Automatic Document Separation: Complete Guide 2026

How automatic document separation works: AI reads each page of a batch scan, finds where a new document starts, keeps multi-page documents together and names every file by content.

A 50-page scan of mixed mail is useless until someone decides where each document starts. Automatic document separation hands that decision to software: it reads every page, marks the ones that begin a new document, keeps continuation pages with their first page, and names each resulting file by what's printed on it, for example Contract_Northwind-Ltd_2025-01-15.pdf or INV-2025-0281_Acme-Corp.pdf. This guide explains how the detection works in Docusplit, which fields the AI reads, and where it still gets things wrong.

Separation by content, not by page count

Classic PDF splitters cut every N pages or at bookmarks. They don't know that page 7 is the back of page 6, so a mixed scan defeats them. Content-based separation looks at each page the way a clerk would: is there a new letterhead, a new date, a new subject line, or does this page carry on a table from the previous one? The result is one file per document, however many pages each document has, and the first page of each document decides its name.

How the detection works, page by page

Every page is read as an image

The tool renders each PDF page to an image and sends it to a vision model (GPT-4.1 via the OpenAI API). No separate OCR step is involved and no text layer is needed, which is why a flatbed scan, a phone-scan PDF and a born-digital PDF are treated the same. Per page, the model answers a short set of questions; in document mode: is this an invoice, what type of document is it, who sent it, what is its date, and does it continue the previous page.

Continuation pages stay with their first page

For every page after the first, the model also sees a one-line summary of the previous page and decides whether the current page continues it. The cues are the ones a human uses: "Page 2 of 3", no new letterhead, a sentence or table that carries on, a page holding only totals and a signature. Pages recognized as continuations are appended to the file before them, so a three-page contract comes out as one file, and so does a two-page invoice whose second page only lists line items. No separator sheets or barcodes are needed.

Each document gets a name from its content

In document mode, the first question per page is whether it belongs to an invoice. Invoices are named InvoiceNumber_Company.pdf. Everything else gets Type_Sender_Date.pdf: a short type label the model writes itself, the company or authority on the letterhead, and the document's own date in ISO form.

Input:  scan_2025-01-20.pdf (50 pages, document mode, no folders)

Output:
INV-2025-0281_Acme-Corp.pdf            (2 pages)
Contract_Northwind-Ltd_2025-01-15.pdf  (3 pages)
Notice_IRS_2025-01-08.pdf
Reminder_City-Utilities_2025-01-12.pdf
Letter_Blue-Harbor-Dental_2025-01-10.pdf
...
Document_Page_37.pdf                   (nothing readable)
_overview.csv
_metadata.json

In invoice mode the classification step is skipped: every page is treated as part of an invoice, and only invoice number, company and invoice date are read. That's why invoice mode runs faster, at about 6 seconds per page against roughly 10 in document mode, by the app's own estimate.

Which document types are recognized

There is no fixed list, because the type label is written by the model from what it sees. A collection letter becomes Collection_letter_..., a termination Termination_..., a utility reminder Reminder_.... In practice the labels cluster around the usual office mail: invoices and credit notes, quotes and order confirmations, delivery notes, contracts, letters and notices from authorities (Notice_IRS_2025-02-28.pdf, Letter_HMRC_2026-02-10.pdf), bank statements, pay stubs, receipts, insurance documents.

What the AI reads per document: type, sender and date, plus invoice number, company and invoice date for invoices. Amounts, tax, account numbers and line items are not extracted. Labels come out in English on the English interface; the documents themselves can be in any language the model can read, with German, English, French, Spanish, Italian and Dutch being the tested ones.

What typically goes wrong

Printed text on a decent scan is the happy path; 200 dpi works, 300 dpi is the safe choice. Outside it, the failures are predictable:

Handwriting and faint thermal receipts: nothing readable, so the page becomes Document_Page_N.pdf, the CSV says "Not detected", and with folders by month it lands in Unknown/.
Blank back pages from duplex scanning: usually attached to the document before them, occasionally a file of their own.
Attachments without a letterhead (terms, a map, a price list): can be glued to the wrong document or split off on their own.
A delivery note directly behind its invoice: in document mode it's classified separately; in invoice mode it's read as an invoice and named after its document number.
Sender-name variants: "Acme Corp" on one letter, "Acme Corporation Inc." on the next, which matters if you sort into sender folders.

There is no correction step inside the tool, no preview of the detected boundaries and no hand-editing before download. You check the CSV and the fallback-named files first, fix the handful by hand, and you're done; for a 50-page batch that's typically a minute of work.

Scan settings that help

Feed the whole stack through an automatic document feeder as one PDF; sorting beforehand is unnecessary. Scan at 200 dpi or better, grayscale for printed pages, color when receipts are on tinted paper, and keep pages straight. Export as PDF, not JPG. The accounting-specific routine, from shoe box to bookkeeper, is in Digitize accounting documents.

Choosing a separation tool

Four questions separate the tools on the market:

Does it split by content or only by page count, bookmarks or barcode separator sheets?
Does it name by content or leave you with part-1.pdf, part-2.pdf?
What happens to unreadable pages: a silent guess, or a marked fallback you can find?
What happens to your data: look for clear deletion and processing terms.

Other tools now split with AI too: iLovePDF added its Smart Range mode in 2026, and AutoSplit for Acrobat or DATEI-BUTLER cover rule-based setups. The difference with Docusplit is that splitting and naming happen in the same upload, without templates or rules. One limitation to know: the output PDFs have no added text layer, so if your archive needs full-text search, run OCR in the DMS afterwards.

Frequently asked questions

How accurate is automatic recognition?

There is no measured accuracy figure to quote. What you can rely on is the fallback: a page the AI can't read is saved as Document_Page_7.pdf, shows "Not detected" in the CSV overview and, with folders by month, lands in Unknown/. Printed text on a scan of 200 dpi or better works well; handwriting and faded thermal paper are the usual failures.

Does it work with handwritten documents?

Only partially. Printed text is read reliably, handwritten additions are mostly ignored, and a fully handwritten page usually comes back as Document_Page_N.pdf for you to name yourself.

Are my documents processed securely?

Files are transferred over TLS, kept only while processing and then deleted from our servers. AI analysis runs through the OpenAI API (data processing agreement with OpenAI Ireland, Standard Contractual Clauses, never used for training; OpenAI retains API data for up to 30 days for abuse monitoring). See our privacy policy for details.

What file formats are supported?

PDF only, up to 50 MB per file. Convert JPG or PNG scans to PDF first; most scanner apps do that on export.

How many pages can be processed at once?

There is no per-PDF limit; only your page quota applies: 30 free pages once after creating an account, then 100, 300 or 1,000 pages a month on Starter, Pro or Max.

Try it on a real stack

Handwritten pages are the hard limit; anything printed is worth a run. Upload your last batch scan in document mode and check the CSV against the pile: the first 30 pages are free after creating an account, no credit card needed. Start for free

Related articles: