Your manager wants the team to ask Copilot questions instead of waiting for a new report page. The question lands on you: is the model ready, and how do you know before people start trusting the answers? Microsoft's own advice is to fix the model first: "Resolve modeling issues in the semantic model because additional instructions can't correct inaccurate relationships or inappropriate aggregations" (Microsoft Learn, read 9 Nov 2026).
This is the hands-on version: a check you can do in about ten minutes, by the clock. We ran every step on a made-up model, Sahel Electronics, a pretend Gulf electronics retailer with four tables, four measures and one report page; no real company and no real data. For why each point matters, see our 7-point check; here we only do the check.
Before the clock starts: can Copilot run here at all?
If the answer is no, the rest can wait. From Microsoft Learn, read 9 Nov 2026 (Enable Fabric Copilot for Power BI):
- Capacity: the report's workspace must be on Power BI Premium (P1 or higher) or a paid Fabric capacity (F2 or higher), or you use a Fabric Copilot capacity. Trial capacities are not supported, and Copilot in Power BI Desktop does not support Premium Per User workspaces (Use Copilot in Power BI Desktop).
- The admin setting: in the Fabric admin portal, Tenant settings, Users can use Copilot and other features powered by Azure OpenAI. Microsoft says it is on by default, and an admin can turn it off or limit it to a group.
- The Gulf part: "If your tenant or capacity is outside the United States or EU data boundary, Copilot is disabled by default" (Microsoft Learn). A tenant in the UAE or Saudi Arabia is in that case: Copilot works only after a Fabric admin turns on the setting that lets data sent to Azure OpenAI be processed outside your region. That is a data-policy decision for your company, not a model fix.
- Q&A on the model: Copilot's data questions use the same engine as Q&A, so Q&A must be on for the semantic model (Microsoft Learn).
None of these is needed to run the check below. It works on the model file, so you can do it today and switch Copilot on later.
Minutes 0 to 2: run the Model Health Check
- In Power BI Desktop: File > Export > Power BI template. A .pbit holds the model's definition, not the data Power BI loaded (small tables typed in with Enter data and values saved in filters and slicers do stay in it).
- Drop the .pbit on the free Model Health Check. It runs in your browser; nothing is uploaded.
- Write down the score and the number of findings. You will run it again after the fixes.
Sahel Electronics scored 79, with 10 kinds of finding on 24 objects. Not all of them are about Copilot. Six kinds are, and they take the next four minutes.
Minutes 2 to 6: the six findings that matter for Copilot
- "ID, year or code columns that sum by default" (6 on our model, among them
Calendar[Year]andProduct[ProductCode]). Put a year that sums by default in a table's values and Power BI adds the years up: 2,026 for every day in the calendar. Microsoft lists inappropriate aggregations among what instructions cannot correct. Fix: set Summarize by to None on each one, and put the numbers people ask for in reviewed measures, as Microsoft advises for business-critical calculations. - "Visible key columns on the many side" (3) and "Columns not used anywhere" (4, two of them the month-number helpers). Fields you hide are left out of the AI data schema when you first set it up (Microsoft Learn). Fix: hide keys and sort helpers; remove the columns nobody uses (our
Sales[Old Notes]). - "Month or day names without a sort column" (2:
Month NameandHijri Month Name). Without a sort column, the Hijri months sort alphabetically: Jumada al-Awwal before Muharram. Fix: Sort by column, by the month number. - "Calendar table not marked as a date table" (1). The classic time-intelligence functions need it, and Copilot can write DAX for questions like year-over-year growth (Microsoft Learn). Fix: Mark as date table, on the
Datecolumn. - "Relationships that filter in both directions" (1, Sales to Branch). Microsoft asks for clear, logical relationships, and instructions cannot repair them. Fix: single direction, unless you know why it must be both.
