Prepare a semantic model for AI

  • 10 minutes

Now that your gold layer is designed with AI in mind, the next step is configuring your semantic model specifically for Copilot. Power BI provides a set of features called Prep for AI that let you add business context, simplify the schema for AI, and predefine answers to common questions. These tools give you direct control over how Copilot interprets and responds to user queries.

Access the Prep for AI features

The Prep for AI features are available in both Power BI Desktop and the Power BI service. In Power BI Desktop, select the Prep data for AI button on the Home ribbon. This opens a unified experience with three configuration tools:

  • AI data schema controls which tables and columns are visible to AI.
  • AI instructions provide business context and rules for AI interpretation.
  • Verified answers predefine responses to common business questions.

In the Power BI service, navigate to your semantic model and select Prep data for AI from the ribbon.

All updates you make through these features save to the semantic model, not to individual reports. This means your AI preparation applies everywhere the semantic model is used.

Note

You can author Prep for AI features in both Power BI Desktop and the Power BI service. Users consume these features everywhere that Copilot in Power BI is available.

Configure the AI data schema

The Simplify data schema section controls which parts of your semantic model Copilot can see in the Prep for AI feature. When you open this feature, you see a list of your model's tables and columns. You can:

  • Hide tables and columns from AI. Technical fields like surrogate keys, ETL columns, and internal identifiers should be hidden so Copilot focuses on business-relevant data.
  • Review field visibility to confirm that all important business fields remain visible.
  • Simplify the schema by reducing the number of fields Copilot needs to process, which improves response accuracy.

This complements the hiding you do in the model itself. Fields hidden in the model are already excluded from Copilot. The AI data schema provides an additional layer of control specifically for AI consumption.

Create verified answers

Verified answers are predefined responses to specific business questions. When a user asks a question that matches a verified answer's trigger phrases, Copilot returns the predefined visual and data instead of generating a response from scratch.

To create a verified answer:

  1. Build a report visual that answers a specific business question (for example, a card showing total sales for the current quarter).
  2. Select the visual.
  3. Select the ... menu, then select Set verified answer (in Desktop) or Set up a verified answer (in the service).
  4. Add trigger phrases that describe how users might ask this question. For example: "What were total sales last quarter?" or "Show me quarterly sales."

Verified answers ensure consistent, accurate responses for your organization's most common questions. They're particularly useful for questions that require specific filters, time periods, or business logic that Copilot might otherwise interpret differently each time.

When to create verified answers:

  • High-frequency business questions that executives or analysts ask regularly.
  • Questions where accuracy is critical and the correct answer involves specific measures and filters.
  • Questions where Copilot's generated responses have been inconsistent or inaccurate.

Tip

Start by identifying the five to ten most common questions your business users ask. Create verified answers for these first, then expand based on user feedback.

Write AI instructions

AI instructions provide Copilot with written business context about your semantic model. They help Copilot understand terminology, business rules, and data interpretation guidelines that aren't captured in table names or descriptions alone.

AI instructions are free-form text where you explain:

  • Business terminology. Define domain-specific terms. For example: "In this model, 'active customer' means a customer with at least one transaction in the last 12 months."
  • Business rules and exceptions. Explain how to handle edge cases. For example: "When calculating year-over-year growth, exclude the first quarter of 2022 because it was a partial reporting period."
  • Data scope and limitations. Clarify what the data covers and doesn't cover. For example: "This model contains sales data for North America only. European and Asian markets are in separate models."
  • Preferred measures. Guide Copilot to the right metrics. For example: "When users ask about revenue, use the 'Revenue (USD)' measure, not the 'Revenue (local currency)' measure."

AI instructions help bridge the gap between what your model contains and what your business users expect when they interact with Copilot.

Tips for writing effective AI instructions:

  • Keep instructions factual and specific. Avoid vague guidance like "use common sense."
  • Update instructions when business rules or data scope changes.
  • Focus on areas where Copilot has produced incorrect or inconsistent answers.
  • Don't repeat information that's already captured in measure descriptions or column names.

AI instructions and AI data schemas save to the semantic model alongside your other metadata. You can also manage them through Git integration or deployment pipelines. After deploying changes through these channels, refresh the model in the Power BI service to sync updates.

Mark your model as Approved for Copilot

After you configure the Prep for AI features and test the results, mark your semantic model as Approved for Copilot in the Power BI service. This designation signals that the model has been reviewed and prepared for AI consumption.

To mark a model as Approved for Copilot:

  1. Go to the Power BI service and find your semantic model.
  2. Select the Settings icon.
  3. Expand the Approved for Copilot section.
  4. Select the Approved for Copilot checkbox, then select Apply.

When a model is marked as Approved for Copilot, the standalone Copilot experience removes friction treatments (warning banners) from answers generated using that model. Reports built on an Approved for Copilot model also inherit this status.

Important

Marking a model as Approved for Copilot is a commitment. It tells your organization that the model has been deliberately prepared and that AI responses from this model are expected to be reliable. Prepare and test thoroughly before applying this designation.

Understand linguistic modeling

Linguistic modeling enhances how Copilot interprets natural language by mapping user terminology to your model structure. It works alongside the Prep for AI features to improve response accuracy.

You set up linguistic modeling through Q&A setup in Power BI Desktop. Key capabilities include:

  • Synonyms. Map alternate terms to fields. For example, map "turnover" and "income" to a measure called "Revenue."
  • Linguistic relationships. Define verbs that connect entities: "Customers purchase Products" or "Employees work in Departments."
  • Field exclusion. Disable specific fields from Q&A and Copilot by deselecting Include in Q&A in the Synonyms view.

Copilot can suggest synonyms automatically. Start with these suggestions, then curate them to match the terminology your users actually use. Share useful synonyms with your organization for reuse across models.

Linguistic modeling and Prep for AI features work together. AI instructions provide broad business context. Linguistic modeling provides precise field-level terminology mapping. Both contribute to the grounding data that Copilot uses during preprocessing.