How accurate is untxt., and how review works
Accuracy depends on the input, so untxt. does not hide behind one number. It scores its confidence on every document, reads clear printed content reliably, and flags anything it is unsure about, with the original beside the extracted data so you check only the few that need it.
Why a single accuracy number is the wrong question
Any tool can quote a percentage on clean invoices. Your inputs are not clean. What matters is not the headline figure, it is whether the tool is confidently wrong or honestly unsure, because a confident wrong value is the one that slips into the books unnoticed. untxt. is built to surface the second kind, not bury it.
Confidence on every document
untxt. attaches a confidence score to each document as it reads it, so the review overview shows where attention is needed and where it is not. High-confidence documents move fast; low-confidence ones are surfaced, not hidden in the pile.
The review queue: original on the left, data on the right
Open a document and the original sits on the left, the extracted data on the right. Every field is editable. You confirm or correct in place and move on, reviewing exceptions instead of re-keying the whole batch.
What lowers confidence
Faded scans, handwriting, and unusual layouts come through with lower confidence and get flagged. untxt. tells you when it is unsure rather than filling the field with a guess, so uncertainty is visible instead of silent.
Accuracy you can audit
Because every document carries a confidence score and a review trail, you can see how much cleared untouched and where your time actually went, rather than trusting a vendor's demo-file percentage.