- "Measures without a description" (4), plus
[Margin %]with no format and a plain/. Copilot reads descriptions, but "Copilot uses only the first 200 characters" (Microsoft Learn): start with the meaning, the unit and what is left out. Microsoft also asks for ratios as percentages and a margin divided from the approved measures. Fix, with the format0.0%:
Margin % =
DIVIDE (
[Total Sales] - [Total Cost],
[Total Sales]
)
Minutes 6 to 9: what no file check settles
A file can show what is missing. These four need you:
- The same name in two tables. Our model had
Namein Product andNamein Branch. Microsoft's tutorial says to avoid duplicate field names across tables (Microsoft Learn); the Health Check does not flag it today. We renamed themProduct NameandBranch Name. - Arabic column names. For Copilot's data questions, "The only supported language is English" (Microsoft Learn). Our Branch table has
اسم الفرعfor the report's Arabic pages. Keep it, and give it an English description that says what it holds: Branch name in Arabic (the same branch as Branch Name). - The Hijri calendar. If people will ask about Ramadan, the model must know it: our Calendar has
Hijri Year,Hijri Month NameandIs Ramadan, and a[Ramadan Sales]measure. Without them, Copilot has no Ramadan to answer about. Our Calendar Generator builds those columns. - Ten real questions. Microsoft suggests starting with focused business questions, each with the expected measure, grouping, date basis, filters and result. Write ten your team asks every month, with the answers you already trust. That is your test for later.
Minutes 9 to 10: write the list
Order it the way Microsoft does: the model first (the six findings and the four checks), then Prep data for AI, then testing with your ten questions. Each fix above is a minute or two in Desktop; the descriptions take longest.
What the fixes did to the made-up model
We made the fixes on Sahel Electronics' model definition and ran the same Health Check engine again. The score went from 79 to 98, and the findings from 10 kinds to 2: a hidden key that no visual uses (fine: it is the row's ID) and a note that the model has no row-level security roles, which is right for a model everyone may see. The DAX for [Margin %] above is the version that replaced the plain division.
Then Prep data for AI (preview)
Microsoft's page still calls it preview (Prepare your data for AI, read 9 Nov 2026). In Desktop it is the Prep data for AI button on the Home ribbon; in the service, on the semantic model's ribbon. Everything is saved on the semantic model, not the report. Three parts:
- AI data schema: the tables, columns and measures Copilot should use. Microsoft notes that when Copilot creates a report page, searches for data or writes a DAX query, it uses the entire semantic model and ignores this schema. So hiding and removing still matter.
- Verified answers: a visual you have checked, linked to the phrases people type (Microsoft suggests five to seven per answer). One catch from Microsoft's own list: "Verified answers don't work if you hide fields in the model", so build them on visible fields (Microsoft Learn).
- AI instructions: your business terms in plain language, up to 10,000 characters, and "There's no guarantee that the LLM will exactly follow instructions" (Microsoft Learn). For our made-up retailer:
Sales means [Total Sales]: net of discounts, excluding VAT, in AED.
Margin means [Margin %]. Never average the margin of single order lines.
Ramadan means the days where 'Calendar'[Is Ramadan] is TRUE. To compare one
Ramadan with the one before, compare by 'Calendar'[Hijri Year], not by [Year].
Branch names in Arabic are in Branch[اسم الفرع]. Answer with Branch[Branch Name]
unless the question asks for Arabic.
When the model is ready, its owner can mark it Approved for Copilot in the semantic model's settings in the service.
What Copilot does not see
- The slicers on the page. Copilot's data questions do not apply the filters or slicers currently on the report page (Microsoft Learn). "Sales this month" means the whole model unless the question says more.
- Hidden fields, pages and visuals, by your choice: Microsoft describes hiding fields to control what Copilot uses from the model, and hiding pages and visuals to control what a report summary reads (Microsoft Learn).
- A description past 200 characters, and a calculation group's items, which the model metadata does not include: list them in the calculation group column's description.
- Your instructions, for the people asking: consumers cannot see the AI instructions or the AI data schema. Tell your team what the model is for, in words.
Where DataArcus fits
The Model Health Check does the measurable part of minutes 0 to 6, free, in your browser. It reads the model's definition, not your data, and it does not run Copilot. When you want someone to do the whole check on a real model, the Copilot readiness audit is a fixed-price review of one semantic model: each finding with its exact fix, the Prep data for AI set-up, and the Gulf calendar if your business runs on it, in English or Arabic.
Questions
Do I need Copilot switched on to do this check?
No. Minutes 0 to 10 work on the model file. You need Copilot only to test the answers afterwards with your ten questions.
Can my team ask Copilot in Arabic?
Microsoft Learn says, for Copilot's data questions, that the only supported language is English (read 9 Nov 2026). Arabic column names can stay for your report; give them English descriptions.
Does the Model Health Check send my model anywhere?
No. It runs in your browser and nothing is uploaded.
AI features in Power BI change every month, so every Microsoft fact here is dated. For the background behind each check, read Is your Power BI model ready for AI? A 7-point check; for the Hijri columns, a Gulf calendar in Power BI.