Tool
Attachment text recognition
Scanned documents can be read on your own computer. Uncertain words are marked rather than silently accepted.
Scanned does not mean unreadable. Uncertain means marked.
Recognition runs on your machine. Every word carries its confidence. Whatever is uncertain is shown, and nothing is applied silently.
From the scan file to the checked reference
Six steps. Pick one to read it.
Scan
A PDF or image from the attachment. The document stays on your machine and is not uploaded for recognition.
Recognition
Text recognition runs locally. Every recognised word gets a confidence from 0 to 100: how sure the recognition is of that word.
Mark uncertain
Words below 60 are marked. If the confidence is missing, the word counts as uncertain, otherwise it would never be checked. You can move the cut-off below.
Suggestion
The AI suggests a correction for marked words. Only typical confusions are allowed, such as O and 0, l and 1, S and 5. Redaction runs first, and the AI only starts once it is approved.
Review
Every suggestion passes six checks (below). Only what passes is offered. You apply each suggestion yourself.
Reference
The attachment comparison shows where the amount sits in the document. The source stays visible, even when recognition was uncertain.
Where the recognition doubts
The bars show the confidence per word. Move the cut-off to see which words would be marked. Synthetic sample data.
5 of 12 words are below 60.
Rechnung96Nr.91RE-2026-041788Lieferant741.25O,0041EUR95Datum941O.03.202652Rechnurg471,5OO3612,5058Büromaterial81
A suggestion must pass six checks
The checks apply to every suggestion from the AI. The examples come from the app's checking code, with synthetic data. Result: 1 of 6 suggestions passes.
-
Passes
1.25O,00→ 1.250,00The word is uncertain. O is read as 0, and the amount stays 1.250,00.
- (a) Word was uncertain and was queried
- (b) Only an allowed confusion: O to 0
- (c) At most 3 characters difference
- (d) Wording changes
- (e) No new letters
- (f) Amount keeps the same value
-
Rejected
1O.03.2026→ 10.03.2026The correction looks right. But a date is checked as an amount, and 10.03.2026 is not a valid amount. So the suggestion is dropped.
- (a) Word was uncertain and was queried
- (b) Only an allowed confusion: O to 0
- (c) At most 3 characters difference
- (d) Wording changes
- (e) No new letters
- (f) A date is checked as an amount and is not one
-
Rejected
Rechnurg→ RechnungHere one letter changes (r to n). The checks do not allow that, even if the correction were right. Letter errors stay with you.
- (a) Word was uncertain and was queried
- (b) Another letter changes
- (c) to (f) not checked
-
Rejected
1,5OO→ 1500The value is not unambiguous: 1,500 has three decimal places. The proposal 1500 would change the amount.
- (a) Word was uncertain and was queried
- (b) Only an allowed confusion: O to 0
- (c) At most 3 characters difference
- (d) Wording changes
- (e) No new letters
- (f) Value is not unambiguous, 1500 changes it
-
Rejected
12,50→ 1.250Comma and point swapped: 12,50 would become 1.250,00, which is a hundred times the value.
- (a) Word was uncertain and was queried
- (b) Only an allowed confusion: separator
- (c) At most 3 characters difference
- (d) Wording changes
- (e) No new letters
- (f) The amount changes its value
-
Rejected
Lieferant→ LieferantenThe word was not uncertain, so it is never suggested. In addition, a letter would be added.
- (a) Word was not uncertain
- (b) to (f) not checked
What the AI sees has been redacted first
Before a passage goes to the AI, identifiers are redacted. The switch shows the example. On the right is what is covered and what is not.
Contact: [email protected], Tel. +43 1 234 5678, IBAN AT61 1904 3002 3457 3201, UID ATU12345678, date 10.03.2026, amount 1.250,00 EUR, document RE-2026-0417, Firma Muster GmbH, Musterstraße 12, 1010 Wien
Blue: will be redacted. Yellow: stays, because redaction does not recognise this item. Date, amount and document number stay on purpose, the check needs them. Sample text.
- Email addresses
- IBAN
- VAT ID (AT and DE)
- Company register number
- Tax number
- Social insurance number
- Credit card numbers
- IP addresses
- Phone numbers
- Names in free text
- Addresses
- Company names in free text
Measured on a published annual report
We checked recognition on 26 pages of a published group financial statement, about 1,600 numbers. The reference was the text of the PDF itself. Around 96 to 98 % of the numbers were read exactly.
The rest are not silently accepted. Some errors, such as lost decimal commas, are marked instead of passed on. What remains is mostly footnote digits and years, which are not amounts. A second selection of pages, not used to set the rules, landed in the same range.
What recognition cannot do
- Measured on one documentThe numbers come from a single annual report. Other scan qualities, such as faxes or photos, have not been checked. The confidence is an indication from the recognition, not proof that a word is right.
- This page computes nothing liveNo recognition runs here. The examples are precomputed, use synthetic data and the same checking code as the app.
- Page limitRecognition covers at most 200 pages per PDF. The remaining pages stay unchecked, and the app says so.
- Dates in correctionsCorrections to dates are currently rejected, because they are checked as amounts. They are then not offered as suggestions.
- Letter errorsThe AI does not correct letter errors. Only confusions of digits and similar characters are allowed.
- Redaction is limitedNames in free text, addresses and company names are not recognised by the redaction. Check excerpts with such details yourself before an AI request.