How to Split Scanned Documents Into Separate Files

TL;DR: To split scanned documents automatically, something has to read the pages and find where one document ends and the next begins. Separator sheets, barcodes, and "split every 4 pages" all depend on preparation nobody does consistently. A Split Document step in The Drive AI takes a plain-English instruction such as "split into separate invoices" and creates a separate file for each document it finds.

Somebody puts forty pages in the feeder and presses scan. What comes out is one PDF holding eleven invoices from seven vendors. Or a single 300-page file that is really a whole loan file: the note, the deed, the disclosures, the title policy, the insurance binder, with no divisions.

Nothing is wrong with the scan. The problem is that a scanner produces files and a business needs documents, and one batch is not one document. Until somebody separates them, that PDF cannot be filed, found by vendor, or attached to the right record.

Document separation is the step that turns one multi-page scan into the individual documents it contains, so each can be named, filed, and found on its own.

How document separation has traditionally been solved

Four methods have been used for decades, and each works until it does not.

Blank page separators. The scanner watches for a blank page and starts a new document when it sees one. Free, and no special paper. It fails in both directions. Somebody forgets a sheet and two invoices merge into one file. A contract with a page marked "this page intentionally left blank" splits itself in half. Duplex scanning turns the back of every single-sided page into a blank, so the detection threshold gets raised, and then a light receipt is swallowed as a separator.

Barcode or patch code separator sheets. Printed sheets carrying a patch code, dropped between documents. This is the most reliable method here, and the barcode can carry a document type or an index number, so separation and classification happen together. The cost is preparation. Someone prints the sheets, keeps a stack at the scanner, and inserts the right one in the right place, and the person at the feeder is often not the one who knows where the boundaries are. None of it exists for scans that arrive as email attachments, from a phone, or from a client who scanned a whole folder at once.

Fixed page counts. Every statement is four pages, so split every four pages. Instant, and correct exactly as long as the assumption holds. One vendor sends a second page of line items, and every document after it is cut in the wrong place.

Manual splitting in a PDF editor. Open the file, scroll the thumbnails, select a range, extract, save, name it, repeat. Accurate, and what most teams do. It also costs three to six minutes per batch, forever. For one-off work that is fine, and a PDF merge tool or a compressor handles the cleanup with no setup.

A fifth option, template-based separation, treats any page with a logo or the word "Invoice" in a configured zone as a first page. It is configured per layout, so every new vendor is a new configuration.

The pattern is the same across all five. Each asks a human to encode the boundary in advance, physically, numerically, or as a template. The boundary is already in the document: page 12 starts with a new letterhead and an invoice number that is not the one on page 11.

How does automatic document splitting work?

It reads the pages and decides where the document changes. In The Drive AI that is the Split Document step, and it is an ai step, meaning it opens the file and reads it.

The step asks one question: "How should the document be split?" You answer in plain English. "Split into separate invoices." "Split into its individual documents." "Split into one statement per account." The panel's own guidance applies: plain English is enough, and the AI works out the rest. It detects the boundaries and creates a separate file for each document.

No separator sheets, no page counts, no zone configuration per vendor. A batch of eleven invoices of one, two, and four pages splits into eleven files because the eleventh reads like a new invoice, not because page 26 was divisible by something.

Every step carries a cost badge, so the tradeoff is visible before the workflow runs. Split Document carries the ai badge: it opens the file, costs a little per file, and can occasionally be wrong.

What happens to the original, once split?

You choose. The setting is "The original, once split", with three options:

• Stays where it is. The scan remains in the folder it arrived in. Use this while you are still checking results, or when the batch scan is itself the record of what was received.
• Goes to trash. A recoverable soft delete, for once you trust the split.
• Goes to a folder you name, Review or Archive/Originals, where the as-scanned batch has to be retained.

The new files are created either way. The setting decides only the fate of the original.

Split, then read, then rename and move

Splitting alone leaves eleven untitled files instead of one. What makes it worth doing is the steps that follow, applied to each new file:

1. Split Document (ai) turns the batch into individual documents.
2. Read Details from File (ai) pulls named values out of each one. You name the fields: "get vendor, invoice number, and invoice date". Those values are then available to later steps.
3. Rename (ai) uses them: "with the vendor name and invoice number".
4. Move (ai) files it: "under Vendors, then the vendor, then the year".

Eleven invoices in one scan become eleven named files in eleven vendor folders, and nobody opens the PDF. The extraction half is covered in the post on extracting data from documents without templates, and the reverse operation in the post on merging documents automatically.

Rescanning a batch is common, so set Move's "If the same file is already there" option deliberately: Move it anyway (renamed if the name is taken), Leave it where it is, or Set it aside for review, which files it into a "Possible Duplicates" folder.

Example workflows for splitting scanned batches

Each is typed as one sentence. The steps then appear laid out on a canvas, where each can be edited.

