Power BI Copilot: What Can It Actually Do in 2026?

Type a sentence, get a finished report page. That's the pitch. What actually happens is closer to handing the work to a fast, confident junior analyst who's never seen your business before — right up until the moment their measure is wrong in a way that looks completely correct.

Copilot in Power BI can genuinely do a lot: build a report page from a description, write a DAX measure or a raw DAX query, summarize a page in plain English, even write descriptions for the measures already sitting in the model. None of that is exaggeration — it works, today, inside both Desktop and the Service.

What most guides skip is the two things that actually decide whether it's worth turning on: the licensing, which isn't what most people assume, and the specific ways it gets things wrong, which are predictable enough to plan around.

Here's what that actually looks like:

• 📄 Real generation, not just Q&A — report pages, DAX, and narrative text, not only answers to questions
• 🔑 Capacity-gated, not a per-user add-on — needs paid Fabric or Premium capacity at the org level
• 🧪 Needs review, every time — confidently wrong DAX is the real risk, not obviously broken DAX

Here's what it can actually do, feature by feature.

Report Creation

Describe a page and Copilot builds the whole thing at once — it reads the semantic model, identifies the relevant tables and fields, chooses visual types, places them on the canvas, and wires them up, instead of a developer dragging each field into each visual by hand. A prompt describing a monthly performance page with a trend line, a top-N breakdown, and a regional split turns into a laid-out page in well under the couple of hours that would normally take.

Visual Generation

Below the whole-page level, Copilot also handles one visual at a time — add a specific chart, change a visual type, adjust what it's grouped or filtered by, all from a description instead of the formatting pane. This is where it's most reliable, since a single visual is a much smaller, more constrained thing to get right than an entire page layout.

DAX Assistance

Copilot writes DAX in two different places. In the model, it writes measures from a description — "revenue growth versus the same period last year" becomes a CALCULATE-based time-intelligence expression, added to the model with an explanation of what each part does. Separately, in DAX Query View, it writes raw DAX queries from a description, useful for ad-hoc exploration without touching the model at all. It can also write descriptions for measures that don't already have one, which is worth combining with a documentation pass rather than treating as a one-off — we're publishing a full breakdown of auto-documenting a model with AI soon, worth a bookmark.

Summaries

The Narrative visual is where Copilot turns a page into a paragraph. It reads whatever visuals are on the page — respecting any filters currently applied — and writes two or three sentences describing what's happening, updating automatically as the underlying data refreshes. It has two modes: Copilot, which generates the text automatically and defaults on for licensed users, and Custom, which uses a template-based approach that doesn't need a Copilot license at all. The generated text has a length limit in the low thousands of characters, enough for a real summary, not enough to replace a full written report.

One request, read against the same semantic model, can come back as a report page, a DAX measure, or a plain-English narrative.

Licensing & Prerequisites

This is the part most guides get wrong, so it's worth stating plainly: Copilot in Power BI is not unlocked by the Microsoft 365 Copilot add-on used in Word, Excel, or Outlook. It's a separate product with separate licensing, gated at the capacity level, not the per-user level.

Limitations

Model Quality Is the Real Bottleneck: Copilot reads field names, descriptions, and relationships to figure out what to generate. A model with cryptic names and blank descriptions produces vague or wrong output — not because Copilot is broken, but because it's working from bad source material.

Confidently Wrong Beats Obviously Wrong: the DAX Copilot writes runs without error far more often than it's actually correct for tricky patterns. A semi-additive measure — an account balance, an inventory level — needs logic like LASTNONBLANK, not a plain SUM; Copilot can default to the simpler, wrong pattern, and the report still renders a number. Just the wrong one.

Security Is Enforced, Which Can Look Like a Bug: row-level and object-level security apply to every Copilot interaction, so a user with restricted access can ask a question and get an answer that looks incomplete or evasive. That's the security working correctly, not Copilot malfunctioning.

It's Off by Default, Region-Gated, and Not on Trials: a tenant admin has to explicitly enable it, the capacity has to be in a supported region, and trial SKUs don't qualify at all. A surprising number of "Copilot isn't working" tickets trace back to one of these three.

Real-World Demo

Here's what a real session looks like, prompt by prompt.

Copilot reads the model, identifies a Sales fact table, a Date dimension, and a Product and Region field, and lays out three visuals — a line chart, a bar chart, and a map or column chart — each connected and pre-filtered. The catch: the fiscal year filter is only right if the model actually has a fiscal-year-aware date table. If it only has a calendar-year one, Copilot applies a calendar-year filter and labels it "this fiscal year" anyway, confidently.

Copilot writes a CALCULATE-based time-intelligence measure, adds it to the model, and drops a card visual showing it on the page. Worth opening before trusting: does it reference the right date table, does it handle the first period in the data gracefully where "last year" doesn't exist, does the sign convention match the rest of the model.

The Narrative visual, in Copilot mode, reads the visuals on the page — respecting whatever filters are currently applied — and writes two or three sentences: which product or region moved the most, in plain English, updating automatically as the data refreshes.

FAQ

Do I need a separate Microsoft 365 Copilot license?

No — that's a common point of confusion. Copilot in Power BI is gated by Fabric or Premium capacity, not by the Microsoft 365 Copilot add-on used in Word, Excel, or Outlook. They're separate products with separate licensing.

Can I try Copilot before paying for capacity?

Not really — trial Fabric and Premium capacities explicitly don't support Copilot, so evaluating it means provisioning at least an F2 capacity, the current entry tier.

Does Copilot work in Power BI Desktop or only the Service?

Both, as of 2026 — though full functionality in Desktop still requires being connected to the Service, since the capacity check happens there.

Will Copilot ever write something that looks right but isn't?

Yes, and it's the main thing to plan for. Anything involving time intelligence, semi-additive measures, or filter context edge cases is worth reading before trusting, the same as reviewing any other developer's first draft.

Does using Copilot expose data to people who shouldn't see it?

No — row-level and object-level security are enforced on every Copilot interaction. It can only generate from, and answer about, data the requesting user already has permission to see.

Weighing whether the capacity cost is worth it for your team? We can walk through the real numbers for your setup.

Closing the Loop

Copilot doesn't replace the judgment a report still needs — it moves that judgment later in the process, from writing the first draft to reviewing one. That's a real time save on the report pages and measures that follow familiar patterns, and no shortcut at all on the ones that don't. The teams getting the most out of it aren't the ones typing the most prompts — they're the ones whose semantic model was already well-documented before Copilot ever touched it.

Copilot's exact feature set, licensing thresholds, and capacity requirements are genuinely one of the fastest-moving parts of Power BI right now — confirm current details in Microsoft's own Copilot documentation before treating any specific number here as fixed.