Power BI Q&A Retires in February 2027: Moving to Copilot, and What It Costs

Last updated: October 2026

Microsoft announced the end of Q&A on 1 December 2025 with a December 2026 date, so teams that planned around December now have roughly two extra months. The main Q&A pages on Microsoft Learn carry the new date, although the developer page for embedded Q&A has no retirement notice yet.

Q&A tends to be in more places than people remember, from Q&A visuals in old reports and a dashboard question box a few managers use every morning to synonyms added years ago and Q&A embedded in a customer portal. Each moves to a different replacement, except app-owns-data embedded Q&A, which has none.

What is retiring, and what replaces each part

Microsoft's September 2026 announcement maps each retiring Q&A experience to an alternative, several of them still in preview.

The standalone Copilot searches reports, semantic models and apps, and Microsoft doesn't list dashboards among them, so we recommend pointing dashboard users at the reports behind their tiles.

Do not switch off Q&A on your semantic models as part of the clean-up. Copilot answers data questions with the same engine as Q&A, and Microsoft's current documentation says Q&A must stay enabled for those answers and for Prep data for AI. Microsoft said in December 2025 that a new file format would remove this dependency, but pages updated in September 2026 still describe it.

Check DirectQuery and Direct Lake models first, because Microsoft says Q&A may need to be turned on manually for them. Power BI indexes the text columns of a DirectQuery model about once every 24 hours, more often if the model is republished, and only for sources that support APPROXIMATEDISTINCTCOUNT. Microsoft's DAX reference lists Snowflake among those sources, although the Q&A limitations page names only SQL Server 2019, Azure SQL Database and Azure Synapse Analytics. Microsoft also says reindexing happens only when Copilot or Q&A has been used on the model in the last 14 days. We recommend including a Snowflake DirectQuery model in the pilot and expecting indexing queries on the warehouse up to once a day while the model is in active use. Our guide to DirectQuery and Import with Snowflake covers the storage-mode choice.

Start with an inventory, and capture your users' questions now

We found no Microsoft tool that finds Q&A visuals or converts Q&A setup content, although Microsoft's December 2025 post promised migration documentation. Part of the inventory can still be scripted.

• Reports with a Q&A visual or button. Search the definitions from the Fabric REST API's getDefinition call, which needs read and write permission on each report, or the Power BI Project files in source control.
• Dashboards where people use the question box, and mobile users who ask questions in the app. Only asking users will find these.
• Embedded solutions that load Q&A through the embed API. A search of the application code for type: 'qna' finds their embed configurations.
• Q&A setup on each semantic model, such as synonyms, phrasings, row labels and Teach Q&A terms. In Power BI Desktop, Linguistic Schema then Export linguistic schema saves the model's linguistic schema, which Microsoft says includes its synonyms and phrasings, to an .lsdl.yaml file. Teach Q&A definitions are stored as synonyms and relationships, so they should be in the same file, but check the export. Row labels are set in each table's properties in Model view, so record them separately.

Before Q&A goes, save the record of what your users have asked it. The Review questions page in Q&A setup, opened from the gear icon on a Q&A visual in Power BI Desktop, lists each semantic model's questions and the words Q&A didn't recognise. It keeps only 28 days and has no documented export, so copy the list now and every four weeks until February. It can be incomplete, because administrators can turn off question storage and users can opt out, but it is still a test set in your users' own words.

If a Q&A visual always shows the same answer, it doesn't need Copilot. Its convert icon turns the result into a standard visual, such as a column chart, which the retirement doesn't affect.

What Copilot needs: capacity, licences and tenant settings

Copilot needs organisational capacity, so teams with only Pro or PPU licences will need to budget for one.

• Capacity. Copilot for Power BI requires a paid Fabric capacity of F2 or higher, or Power BI Premium P1 or higher, and because Microsoft is retiring P SKUs a new purchase should be an F SKU. Trial capacities, free SKUs and sovereign clouds aren't supported. Copilot can take up to 24 hours to recognise a new or resized capacity.
• Fabric Copilot capacity. This can be created on F2 or higher. It bills a chosen group of users' Copilot usage to one capacity, including use in Pro and PPU workspaces, and takes precedence over the workspace's capacity. It must be in the tenant's home region and doesn't support the Embedded licence mode.
• Viewer licences. On F SKUs below F64, everyone who views Power BI content still needs a Pro, PPU or trial licence, so an F2 bought for Copilot doesn't replace Pro licences.
• Microsoft 365 Copilot Chat. Outside Microsoft's Q&A mapping, Fabric IQ in Microsoft 365 Copilot Chat (generally available) answers questions from Power BI reports and semantic models without the user needing access to Copilot in Fabric. It needs a Microsoft 365 Copilot Premium licence for each user, and outside the US and EU it needs cross-geo processing. It doesn't cover dashboards, Embedded workspaces or Power BI-only regions, and Microsoft doesn't say whether its answers consume Fabric capacity.
• Tenant settings. The main Copilot setting is on by default and cross-geo processing is off. The standalone Copilot has its own tenant setting. Microsoft's admin page lists it as disabled by default, while the enable page says it has been on by default since September 2025 where Copilot is on, so check it. The standalone Copilot also needs the main Copilot setting on at tenant level, because delegating that setting to a capacity isn't enough.
• Language. Copilot data questions are officially supported only in English.
• Networking. Copilot isn't currently supported with Private Link or in closed network environments.

