A PDF containing several invoices is not necessarily one business document for AI Builder. Microsoft's document-processing guidance recommends separating multiple invoices before sending them to the model. Partial output can result from treating a combined scanning batch as a single form.
Use invented documents and preserve the original batch. This guide concerns custom document-processing input preparation, not a guarantee that every invoice layout will be extracted correctly after splitting. Confirm the model and action actually used by your flow.
Identify business boundaries, not just page boundaries
Count the invoices in the PDF and the pages belonging to each. One invoice can span several pages, while another occupies a single page. Splitting every page into a separate file can therefore be just as wrong as submitting the entire batch as one document.
The official document processing FAQ explicitly gives multiple invoices as an example of material to separate into individual files. Build the intake rule around the business document's identity and page continuity, rather than assuming that the .pdf container defines a single transaction.
Preserve multi-page continuity
Check continuation pages, repeated headers and totals. A later page may contain the final amount or line items missing from the first. Do not detach it merely because it has no repeated invoice number in the same location.
Keep cover sheets identifiable
A batch can include a scanner cover page, email printout or supporting receipt. Decide with the process owner which pages belong to each invoice and which are supporting material. Do not discard them automatically without a documented retention decision.
Compare one complete invoice with the combined batch
Create an approved working copy containing one full harmless invoice and run a controlled test. Compare its fields and table output with the expected answer. Then inspect the earlier combined-batch result to determine whether the failure concerned document boundaries or recognition within a single layout.
The flow action guidance describes a Pages parameter and warns about page ranges spanning more than one form. A range can select a complete form for processing, but an arbitrary range cutting through two invoices is not a substitute for correct segmentation.
| Input shape | Preparation question |
|---|---|
| One invoice on several pages | Are all continuation pages included? |
| Several invoices in one PDF | Where does each business document begin and end? |
| Cover sheet plus invoice | Which pages contain extraction data and which are retained context? |
| Mixed or unclear boundaries | Who reviews the split before downstream processing? |
Design a review path for uncertain boundaries
If automation cannot confidently determine where documents begin and end, route the batch to an authorized reviewer. Do not infer a universal split rule from one supplier's layout. A mistaken boundary can combine one supplier's header with another invoice's totals.
Keep source-to-output traceability
Record which original batch and page range produced each derivative. Use internal references rather than sensitive details in filenames. This lets a reviewer return to the source if an extracted line seems inconsistent.
Avoid overwriting the source
Create clearly named working derivatives through the approved PDF workflow. Preserve the received file when the records process requires it. Splitting for processing should not silently replace the original evidence of what arrived.
Check every resulting document
Before enabling production writes, verify page count, invoice identity, complete table rows and totals for the sample derivatives. Include a two-page invoice followed by a one-page invoice in the acceptance set. That test is more useful than three unrelated single-page samples.
For an approval process downstream, our SharePoint workflow guide explains how a review request relates to document approval. Make sure the approver receives the intended complete invoice rather than a partial extraction or an accidentally combined batch.
Processing activity can consume licensed capacity, including pages without useful extraction data. Review the current commercial terms before scaling; do not assume splitting or page selection makes the service free or establishes a specific monetary saving.
Have the reviewer confirm each derivative's final page before approving the sample set.
Document boundaries and selected physical pages must agree. The related page-selection guide is Document Intelligence Pages Parameter: Page Positions.
Keep a small segmentation policy describing complete documents, supporting pages and exception handling. Accept the workflow when each extracted record can be traced to one intended complete invoice and ambiguous batches stop for review rather than producing plausible partial data.
