Microsoft has just set out Power BI's next chapter: a new experience in Power BI Desktop that builds data apps from your semantic model in plain language, and Copilot answers grounded in governed Power BI reports and semantic models (Microsoft's announcement). Every one of these features starts from the same place: your semantic model.

Microsoft is clear about what happens when that model isn't ready: "Without proper preparation, Copilot can generate less helpful and inaccurate results" (Microsoft Learn). In other words, the AI is only as good as the model behind it. Here are the seven things to check first, in the order we'd fix them.

1. Names a person can read, and no duplicates

AI reads your field names the way a new colleague would. Sls_Amt_2 means nothing; Sales Amount does. Watch for the same name in two tables, like Name in Customer and Name in Store: Microsoft lists duplicate field names across tables among the things that make accurate answers hard. Rename them so each field says exactly what it is.

2. Descriptions on the fields that matter

A one-sentence description on a table, column or measure tells the AI what it means and when to use it: "Net sales after returns, excluding VAT." Start with your top 20 measures and the columns people filter by most. In our experience, it's one of the cheapest improvements you can make.

3. Hide the technical fields

IDs, keys and helper columns such as sort-order columns are needed by the model but confuse the AI. Hide them, or leave them out of the AI's view (step 7). In Microsoft's own tutorial, the first fields removed from the AI's view are an account ID and the day and month sort-order helper columns.

4. Remove what nobody uses

Microsoft recommends simplifying models by removing unused tables, columns and measures, because they add uncertainty to answers. Unused objects are also the hardest thing to find by hand. Our free Model Health Check lists every column and measure that no visual, filter, measure or relationship uses, and writes the cleanup steps for you.

5. A clean star schema, without ambiguous paths

Fact tables in the middle, dimension tables around them, one clear path between any two tables. Ambiguous or many-to-many relationships make a person hesitate, and they make the AI guess. Microsoft's guidance is the same as for any good model: follow star-schema design and avoid ambiguous relationships.

6. Formats and sort orders, because the AI repeats them

If a margin measure has no format, a card shows 0.34 for a margin of 33.8% (0.3378): no percent sign, and two decimals that round the value. If month names have no sort-by column, a chart by month runs April, August, December. Copilot reads format strings from your model, and its tables use the model's formatting (Microsoft Learn), so the same 0.34 can reach whoever asked. The Model Health Check flags month and day names without a sort column and measures without a format.

7. Use Prep data for AI

Power BI now has a Prep data for AI button (in preview): on the Home ribbon in Desktop, and on the semantic model's ribbon in the service (Microsoft Learn). It has three parts:

  • AI data schema: choose which tables, columns and measures the AI should use, and leave the rest out.
  • Verified answers: link common questions to a visual you've already checked, so the answer is the one you approved.
  • AI instructions: explain your business terms in plain language, for example what "active customer" means in your company.

Two practical notes from Microsoft's documentation: Q&A must be turned on for the model, or the tabs are disabled; and AI answers are nondeterministic, so preparation improves them but can't guarantee the same answer every time.

If your business runs on the Gulf calendar

Ask an AI to compare "this Ramadan with last Ramadan" and it can only answer if your model knows when Ramadan was. Put Hijri months, Ramadan and Eid flags and your weekend in the calendar table (our free Calendar Generator builds one), then say in the AI instructions how your business compares those periods. We showed the DAX in Ramadan sales in Power BI.

Start with a ten-minute check

  • In Power BI Desktop: File > Export > Power BI template (a .pbit has the model but no data).
  • Drop it on the Model Health Check. It runs in your browser and nothing is uploaded.
  • Fix the unused fields, missing sort orders and formats first, then add names, descriptions and Prep data for AI.

AI in Power BI is moving fast. The models that get good answers will be the ones that were clear for people first.