Can AI use the approver’s final corrections as ground truth for learning?

During invoice processing, the import user may correct the extracted data before submitting it for approval. However, these corrections may still be incorrect and may subsequently be revised by the approver. We therefore consider the values corrected and confirmed by the approver to be the ground truth.

Can the AI automatically learn from the final approved values to improve extraction for subsequent invoices?

Please clarify the following:

1. Does the AI learn from the import user’s corrections before approval? If so, can these corrections be treated as provisional or excluded from learning until approval?
2. If the import user and approver enter different values for the same field, does the approver’s final value supersede the earlier correction in the learning data?