A fallback should help someone continue the work. “Something went wrong” is rarely enough.

Preserve the user's input, summarize the completed steps and identify the point that needs attention. Show the source information or draft where it is safe to do so. Avoid automatically repeating an action that might already have succeeded.

Test the fallback with the people who will use it. They should be able to recover without reading an engineering log or reconstructing the original request. Error handling is part of the workflow, not a separate screen added at the end.