DE EN
Intro call

Redaction with human approval

The AI proposes. Your team decides.

The feld.ai agent finds personal data in PDFs and scans, including logos, stamps and signatures. Your team sees every finding on the document and decides what gets redacted. Only the download creates the redacted copy — technically removed, not covered.

Document redaction: the copy test, before and after (fictitious demo data)

A black bar is not a redaction

Many “redacted” PDFs still contain the text — underneath the bar, copyable with Ctrl+A. In file access requests, subject access requests or disclosure to third parties, that is not a cosmetic flaw but a reportable incident. And anyone who redacts by hand knows the page-247 risk: one missed occurrence in a 300-page file.

Both failure modes — the invisible residual text and the missed occurrence — are what this workflow is built against.

You can run the test from the video yourself: open one of your redacted PDFs, select the area over the bar, copy it and paste it into a text field.

The principle: three steps, three responsibilities

1
Detection is probabilistic
The AI finds names, addresses, identifiers and your own custom fields — measured so far at over 93% precision, up to 99.5% depending on document type. Nobody honestly promises you 100%.
2
Deciding is human
Precisely because of the remaining 0.5–7%, the agent shows your team every occurrence directly on the original. Your team decides what gets redacted.
3
Removal is deterministic
On download, every finding in the ticked categories is technically removed from the file, text and pixels. The redacted text can no longer be extracted afterwards: the copy test in the video finds nothing.

You see exactly what goes out

This is what your team sees before the download:

  • All findings are listed next to the document, sorted by category, with hit counts.
  • One tick per category decides: untick what may stay. Preview and download only redact what is ticked.
  • Original, Redacted and Compare: the slider shows before and after on the same page.

Two files per case

The unchanged original

Preserved and untouched — the reference for every decision.

The redacted output file

Technically cleaned — no text under the bar.

What the agent does not do

  • It passes nothing on by itself: a redacted file is only created when someone requests it.
  • It does not promise 100% detection — instead, a workflow that handles the remainder.
  • It does not judge what legally has to be redacted in your specific case. That stays with your responsible team; the agent applies your rules.

Facts

Operations & data location

feld.ai infrastructure in Feldkirch, Austria. Your documents never leave the agreed hosting location — no third-party providers in the processing path.

Information security

ISO 27001: in certification process.

Measured precision

Over 93% across all fields measured so far, up to 99.5% depending on document type — indicative. The number only becomes reliable on your own documents, which is exactly where we measure it in the demo.

Which documents do you need to redact?

30 minutes, ideally on your own sample documents: we cover document types and rules — and you see the measured precision on your material, not ours.

Book an intro call