Error message when trying to prep model for AI

This error looks service-side. “NullSchemaPackageEvenAfterForcedUpdate” with HTTP 500 usually means Power BI failed to generate or retrieve the AI schema package for the semantic model, even after forcing a refresh.

Try these in order.

1. Confirm the basics are truly met

• The semantic model sits in a Copilot-enabled workspace on supported capacity.

• You have authoring rights on the semantic model (Member/Contributor in the workspace and Build permission on the model). The Prep data for AI experience requires authoring permission, not Viewer access.

• The model refresh completes successfully (no refresh failures, no credential prompts).

2. Force a clean “schema + metadata” update from Desktop, then republish
   This often bypasses a broken service-side package.

• Update Power BI Desktop to the latest.

• Open the PBIX that publishes this semantic model.

• Refresh, then publish again to the same workspace.

• After publish, open the semantic model in the service and try Prep data for AI again.

3. Try authoring Prep data for AI in Desktop instead of the service
   Microsoft supports authoring these features in both Desktop and Service. Desktop authoring sometimes succeeds when the service dialog fails.

• In Desktop: Home ribbon, Prep data for AI.

• Make a tiny change (add one AI instruction or one schema entry), save, publish.

• Then validate in the service.

4. Eliminate model metadata edge cases that trip schema packaging
   I would check for:

• Duplicate display names across tables/columns/measures (same visible name in different places).

• Hidden key columns used in relationships without a readable description.

• Columns with mixed data types or “Any” type coming from Power Query.

• Measure names or table names with unusual characters.
  Fixing duplicates and cleaning names is low effort and improves Copilot quality anyway.

5. Quick isolation test
   Create a minimal copy to see if the issue is model-specific or tenant/service-wide.

• Save a copy of the PBIX.

• Remove all but 1–2 tables and a couple measures.

• Publish to a new test workspace (same capacity).

• Run Prep data for AI.
  If the small model works, the issue sits in the original model metadata. If it fails everywhere, the issue is tenant/service-side.

6. Collect the right evidence and escalate with Microsoft support
   If none of the above works, you need the activity details from the failure and a diagnostic capture.

• When the error pops, copy the Activity Id / Correlation Id (if shown).

• Use the Copilot pane diagnostics option in Desktop (Download diagnostics) and attach it to the support request.

Thank you!
Proud to be a Super User!
📩 Need more help?
✔️ Don’t forget to Accept as Solution if this guidance worked for you.
💛 Your Like motivates me to keep helping