AI capture

AI capture is the side of Scanix Desktop that reads your documents the way a person would: it works out what each document is, pulls the fields you care about, separates a mixed stack into individual documents, and gets steadily better as you correct it. Instead of drawing zones for every layout, you point Scanix at a document type and let the AI fill in the fields.

This page explains the mental model — what AI capture does, where it runs, and how it stays private. When you're ready to set it up, each section links to a focused how-to.

What AI capture does

Think of AI capture as four jobs that run together on a stack of pages:

• Classify — recognise what each document is (an invoice, a passport, a receipt) so the right extraction rules apply.
• Extract — pull the fields for that document type and return them in clean, canonical form: dates as YYYY-MM-DD, money as an amount plus a 3-letter currency code, phone numbers in international format, IBANs checksum-validated, and so on.
• Split — given a single stack of mixed documents with no separator sheets or barcodes, read each page and decide where one document ends and the next begins, then break the stack into separate typed documents.
• Learn — feed the corrections your operators make back into future runs, so extraction adapts to your layouts, languages, and field labels over time.

The building blocks you'll meet across the docs are AI Services (pre-built extraction schemas for a document type), Smart Templates (the AI builds a whole template from one sample), Analyse with AI (a one-click Viewer action), and Job Groups (mixed-batch automation). They all draw on the same catalogue and the same provider settings.

AI Services: the document-type catalogue

An AI Service (also called a module) is a ready-made extraction schema for one class of business document. It already knows which fields to read and how to normalise them, so an operator never draws a zone — they pick the matching service, run it, and the fields come back filled.

Scanix ships 27 built-in services today, grouped in Settings → AI Services under Finance & Accounting, Procurement & Logistics, HR & People, Legal & Governance, Identity & Compliance, and Other. They cover everyday document types — Invoices AI, Receipts AI, Identity Documents AI, Bank Statements AI, Contracts AI, and many more — plus one special service, Natural Language Query, which has no fixed schema: you type a request in plain English ("Find the customer email, the order number, and the shipping address") and it builds the schema at run time.

Several services also run cross-checks after extraction. An invoice's line items must sum to its total; a bank statement's opening balance plus credits minus debits must equal its closing balance; an ID document's printed fields must match its machine-readable zone. Failed checks surface in the field-verification panel so a person can take a look.

The catalogue is fixed in the build — you can't invent a new service type or rewrite a service's fields from the UI — but you can turn whole services on and off, and override whether individual fields are required.

Where AI capture shows up

The same services appear in a few places, each suited to a different job:

• On a template — bind one service to a template so it runs automatically every time that template runs. In the Viewer, the primary button changes from OCR to Analyse with AI.
• In a Job Group — select several services for a batch that mixes document types; the group splits the stack, routes each piece to the right template, and runs the services across all of them.
• Ad-hoc in the Viewer — run a service on a one-off import with no template applied, straight from the Analyse with AI button.
• A whole template, built for you — a Smart Template has the AI look at one sample page and propose the document type, extraction zones, index fields, a splitting strategy, and default outputs. You review, toggle items on or off, and save — no manual zone-drawing.

Bring your own cloud — or run it entirely on-device

This is the part that matters most for privacy and cost. Scanix is bring-your-own-cloud: when a service uses a cloud model, the request goes straight from your machine to your own Anthropic or OpenAI account. Scanix is never the account holder — it never sees your API key, never sees your documents on that call, and never bills you for tokens. You pay your own provider, and token counts are shown locally for your own visibility.

You can also run extraction fully on-device with a local model, so documents never leave the machine at all. That suits sensitive material — for example, Medical Forms AI values are treated as protected health information and pair well with an on-device-only setup.

A routing policy decides, per template or per Job Group, which model handles each document:

A Privacy mode: local only setting overrides everything above — it hard-blocks any document from leaving the machine, even if a routing policy would otherwise allow the cloud. When cloud calls are allowed, you must explicitly tick I acknowledge cloud egress before Scanix will make them.

Straight-through processing and the learning loop

AI capture is built to run hands-free where it can. In Settings → AI Services, an Auto-accept above N% confidence threshold (default 97%) decides which results sail through: documents whose fields all clear the bar export automatically, while the rest are held for a person to review. Raise the bar for more checking, lower it for more automation.

Behind the scenes, the corrections operators make become future examples. The model learns this customer's layouts, languages, and field labels over time — but it always reads each new document's own values rather than memorising answers. The more you use it, the less you correct.

When to use AI capture vs a template

AI capture and classic templates aren't rivals — they solve different problems, and the best setups combine them.

In practice many operators do both: a template captures the few fixed zones it can be sure of, and an AI Service fills the rest. A Smart Template is a fast way to get that hybrid started from a single sample.

Next steps

Related

• Templates & the Designer — define zones, index fields, and anchoring by hand.
• Job Groups — split, route, OCR, extract, and export a mixed batch in one pass.
• OCR — recognise text and produce searchable PDFs, with or without AI.