Answers/Power BI

Is Power BI still worth investing in now that there’s AI?

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

I get asked this a lot lately, usually by somebody who was just told that dashboards are dead. The worry is reasonable. If anyone can ask a question in plain English and get an answer back, why keep paying people to build report pages?

Because the answer has to come from somewhere, and it has to be the right one.

I have an obvious bias, so let me put it on the table

I spent 13 years at Microsoft leading the BI work in Excel. I designed much of Power Pivot, and I was a founding engineer on Power BI. Then I left for the customer side of the same software, founded P3 there in 2013, and have been on that side ever since. So yes, I’m attached to it. Weigh what follows with that in mind.

I’ve also been paid for a long time now to tell companies when something they bought isn’t working, and I’m happy to do it. The part of Power BI I’d defend hardest is the one most executives never see. It’s the semantic model sitting under the reports.

What the semantic model actually does

Every report you’ve ever opened in Power BI sits on a model. The model says which tables exist and how they connect. It holds the measures, the formulas that say what revenue means, what counts as an active customer, and how year-to-date works when your fiscal year starts in October. It also carries row-level security, so a regional manager sees her region and nobody else’s.

When we built it, the reader we had in mind was a person making a report. Write the business logic once, in the model, and every visual on top of it agrees with every other visual. That turned out to be a good idea for a reason I didn’t see coming.

An AI needs what a report needed, only more of it

Ask an assistant “what was revenue in the Southeast last quarter” and it has several jobs to do before it can give you a number. Each of them is something the model either handles or leaves to chance.

A well-built Power BI model already does every one of those. I think that’s the most underrated fact in the whole AI conversation. Companies are out shopping for an AI layer, and a lot of them already own the hard part. I’ve written up the four ways this breaks, and how to tell which one you’ve got, in why your AI can’t answer questions about your own data.

A good semantic model is where your business wrote down what its own words mean. An assistant can read that. It can’t read the version in your analysts’ heads.

What changes for your BI team

Some things really do get smaller. I’d rather say that now than have you find out in a budget review.

Fewer tickets for new visuals

A big share of any BI queue is some version of “can I get the same thing, cut a different way.” Once people can ask for that in plain language, a lot of those tickets stop coming. The report canvas matters less than it did. I helped build that canvas and I still think so. There’s a whole separate argument about why business reports all look the same and what replaces them, and I’ll spare you the rerun here.

More modeling work, and a different kind

The work that grows is in the model. More people asking more questions means more of them land in corners nobody has touched since the model was built. And the reader has changed. A person can squint at a measure called Sales_Amt_2 and know from experience to use the other one. An assistant takes it at its word.

Models tuned for a literal reader

In practice that comes down to a handful of habits.

• Names that mean something. Net Revenue, spelled out, instead of whatever the source column happened to be called.
• Descriptions that carry the business rule. “Excludes intercompany. Fiscal year starts in October.” A description that repeats the name helps nobody.
• Relationships that are right, including the ones that were close enough back when a person was checking every output.
• One measure per metric, with the retired versions hidden or deleted so nothing reaches for them.
• Security tested by real people in different roles, before anyone else relies on it.

None of this is exotic. Good modelers have been nagging about most of it for years. What changed is that skipping it now costs you something visible, every time someone asks a question and gets a confident wrong answer.

Where the money should go

If you’re setting budget right now, my honest take is to spend less on building report pages and more on the model under them. The pages keep getting cheaper to make. The model is what every one of those pages depends on, and every answer from every assistant too.

And it matters less than you’d think which assistant you pick. Copilot, ChatGPT, Claude, whatever shows up in your company next year. They all read the same model. Get it right once and you’re covered for the ones you haven’t bought yet.

People ask whether Fabric changes the picture. In Fabric the semantic model is still the layer people and assistants query, so the same work pays off there.

When I’d tell you to hold off

If you have no semantic models at all, and every report is a flat export somebody dresses up in a visual, don’t start by buying AI licenses. Start with one model.

And if your Power BI estate is a tangle nobody trusts, AI will make that louder. That’s a repair job first. I’ve written about why people stop trusting the numbers and what fixing it looks like.

Four things you can check this afternoon

Where I’d start

Pick the one model your company leans on most, the one finance and sales both pull from. Clean up the names. Write descriptions on the measures people actually use. Hide the dead ones. Check the relationships against how the business really works.

Then point an assistant at it and have a handful of people ask it their real questions for a week or two. What they ask tells you what to fix next. It also tells you, faster than any roadmap would, whether your Power BI investment is paying off.

My bet is you already own more of this than you thought.