Copilot in Power BI

What it does

Copilot in Power BI builds and edits report pages from natural-language prompts, summarises and answers questions about report data, helps with DAX and semantic-model prep, and can drop a narrative visual straight onto a page. It works in both Power BI Desktop and the service.

Key facts

Report creation: describe the page you want, or hit "Suggest content for this report" and Copilot picks visuals off your model. It can add, change and delete visuals on existing pages, with undo/redo.
You can start report creation from a screenshot or image reference, not just a text prompt.
The Q&A feature switch on the semantic model must be on for report creation and editing to work.
Summarise and ask questions: Copilot writes plain-language summaries of a report and answers data questions in the Copilot pane.
Narrative visual: Copilot generates a dynamic, prompt-driven text visual that updates with the data and can be pinned to the page for all readers.
Semantic-model help: Copilot can draft measure descriptions and uses model metadata (DAX expressions, descriptions truncated at 200 characters, data types, format strings, relationships) as grounding.
DAX: Copilot assists DAX query writing in the DAX query view. On live connections to shared models it can't see measure DAX or hidden/private objects, so suggestions are weaker there.
Model quality drives everything. Descriptions, synonyms and sample values measurably improve the accuracy of Copilot's DAX and answers.

When to use / skip

Use Copilot in Power BI to get a first-draft report page in front of a client fast, and to hand report consumers a self-serve summarise/ask experience. It's genuinely good at the boring first 70% of a page. Skip it as a substitute for a well-built semantic model. If the model is thin on descriptions and synonyms, Copilot guesses, and it guesses confidently. Also skip relying on DAX suggestions over live connections to shared models, where it's half-blind.

Configuration decisions

Turn on the Q&A feature switch for any model that report authors will use with Copilot.
Decide a house standard for measure and column descriptions before rolling Copilot out to authors.
Choose whether the narrative visual is author-only or pinned for all readers, and who signs off its wording.
Set expectations on which capacity the Copilot usage bills against.

Gotchas

Descriptions are truncated at 200 characters for grounding. Front-load the meaning; don't bury it.
The narrative visual reads confidently even when the underlying figure is wrong. Someone has to sanity-check the numbers it narrates.
Copilot only sees non-hidden model objects. Hiding a measure hides it from Copilot too.
Report-creation quality swings hard on model prep. A demo on a clean model flatters what the client will get on their real one.
Prompt-to-report is non-deterministic. The same prompt twice can give different layouts.

Consultant notes

Budget model-prep time explicitly. "Copilot report creation" is mostly a semantic-model-quality project wearing a Copilot hat.
The summarise/ask features are often the easier win with business users than author-side report generation. Lead with those in demos.
Treat generated pages as scaffolding to refine, not finished deliverables. Set that expectation with clients up front.
Keep an eye on the agent-skills preview for report design-to-deployment; it's moving.

Confirm screenshot-to-report and narrative-visual availability in the client's region before you demo them.