was made. Nobody ever stepped back and looked at the whole thing, because there was never a day when that was the most urgent job. A few years on, the model is a record of every request it ever got.

I wrote the book a lot of people learned DAX from, so some of this is my fault. We taught a generation of analysts how to write a measure. Far fewer of us talked about when to delete one.

What piles up

When I open one of these models, I tend to find the same handful of things.

Measures copied and tweaked

Someone needed revenue without returns, so they copied the revenue measure and edited it. Someone else needed it without one big customer. Now there are several, the names barely differ, and each one is right for the person who wrote it.

Test calculations left in

A measure called Test, or Copy of Margin, or something with a person’s initials on the end. It was scaffolding. Nobody took it down, and now it sits in the field list right next to the real ones, looking exactly as official.

Relationships guessed

Two tables got joined on the column that looked right. Or a relationship was switched to filter in both directions to make one visual behave, and now filters flow places they shouldn’t. The numbers come out close. Close is the hardest kind of wrong to catch.

Logic buried in Power Query

A business rule got applied as a filter step deep inside a query, where nobody will ever think to look. The report says revenue. What it shows is revenue minus a division somebody filtered out years ago, for a reason that made sense at the time.

Every report on its own dataset

This one costs the most. Instead of many reports sitting on one shared model, each report came with its own copy of the data and its own measures. Fix revenue in one and the other nine still have the old version. It’s how two departments end up with honest, different answers to the same question.

A model that grew on its own ends up saying several things at once. People notice, and they stop believing any of it.

It matters more now that AI reads the model

People working in a messy model get by on instinct. They know Revenue 2is the one finance uses. They skip the test measures without thinking about it. They’ve learned which report to trust.

An assistant has none of that. It reads the model exactly as written. Ask it for revenue and it will pick one of your four, give you a clean confident answer, and never mention the other three. Ask again tomorrow with different wording and it might pick another one.

So the mess that used to cost you twenty minutes in a meeting now shows up for everyone who asks the assistant anything. I’ve gone into that in more detail in why your AI can’t answer questions about your own data, and into whether Power BI is still worth the investment given all this. It is, and the model is the reason.

Repair usually beats starting over

When trust gets this low, the instinct is to throw the model out and rebuild it properly. Sometimes that’s the right call. More often a rebuild brings its own mess: months where the old model keeps running, the new one isn’t finished, and people now have two sources to argue about.

Most of what’s in an old model is correct. It’s tangled up with the parts that aren’t. A repair keeps what’s right and gets rid of what isn’t, and the reports people already use keep working the whole time.

It’s unglamorous work. It also tends to go quicker than people expect, because you’re editing something that mostly works.

Somebody has to own it afterward

A cleaned-up model starts growing sideways again the week after you finish, unless someone is watching it. One named person per model is usually enough. Their job is to approve new measures and to say no to a fifth version of revenue. Bigger governance programs can come later if it turns out you need them.

Five things you can check this afternoon

Where I’d start

Leave the whole estate alone for now. Pick the one number that causes the most arguments, which is usually revenue, and trace it. Find every measure and every dataset that calculates it. Get the people who argue about it in one room and decide which definition wins. Then fix the model so that definition is the only one anybody can reach.

That one fix does more for trust than a dozen new reports. Once people watch a number stop moving between meetings, they’ll tell you which one to fix next.