Learning from your corrections

Every time you fix a field that the AI got wrong and confirm the ones it got right, you're not just cleaning up one document — you're teaching Scanix how your paperwork is laid out. This page explains how that learning works, what you do to benefit from it, and where its limits honestly lie. There are no extra buttons to press: the only thing the learning loop asks of you is the work you were already doing.

The idea in one minute

When an AI module reads a document, it doesn't just guess from nothing — it works from examples of what "right" looks like. The learning loop (sometimes called few-shot learning) takes the corrections your operators make and turns them into those examples. A correction you verify today quietly becomes a worked example that guides the AI on the next, similar document.

Think of it the way you'd train a new colleague. The first few invoices from an unfamiliar supplier, they might put the wrong number in the Invoice total box. You correct them once or twice, they see the pattern, and from then on they get that supplier right. Scanix learns the same way — from the same corrections — except it never gets tired and never forgets.

What you do to benefit

The loop runs on the everyday review work, so there's almost nothing new to learn:

• Correct the fields that are wrong. When you Analyse a document with AI, the extracted values land in the Fields panel. Fix any that came back wrong or empty, exactly as you normally would.
• Verify the fields that are right. Confirming a value is a signal too — it tells Scanix "yes, that's the pattern." Documents whose fields all clear your confidence bar can be accepted automatically; the rest are held in the Verification Queue for you to review and confirm.
• Keep going. Each verified document adds to the pool of examples. The benefit compounds quietly over a batch and over time — you don't switch anything on, and there's no "train now" step to remember.

Why it gets the next document right — without copying the last one

A fair question: if Scanix is learning from past documents, won't it just paste yesterday's values onto today's invoice? No — and the distinction matters.

The learning loop teaches the AI about the mapping: where on this kind of document the total tends to sit, what your supplier calls the invoice date, which label introduces the VAT number, how your forms are structured. It does not memorise the values. Every new document is still read on its own — the AI extracts that document's own number, that document's own date, that document's own amounts. What it carries forward from your corrections is the know-how to find them, not the answers themselves.

So a sharper model means fewer documents land in the Verification Queue and more sail through correctly the first time — not that two different invoices come back with the same numbers.

Keeping your expectations honest

Learning helps most where there's a pattern to learn, and it isn't magic. A few things worth knowing:

Next steps

Related

• Analyse a document with AI
• Edit & correct recognized text