Classification & AI splitting

When you feed Scanix a stack of mixed paperwork, two things have to happen before anything can be extracted: each document has to be recognised for what it is, and the stack has to be broken into the right individual documents. This page explains how Scanix does both — automatic classification and AI document splitting — and how they hand off to a Job Group so the rest of the pipeline knows where every page belongs.

This is a concept page. When you're ready to set it up, follow the linked how-tos: How routing works for the Job Group side, and Document splitting for the per-template rules.

Two problems, one goal

Picture the everyday case: five invoices from three suppliers plus a couple of receipts, all scanned in one pass. To process that batch without pre-sorting it by hand, Scanix has to answer two questions for the pile:

• Where does each document begin and end? (splitting)
• What type is each one, so the right rules apply? (classification)

Get those right and everything downstream — OCR, field extraction, and filing — falls into place. Get them wrong and pages land under the wrong type or run together. Scanix gives you separator-based splitting for predictable batches and AI-driven recognition for messy ones, and you can lean on whichever fits the work.

Automatic Document Recognition

Automatic Document Recognition is how Scanix decides which template a document belongs to based on its content, rather than asking you to sort the stack first. Inside a Job Group, each member template carries its own splitting rule, and Scanix walks the member list from top to bottom for every page: it asks member 1 "does your rule match this page?", then member 2, and so on. The first match wins, and the document is routed to that template. The order you give members is the routing priority — put the most specific matchers near the top and any catch-all at the bottom.

Recognition leans on each template's own splitting configuration (a barcode, a patch code, a fixed page count, and so on), which is why the matchers live in the Template Designer, not in the Job Group. A page that no member claims is set aside in the review bucket rather than being forced into the wrong type.

AI document splitting

Sometimes there are no separators at all — no barcode pages, no patch sheets, just a continuous scan where one document runs straight into the next. That's where AI document splitting comes in: instead of looking for a marker, the AI reads each page's text and works out where one logical document ends and the next begins, then breaks the stack into separate typed documents. A run like [invoice, invoice, two-page delivery note, receipt] comes apart into four documents, each with its own pages.

Because the AI needs text to reason over, splitting runs after a first OCR pass. That OCR isn't wasted — it's reused for the extraction step, so a document is never read twice. In an AI-only group (one with AI Services configured but no member templates), this boundary detection is the splitting strategy: Scanix OCRs the batch and runs AI boundary detection to carve the mixed scan into typed segments, with no template routing in between.

How classification feeds Job Group routing

Classification and splitting aren't standalone tricks — they're the first stages of the Job Group pipeline, and everything after them depends on getting the type right. When you run a group, Scanix moves through these stages in order:

1. Split and route — pages are separated and routed to the matching member template by priority (first match wins), or, for an AI-only group, boundary-detected into typed segments.
2. OCR — text is read for each resulting document.
3. AI-classify the unrecognised (optional) — see below.
4. Extract — each configured AI Service runs on the documents it applies to.
5. Resolve cross-template mappings — values that flow between sibling documents are filled in one sweep.
6. Stage for review — documents are placed in their type buckets, ready for you to check.

Nothing is written to disk during this run. Export happens only when you review the results and click Process — that's the moment files reach their destination. For the full routing model, see How routing works.

When no matcher claims a document

If a document doesn't match any member's splitting rule, you can let the AI step in. The Members tab has an AI-classify unrecognized documents toggle, off by default. Turn it on and Scanix uses AI to pick the best-matching member template for any document the matchers missed — choosing between, say, an invoice and a receipt — and its bucket flips from Unrecognized to the matched type.

It's off by default for a reason: it spends AI tokens on every unrecognised document. As the in-app description puts it, leave it off to rely only on your configured matchers. Reach for it when your batches are too varied for fixed rules to catch everything.

Confidence and what you review

Classification and extraction aren't all-or-nothing — Scanix tracks how sure it is and lets you decide where the bar sits.

• Per-template match confidence. A member template's auto-match confidence is adjustable from 50% to 99%. Lower means looser matching (more documents claimed, with more risk of a wrong call); higher means stricter.
• Auto-accept threshold. In Settings → AI Services, the Auto-accept above N% confidence setting (default 97%) governs straight-through processing: a document whose fields all clear the bar can export automatically, while anything below it is held back for a person to check.

Either way, nothing slips through unseen. After a group run, documents are staged in type buckets for review, and anything Scanix couldn't place confidently lands in a dedicated review bucket — Unrecognized for documents in a routed group that no matcher claimed, or Unclassified for documents in an AI-only group that the classifier couldn't confidently route. From there you can reassign a document to the right type before exporting.

How this differs from separator splitting

It's worth keeping two kinds of splitting straight, because you'll meet both:

• Separator-based splitting lives on a single template, in the Template Designer. You tell Scanix to break the stack on a Barcode, Patch Code, Blank Page, Fixed Pages, Zonal OCR, or Specific Doc marker. It's exact and predictable, and it uses no AI. See Document splitting.
• AI document splitting lives inside a Job Group run. It reads the pages themselves to find boundaries when there are no markers to rely on, and it pairs naturally with classification to route each piece to the right type.

Use separators when your batches are structured and you control how they're prepared; lean on AI splitting and classification when the stack is mixed and unpredictable.

Next steps

Related

• How routing works — how a Job Group decides each document's type, and the read-only routing health check.
• Document splitting — separator-based splitting rules configured per template.
• AI capture — where classification and splitting fit among Smart Templates, AI Services, and the learning loop.