In India and the UAE, the processing region can decide whether the project goes ahead. The Azure OpenAI service behind Copilot runs in the US and the EU data boundary, and capacities in India and the UAE are processed in the US. Outside those two areas Copilot is disabled by default until a Fabric administrator enables cross-geo processing, and UK capacities, though processed in the EU boundary, need the setting too. Microsoft's Fabric region table doesn't list a Saudi Arabia region. The standalone Copilot isn't yet available in the Power BI-only regions India West, UAE Central and Qatar Central, and it needs a home tenant in a full Fabric region such as Central India or UAE North.

Microsoft says what is sent includes prompts, schema and conversation history, and that table contents aren't sent unless you direct it. Its privacy page also says grounding data can include specific data points, so we would assume values used in answers are processed in the US too. We recommend agreeing the cross-geo setting with whoever owns your data residency commitments before the pilot starts.

What Copilot costs

Copilot in Power BI doesn't have a per-user price. It consumes capacity units based on tokens: 100 CU seconds for every 1,000 input tokens, 10 for every 1,000 cached input tokens, and 400 for every 1,000 output tokens. These background operations are smoothed over 24 hours and appear in the Fabric Capacity Metrics app as "Copilot in Fabric", and refreshes or DAX queries that Copilot triggers are billed separately.

Microsoft's worked example is a request with 2,000 input tokens and 500 output tokens, which costs 400 CU seconds. By our arithmetic, an F2's 48 CU hours a day cover roughly 432 such requests if nothing else runs, and at the East US pay-as-you-go rate on 23 October 2026 (USD 0.18 per CU hour) each costs about two US cents. Real requests vary, so Microsoft recommends measuring a small pilot group's consumption in the Metrics app before a wider roll-out.

We calculated these prices from Microsoft's Azure retail price list on 23 October 2026, at USD 0.18 per CU hour in East US and USD 0.20 in Central India, for a 730-hour month. A one-year reservation came out about 40% cheaper in every region we checked. Copilot is billed at the CU price of your billing region even when requests are processed elsewhere, so cross-geo processing doesn't change the rate. Confirm current prices in the Azure pricing calculator before you budget.

Copilot consumption counts against the same capacity as the reports on it, and Microsoft warns that broad deployment can quickly exhaust a capacity, at which point every workload on it is throttled. We recommend a Fabric Copilot capacity whenever Copilot would otherwise share a capacity with production reports, because Copilot's token consumption is then billed to it, while downstream operations Copilot triggers, such as DAX queries, stay on the semantic model's capacity. For a team on Pro licences, the smallest set-up is an F2 Fabric Copilot capacity with the Copilot users assigned to it, while their reports stay in Pro workspaces.

Capacity overage can pay for the excess on F SKUs at three times the pay-as-you-go rate, up to a rolling 24-hour threshold that Microsoft says isn't a hard spending cap. Its overage page calls the feature opt-in and also says it is enabled by default on new Fabric capacities, so check the setting. Our Fabric and Snowflake comparison explains how smoothing, throttling and overage work.

Rebuild Q&A setup as Prep data for AI

Prep data for AI is Microsoft's replacement for Q&A setup. It is in preview, it is saved on the semantic model rather than on a report, and it applies everywhere Copilot uses that model. It has three parts.

• AI data schema. This is a subset of fields that Copilot gives priority to when answering data questions. It doesn't apply to report summaries, report page creation, search or DAX queries.
• Verified answers. These are approved visuals linked to trigger phrases, up to 250 per model with up to 15 phrases each. They don't support relative date filters such as year to date, hidden fields or report measures, and aren't returned when Fabric IQ reasoning is on in the Copilot pane. In the preview, row-level and object-level security aren't fully supported for them and data might be exposed, for example through the file format in Git, so we recommend keeping them away from RLS- or OLS-protected data.
• AI instructions. These hold up to 10,000 characters of business context and rules, such as which date column to use for revenue questions, or that busy season means June to August. Microsoft says there is no guarantee Copilot follows them.

Verified answers don't support Git integration, REST APIs or deployment pipelines, so plan on recreating them in each environment. AI instructions and the AI data schema travel in the model's linguistic schema, but after a Git or pipeline deployment they apply only once the model is refreshed in the service, and on DirectQuery and Direct Lake models that sync happens once a day.

No Microsoft page we found says whether existing synonyms and Teach Q&A terms will keep influencing Copilot after February 2027. The Copilot guidance on Microsoft Learn still tells authors to add synonyms, so we wouldn't delete them, but we would move key business terms into AI instructions and turn the most common saved Review questions into verified answers.

Microsoft advises fixing the model before preparing it, because instructions can't repair wrong relationships or bad aggregations. In practice that means a star schema with clear relationships, fact tables with a documented grain, descriptive measure names (Average Customer Rating, not AvgRating), and descriptions that put the important part first, because Copilot reads only the first 200 characters. Our Power BI performance guide covers the model problems that also make reports slow.

Once a model is prepared, you can mark it Approved for Copilot in the semantic model settings, which removes the low-quality warning the standalone Copilot shows for unprepared models.

How to test Copilot's answers before users rely on them

Asking a question twice is a weak test, because Microsoft's Copilot overview documents a cache. An identical prompt on an unchanged semantic model within a rolling 24-hour window is answered from the cache, even after clear chat, so a second identical answer doesn't show that the answer is stable. The checklist below adapts advice from Gulab Chand Tejwani's Towards AI post of 20 September 2026.

The same page says Copilot answers from the semantic model when a question relates to its data, and otherwise from the language model's general knowledge. Both kinds of answer appear in the same pane, and if a user's wording doesn't match your field names, Copilot may not treat the question as being about your model and can answer from general knowledge instead. When Fabric IQ is off, answers from the model include How Copilot arrived at this, which shows the fields, measures and filters Copilot selected. When Fabric IQ is on, Copilot shows its reasoning steps instead, including the DAX query it generated and the query results.

• Ask the same business question in three different wordings, and repeat important questions after a refresh or more than 24 hours later.
• Check answers against the source system or a trusted report, never against an earlier Copilot answer.
• Open How Copilot arrived at this, or the DAX query with Fabric IQ on, for every test answer and check the measure, filters and date basis. Treat an answer without it as ungrounded.
• Test as a user under row-level security. Microsoft says Copilot only accesses data the user may see, but we found no Power BI page on the report pane's behaviour under RLS or OLS.
• Use your saved Review questions. Also deliberately ask questions your model can't answer, so you know what an ungrounded answer looks like.
• Compare answers to data questions with an unfiltered view of the report, because they don't apply the page's filters and slicers. In-report Copilot on mobile uses the current state of the report, so a phone can show different numbers.
• Run the tests in the Power BI service, because Microsoft's data-question page says data questions are service-only for now, although the Prep data for AI pages describe testing in Desktop.

We recommend telling business sponsors at the start that Copilot can return inaccurate content, and that Microsoft's privacy documentation says people who can judge accuracy should review its answers.

Embedded Q&A: the gap with no direct replacement

According to Microsoft's September 2026 post, there are currently no plans for a Copilot replacement for app-owns-data embedded Q&A, although in December 2025 Microsoft had said PaaS support was coming.

The options Microsoft lists are not direct replacements. The first is embedding for users who sign in to Power BI. Copilot can then be shown inside an embedded report when all of the following are in place.

• The report shows the action bar.
• The app has the MLModel.Execute.All permission.
• The embed URL includes reportCopilotInEmbed=true.
• The Copilot tenant setting is on.
• The content sits on Premium or paid Fabric capacity.

Secure embedding in a portal or website uses the Enable Copilot option in the embed dialog. For a software product this is a large change, because every customer user then signs in to Power BI and, on F SKUs below F64, needs a Pro or PPU licence, which app-owns-data didn't require.

The second option is the narrative visual and smart narrative summaries. The third is a custom question interface built on the Fabric IQ MCP Server. It doesn't support service principals or app-only authentication, so every user signs in with a Microsoft Entra work or school account and queries run under that user's permissions. It has no natural-language answering tool, so your application supplies the agent that writes the DAX, and it isn't available to tenants whose home region is Power BI-only, such as India West. Each option changes the product, so the decision belongs with product owners as well as BI engineers. Our embedded analytics guide covers the capacity side of Power BI Embedded.

A timeline that leaves room before February

Microsoft hasn't named a day, so we recommend planning as though Q&A stops on 1 February 2027.

• October and November 2026: build the inventory, save the first Review questions list, convert fixed-answer Q&A visuals, decide on capacity and cross-geo processing, and fix and prepare one or two models for the pilot. Buy the paid capacity at least a day before the pilot.
• December 2026: run a Copilot pilot with a small group on the prepared models, measure consumption in the Metrics app, test with the methods above, and save the next Review questions list.
• January 2027: roll Copilot out to the wider audience, remove the remaining Q&A visuals, tell users where their question box has gone, and save the last Review questions list.
• February 2027: Q&A stops working, and any remaining Q&A visual displays an error.

Teams that rarely used Q&A may decide Copilot isn't worth a capacity yet. They still need to convert or remove Q&A visuals before February.