Reviewing and correcting extractions

Every field the model reads carries a confidence and, where relevant, a reason it was flagged. The review screen is where you fix what it got wrong — and it's built so that fixing something is never lost to a later re-extraction.

The review screen

Alongside the extracted fields sits the original image, so you can check a figure against the actual document without leaving the screen. A document that needed review shows exactly which check failed — a mismatched line total, an ABN that doesn't check out, a date that looks wrong — rather than a generic "low confidence" notice.

Making a correction

Edit any field directly. What the model originally said is never overwritten or discarded — it stays attached to the record alongside your corrected value, so there is always a full trail of what was read, what it was changed to, and who changed it.

Why your edit always outranks a later re-extraction

Extraction can be re-run later — over your whole history, if the model improves. Without a safeguard, a re-run could silently overwrite a field you had already fixed, which would waste the attention you gave it. So once you correct or confirm a field, it is locked: a later machine run can disagree with it, and if it does that's raised for you to look at, but it can never overwrite it on its own. A machine run can still add new flags for fields it hasn't seen before — it just can't undo a human decision.

Rejecting a mis-scan

Photographed the wrong thing, or a duplicate you don't want to keep as a separate record? Reject it. The original image is retained under the same retention rules as everything else, but it's taken out of your working list and never proposed as a transaction.

Nothing you see is ever lost

Once a document has been confirmed into a posted transaction, extraction changes stop silently altering it. If a re-extraction would touch a posted transaction, it raises a review task for you instead of rewriting your books behind your back.