When AI Builder stops returning a form field to Power Automate, the first question is not whether the whole model is broken. The field may still exist in the model, but it may be missing from the latest extraction, renamed after a model edit, hidden by a layout variation, or mapped incorrectly in the flow.
Keep one failed document and one known-good document open while you troubleshoot. If you are still building the initial extraction process, the broader AI Builder document processing notes are useful context before you change a production flow.
Look at the extraction result before the flow
Open the AI Builder model and test it with the same file that caused the failed flow run. If the field is missing in the model test, Power Automate is only passing along that result. If the field appears in the model test but not in the flow output, the problem is in the flow action or dynamic content mapping.
- Use the same file type and layout that failed in the flow.
- Check whether the field is blank, misread, or absent from the output.
- Compare the failed document with a document that still extracts correctly.
- Note whether the field label changed after the model was edited or republished.
This split matters. Retraining the model will not fix a flow action that still points to an old field token, and remapping the flow will not fix a model that never sees the field on the document.
Find out whether the field moved, changed, or was never trained
Form fields often disappear after a vendor changes the layout, moves a label, adds a second table, or changes the wording near the value. AI Builder may still recognize the document, but the field confidence can drop enough that the value is empty.
Review the labeled training samples. If the missing value appears in only one sample layout, add more examples of that layout before retraining. If the field name was edited, republish the model and expect Power Automate to need a remap because the old dynamic content reference may no longer match.
For fields inside tables, check whether the value belongs to a table column rather than a top-level field. A flow that expects a single field will not behave the same way when the model returns a collection of rows.
Repair the Power Automate mapping
After the model output looks correct, open the latest failed flow run and inspect the AI Builder action output. Then edit the flow and reselect the field from dynamic content instead of typing the old name. This forces the action to use the current model schema.
- Open the AI Builder action in the flow designer.
- Confirm it points to the current published model and environment.
- Remove the old field token from downstream actions.
- Select the field again from dynamic content.
- Run the flow with the same document and compare the output.
If the extracted value is sent into Forms, Excel, or an approval step, test that destination separately. For workflows that also collect responses from Microsoft Forms, this Forms and Teams workflow is a helpful comparison for checking whether the receiving step is the actual point of failure.
Questions before retraining the model
Why are only some form fields blank?
The model may recognize the document but lose confidence on fields whose label, position, or surrounding text changed. Those fields need better sample coverage, while stable fields can continue working.
Do I need to retrain the whole model?
Retrain only when the model test cannot extract the field reliably. If the model test is correct and the flow output is wrong, remap the dynamic content in Power Automate first.
Why did Power Automate lose a field after I edited the model?
Changing field names or republishing a model can leave downstream actions pointing to old tokens. Remove the old token and select the current field again from dynamic content.
The fastest recovery is to separate model extraction from flow mapping. Once you know which side lost the field, you can fix that layer without rebuilding the entire document process.