Power BI Q&A Is Being Retired.
Here Is What To Do About It.

If your users type questions into a Power BI dashboard tile and get a chart back, that feature is going away. Microsoft has confirmed the retirement of Q&A, and the date has already moved once.

This is not a feature quietly losing support. On the retirement date the Q&A visual, dashboard tiles, mobile Q&A and embedded Q&A stop working. If people in your business rely on typing questions at their data, they will arrive one morning and find it gone.

What Is Happening

Microsoft announced the deprecation of Q&A, its original natural language feature, and has since extended the retirement to February 2027 to give organisations more time to assess dependencies. The direction is settled; only the date moved. You can read Microsoft's deprecation notice directly.

The replacement is Copilot, which is a genuinely more capable tool. It handles multi-turn conversation, generates whole report pages from a description, writes DAX, and produces narrative summaries. It is also a different product with different requirements, which is where the migration work sits.

What Actually Breaks, and What Replaces It

Not every Q&A surface has a like-for-like replacement. This is the part worth checking against your own tenant before the deadline:

That last row is the one to check first. If you embedded Q&A into your own application using the PaaS embedding model, Microsoft has stated there are currently no plans for a Copilot equivalent. You are not migrating; you are rebuilding that capability another way.

What Copilot Actually Requires

Q&A was available broadly. Copilot is not, and this surprises people mid-migration.

Copilot runs on paid Fabric capacity, and the entry point has come down considerably since launch — it now runs from F2 upward rather than requiring F64. Premium P1 also qualifies, as does Premium Per User at around $20 per user per month with the Fabric trial enabled.

Your tenant admin must also explicitly enable Copilot in the admin portal. It is off until somebody turns it on, which is a common reason a migration appears to have failed when it has simply not been switched on.

The practical point: budget for capacity and check your licensing position before you promise users a replacement date.

There is a second-order effect worth planning for. Because Copilot is capacity-based rather than broadly available, the set of people who can ask questions of the data may be smaller than the set who could before. If Q&A was your way of giving occasional users self-service access, work out who loses it, and decide whether they need a licence or a better-designed report instead. That decision is easier made now than during a rollout.

The Catch Nobody Mentions

Here is the part that decides whether your migration succeeds, and it has nothing to do with licensing.

Copilot answers questions by reading your semantic model. Its accuracy depends overwhelmingly on how that model is built: field naming, table structure, measure descriptions and synonyms. Industry guidance puts roughly 80% of Copilot answer quality down to model quality rather than the AI itself.

Which means if your model has tables called Table1, columns called Column3, measures named Measure 2 (copy), and a flattened structure with no proper relationships, Copilot will produce confident, wrong answers. Users will try it twice, get nonsense, and never open it again.

This is the same reason Q&A frequently disappointed people. The tool changed; the dependency on a well-built model did not.

Preparing Your Model: The Actual Work

Four things, in order of how much difference they make.

Rename in the semantic model, not downstream

// Column names Copilot can match against natural language
DOC_AMT_LC        ->  Invoice Amount
CUST_NM           ->  Customer Name
POST_DT           ->  Posting Date
MATNR             ->  Material Number

Description as Copilot context

Measure:      Net Revenue
Description:  Gross revenue less returns and intercompany.
              Excludes freight. Matches the figure in the
              monthly board pack. Use this for external reporting.

What Copilot Still Cannot Do

Worth setting expectations before you roll it out, because an oversold tool gets abandoned faster than an undersold one.

It cannot invent a measure that does not exist. Ask for gross margin by region when nobody has defined gross margin in the model, and you will get either a refusal or a guess assembled from whatever looked close. The guess is the dangerous outcome, because it looks like an answer.

It cannot resolve a definitional disagreement. If finance and operations mean different things by revenue and both definitions live in the model, Copilot will pick one. It has no way to know which the person asking meant.

And it does not remove the need for designed reports. Copilot is strong at ad hoc questions and exploration. The recurring weekly number that ten people check still belongs on a built page, because consistency matters more than flexibility for a metric people track over time.

Positioned as a way to answer the long tail of one-off questions, it lands well. Positioned as a replacement for your reporting, it disappoints.

What We Would Do

A Copilot readiness engagement is mostly modelling work, not AI work. We audit the semantic model against what Copilot needs, rename and describe fields in business language, add synonyms, correct the structure where it is flattened, then test against the real questions your users ask and iterate until the answers hold up.

We also map your Q&A exposure first, because the PaaS embedding case needs the most lead time and is the one most often missed until late.

If you have until February 2027 and a model that was built for reports rather than for questions, the modelling work is the long pole. It is worth starting before the deadline is close enough to be a problem.