Of all the places Microsoft has bolted Copilot onto the data stack this year, the one I find most genuinely useful is also the least glamorous: Copilot for DAX. Not the "ask your data a question" feature that gets the applause — the quieter one that helps you write the formula. And I think it's the most useful precisely because of the thing that makes it unglamorous, which is worth unpacking, because it also carries a risk nobody's talking about.

Quick translation for anyone who doesn't live in this world: DAX is the formula language behind every calculation in Power BI. When a report shows "year-over-year revenue growth," some human wrote a DAX expression to define exactly what that means. DAX is powerful, and it is notoriously unforgiving — a language that looks like spreadsheet formulas and behaves nothing like them, with a learning curve that stops a lot of capable analysts at "I can build a chart but I can't write the measure behind it."

Why this is the Copilot that works

Copilot for DAX takes a plain-English description — "give me a measure for revenue this quarter versus the same quarter last year" — and writes the DAX for it. And here is the key property that makes it different from the flashier Copilot features: the output is checkable. You get a formula, you can read it, you can test it against numbers you already trust, and you can see whether it's right.

Contrast that with Copilot answering a business question in natural language, which I've argued is the risky use — because there, a confident wrong answer arrives as a number, with nothing exposed to inspect. A wrong DAX measure arrives as code you can examine and validate. The failure is visible and correctable, not buried and plausible. That single difference — checkable versus opaque — is the whole reason one of these Copilots earns my trust for daily use and the other earns my caution.

So for the analyst who knows what they want but stalls at the syntax, this is a real unlock. It collapses the distance between "I understand the business logic" and "I've expressed it correctly in DAX." It drafts the pattern, handles the fiddly time-intelligence boilerplate, and — my favourite use — explains an existing measure someone else wrote, turning an inherited wall of DAX into something you can actually understand before you touch it.

The new failure mode nobody's flagging

But there's a shadow to this, and I'd be doing my job badly if I sold the upside without it. When you lower the barrier to producing something, you also lower the barrier to producing something you don't understand. And a DAX measure you don't understand, shipped into a report people make decisions from, is a genuine liability.

Here's the trap. Copilot writes a measure. It looks plausible. The number it produces looks roughly right. So it goes into the report. And nobody on the team can actually explain why it computes what it computes, what edge cases it mishandles, or how it behaves when the filter context shifts — which in DAX is where all the real subtlety lives. You've shipped a calculation your organisation now depends on and no human fully owns. The measure is right until the day it's subtly wrong, and on that day nobody knows how to find the problem because nobody understood the formula in the first place.

That's not an argument against the tool. It's an argument about how you use it.

  • Use it to accelerate understanding, not to replace it. Have Copilot draft the measure and explain it, and don't ship anything you can't paraphrase back in your own words. If you can't explain what it does, you don't own it yet.
  • Keep testing non-negotiable. A drafted measure gets validated against known-good numbers exactly like a hand-written one. Copilot changes who writes the first draft; it changes nothing about your obligation to verify.
  • Treat the semantic model as the thing that makes it good. Copilot writes better DAX against a clean, well-named model — because it, like everything else, works from your definitions. Tidy model, better drafts. Messy model, plausible drafts built on the wrong column.
Copilot for DAX doesn't make DAX expertise obsolete. It makes the judgement to check DAX more valuable — because now there's a lot more DAX, written faster, by people who didn't have to understand it to produce it.

What it means for your power users

The phrase in the title — "quietly transforming power users" — is the real story, and it's a genuinely positive one if handled well. The power-user community in most organisations is where a lot of real analytical value gets created and also where a lot of un-governed, un-owned logic accumulates. Copilot for DAX pours accelerant on both. Your capable analysts get dramatically more productive at expressing sophisticated logic. And your estate accumulates sophisticated logic faster than anyone's ability to review it, unless you build the habit of understanding-before-shipping into how the team works.

My advice to anyone leading an analytics function: embrace this, enthusiastically, with one rule attached. Copilot can write the measure. A human still has to understand it before it ships. Hold that line and Copilot for DAX is one of the best productivity gains the platform has offered your power users in years. Drop it, and you're just industrialising the creation of formulas nobody can maintain — faster, more confidently, and with the machine to blame when one of them quietly breaks.