A scanner folder of batched invoices:

When a PDF lands in Scans/Inbox, split it into separate invoices, get the vendor and invoice number from each, name each file with both, and file it under Vendors, then the vendor, then the year. Move the original scan to Archive/Originals.

A signed closing package, split into its individual documents:

When a closing package is added to Transactions/Inbox, split it into its individual documents, name each with the property address and the document type, and file them under Transactions/1420 Oak Street/Closing. Ask Dana to approve first, and keep the original where it is.

A monthly bank statement bundle:

Every month, go through Finance/Statements/Incoming, split each bundle into one statement per account, name each file with the account and month, and file them under Finance/Statements/2026.

An HR packet, split into its forms:

When an onboarding packet arrives in People/Onboarding/Inbox, split it into its individual forms, get the employee name and form name from each, name each file with both, and file it under People, then the employee. Send the original to trash.

Ways to split scanned documents, compared

Limits, and what splitting will not do

ai steps can occasionally be wrong. A page opening with a large logo can read like the start of a new document, and a continuation sheet with no header can read like part of the previous one. On ambiguous batches, expect the occasional bad cut.

That is what Ask a Person to Approve is for. Put one after the split and nothing downstream runs until someone decides. You set What should the approver decide? ("Do these eleven files look like eleven invoices?"), Who decides? (Admins, Specific people, or Everyone), and When to ask?: Always ask, or Only when unsure, where confident filing goes ahead and is logged while new folders, unsure matches, and anything destructive still ask. Approvals are not open-ended: Wait up to 24 hours, 3 days, or 1 week, then the file takes the Expired exit. The human-in-the-loop approval post covers the three exits and the caution presets.

Other honest limits. There is no OCR configuration step, no summarize step, no tagging step, and no translate step. Files in a connected Google Drive, OneDrive, SharePoint, or Dropbox folder are worked on in place inside that storage, and nothing is copied into The Drive AI unless you add a Save a Copy step. Workflows on uploaded files work on every plan; connecting a Gmail, Outlook, Slack, Teams, or cloud storage account needs the Max plan on a personal workspace, or any Team plan. Every file a workflow processes counts once against a monthly allowance.

Setting up a split workflow

1. Connect the account holding the scanner folder, or upload the scans to The Drive AI. For a scanner that writes into a shared drive, the Google Drive workflow post covers what a source watches.
2. Type the sentence: "When a PDF lands in Scans/Inbox, split it into separate invoices, name each with the vendor and invoice number, and file it under the vendor and year."
3. Open the Split Document step and check the answer to "How should the document be split?". Be specific: "split into separate invoices, one per invoice number" beats "split this up".
4. Set "The original, once split". Start with Stays where it is, and switch to Goes to trash only after a few dozen batches come out right.
5. Add Read Details from File, name the fields you need, and let Rename and Move use them.
6. Add Ask a Person to Approve, set to Always ask, Wait up to 3 days. After two weeks of clean results, move to Only when unsure.

Every run is recorded, can be replayed on the canvas exactly as it happened, and can be retried from the step that failed. That is how a splitting instruction gets tuned. For the wider picture, read the complete guide to automated document workflows.

Frequently Asked Questions

Can I split a PDF into separate documents without separator sheets?

Yes. Separator sheets exist because older scanning software could only detect a physical marker between documents. A Split Document step reads the pages instead, so the boundary comes from the content: a new letterhead, a new invoice number, a new form title. You describe what the batch holds, and the step creates one file per document.

What happens to the original scan after it is split?

You pick one of three outcomes in the "The original, once split" setting. It stays where it is, goes to trash as a recoverable soft delete, or goes to a folder you name such as Review or Archive/Originals. The new files are created in all three cases. Regulated work usually keeps the as-scanned batch, because it is the record of what was received.

Can it split a multipage TIFF scan, not just a PDF?

The instruction you write does not change by file type, so a multipage TIFF scan is described the same way as a PDF one: say what the batch holds and let the step find the boundaries. Older scanners still produce multipage TIFF, and the problem is identical. Treat the split output as the working files and archive whatever the scanner produced.

How accurate is AI document separation?

It is an ai step, so it reads the file and can occasionally be wrong, most often on pages that give few clues: a continuation sheet with no header, or a cover page with a large logo that reads like a new document. It is more accurate than fixed page counts on variable-length batches, and less predictable than a barcode separator.

Does splitting happen inside Google Drive or SharePoint, or does it copy files out?

A workflow that watches a folder or library works in place, inside that storage. Nothing is imported into The Drive AI unless the workflow includes a Save a Copy step, and then only to the place you name. Nothing outside the folder you chose is touched, and Save a Copy lets you choose whether later steps act on the copy or the original.

Start with one batch

Take the folder your scanner writes into and one week of its output. Describe the split and the filing in a sentence, keep the original where it is, and add an approval step so you see every result before it lands. Two weeks is enough to know whether the boundaries are read correctly. Create a free account or read the Workflows page.