In Short: A Useful Assistant on a Good Model, a Confident Liar on a Bad One
Copilot in Power BI does four distinct jobs: it drafts report pages from a prompt, summarises what visuals show, answers natural-language questions against your semantic model, and helps authors write and understand DAX. Used on a clean, well-described model, each of those saves real time. Used on the average inherited dataset, it produces fluent answers that are wrong in ways a casual reader cannot detect.
That is not a reason to avoid it. It is a reason to treat Copilot readiness as a data quality project rather than a licence purchase. The switch is easy to turn on; the trust is earned in the semantic model.
What Copilot Actually Does Today
Report drafting. Give Copilot a prompt like "create a page summarising sales performance by region and channel" and it produces a first-cut report page using the fields and measures in your model. The output is a starting point: layouts are serviceable, visual choices are reasonable, and an experienced author will still rearrange most of it. As a way past the blank page, it works.
Summaries and narratives. Copilot generates plain-language summaries of a visual, a page or a whole report: trends, outliers, notable movements. For executives who receive reports by email and want the reading done for them, this is quietly one of the most-used features. The summaries describe what the data shows; they do not know why, and they occasionally emphasise the wrong thing.
Data Q&A. Viewers can ask questions in natural language and get answers resolved against the semantic model, including simple visuals generated on the fly. This is the feature that inherits every strength and weakness of your model, and the one connected to the broader Fabric IQ grounding story that carries the same answers into Teams and Outlook.
DAX and modelling assistance. In Power BI Desktop, Copilot writes measures from a description, explains what an unfamiliar measure does, and suggests descriptions for model objects. For developers this is the least glamorous and most reliably useful capability: explaining an inherited model's 400-line measure is exactly the kind of work machines should do.
Where It Falls Short
A candid list, from experience across client estates.
- It cannot fix a bad model. Copilot resolves your words against the model's names and relationships. Cryptic names, missing descriptions and wrong relationships produce wrong answers delivered with total confidence
- Summaries lack business context. A revenue dip caused by a known contract timing issue will be narrated as a concerning trend, because Copilot does not know what you know
- Report drafts are generic. Useful scaffolding, but they follow no design system and ignore your organisation's reporting conventions
- Capacity dependency. Copilot consumes Fabric capacity units. Heavy use on a small capacity competes with refreshes and queries, and the accounting surprises teams that have not planned for it, something we cover in what Copilot in Fabric costs
- Non-determinism. The same question can produce differently phrased answers on different days. For governed reporting where numbers must be reproducible, that property needs managing
The Preparation That Makes It Work
Every improvement that helps Copilot also helps humans, which is why we recommend the preparation regardless of AI plans.
- Rename tables, columns and measures into the language the business actually uses
- Write descriptions on every measure that matters: what it means, how it is calculated
- Verify relationships and keep the model a proper star schema, per our star schema guide
- Hide keys, technical columns and deprecated objects so Copilot cannot select them
- Test row-level security under a viewer account, because Copilot answers respect RLS and a misconfiguration becomes very visible
The full checklist, including linguistic tuning and endorsement, is in our guide to preparing your semantic model for AI.
Should You Turn It On?
For most organisations on paid Fabric capacity: yes, for authors first. The DAX and documentation assistance carries near-zero risk and immediate productivity value. Viewer-facing Q&A and summaries deserve a staged rollout: pick two or three models you trust, prepare them properly, enable Copilot for a pilot group, and review what people ask and what they were told. Expand on evidence, not enthusiasm.
The teams that regret Copilot are almost always the ones that switched it on estate-wide against unprepared models and spent the next quarter explaining discrepancies. The technology was fine. The foundations were not.
Sources and Further Reading
Frequently asked
Four things in practice: draft report pages from a prompt, summarise what a visual or report shows in plain language, answer data questions against the semantic model, and help authors write and explain DAX measures. Authoring features live in Power BI Desktop and the service; consumption features appear for report viewers.
Yes. Copilot features are billed against Fabric capacity, and your tenant needs a paid capacity with Copilot enabled by an administrator. The original requirement of a large F64 capacity has been relaxed so that smaller paid SKUs can use Copilot, with consumption metered against capacity units. Check the current minimum in Microsoft's documentation, as it has changed over time.
It is as accurate as the semantic model it queries. Against a well-named, correctly related model with described measures, answers are generally reliable because they resolve to real queries. Against a cryptic or poorly modelled dataset, Copilot guesses at meaning and can produce confident, wrong answers.
No. It accelerates a developer - drafting DAX, explaining unfamiliar measures, producing first-cut report layouts - but the judgement work of modelling, performance tuning and understanding the business question remains human. Treat it as a productivity multiplier, not a replacement.
The usual reasons: your workspace is not on a paid Fabric capacity, an administrator has not enabled Copilot in the admin portal, your region requires cross-geo processing consent that has not been granted, or the feature is limited for your capacity size. Work through those in order.
Rename tables and measures into business language, add descriptions, verify relationships, hide technical clutter and test row-level security. Microsoft also lets you mark preferred (endorsed) models. The same preparation improves human self-service, so none of the effort is wasted.
