You blacked out the names. Did you catch the account number on page nine?
You blacked out the names. You blacked out the signature. You sent the file.
Then, a week later, you notice the account number sitting in the footer of page nine, in grey six-point type, exactly where you never thought to look.
Redaction failures are rarely about the boxes you drew. They are about the ones you did not, because you are a person reading a long document with your eyes, and personal data does not politely gather itself on page one. It hides in headers, in a "reference: " line, in an email address inside a forwarded thread, in a date of birth three tables down. Finding all of it by hand is the part humans are worst at, and it is the part that gets people caught.
Why manual redaction misses things
When you redact by reading, you are running a search-by-attention over hundreds of small strings, and attention is a finite budget. You spend it on the obvious sensitive block at the top of the page and quietly run out before the boilerplate at the bottom. The document does not help you: a sixteen-digit card number and a sixteen-digit order reference look identical to a tired reader, so you either redact both or, more often, neither.
This is not carelessness. It is the wrong tool for the job. Catching every occurrence of a pattern across a whole document is exactly the kind of work you should hand to the machine, and exactly the work most redaction tools leave entirely to you.
Let the document scan itself
Our Redact PDF tool reads the text that is already in your file and proposes redaction boxes for the personal data it recognises, so the review starts from a checklist instead of a blank page. It works in two layers.
The first layer is a pattern scan. It looks for the structured information that has a recognisable shape: email addresses, card numbers, ID and Social Security numbers, IBANs and bank codes, phone numbers and dates. Card numbers get an extra check. A run of digits is only flagged as a card if it passes the checksum that real card numbers are built to satisfy, so an invoice number that happens to be sixteen digits long does not get boxed by mistake. Precision matters here: a scanner that flags everything trains you to click "dismiss" without reading, which is how the one that mattered slips through.
The second layer catches what no pattern can. A name is just words; an address is just more words. There is no regular expression for "this is a person." So when on-device AI is available, a second pass reads each page and proposes boxes over the people’s names and full postal addresses it finds. It is deliberately conservative: it can only ever suggest a box over text that actually appears on the page, so it cannot invent a redaction or remove something that was never there. You stay in control of every box; nothing is applied until you export.
The part every competitor gets wrong
Automatic PII detection is not new. What is rare is doing it without uploading the document first.
Think about what "AI-powered redaction" means almost everywhere else. To find the sensitive text with a server-side model, the tool has to send the text, or the whole file, to that server. So the very first thing that happens, before a single name is removed, is that your unredacted original is transmitted and processed somewhere you cannot see. You are handing over the raw version to get the safe version back. For a document sensitive enough to redact, that is precisely backwards.
Ours never makes that trade. The pattern scan is local. The name-and-address pass runs an AI model on your own device. The unredacted file, the version you are most worried about, stays in the tab the entire time and is gone when you close it. You do not have to take that on faith either: the live counter we put on the page measures the actual bytes leaving your browser, and it stays at zero while the scan runs.
Cleaning a document before you paste it into AI
There is a newer reason people suddenly need to strip personal data from a file, and it is not a lawyer or a leak. It is that they want to paste a document into ChatGPT or Claude to summarise it, and they do not want to hand a chatbot a customer’s address or their own bank details in the process.
That is the same problem in a friendlier hat, and it takes two steps. The scan above removes the personal data you can see. A PDF can also leak invisibly, through author metadata, comments, form-field data and hidden layers, so after redacting, run the file through Clean PDF for AI to strip those too. Redact the visible, clean the invisible, and what you paste is all you meant to share, without the file ever leaving your device. The irony is worth saying out loud: it takes local AI to make a document safe for someone else’s cloud AI. We wrote the full flow up in Redact a PDF before pasting it into an AI.
How to actually use it, in one pass
- Open your file in Redact PDF. It stays in your browser.
- Let it scan. Review the suggested boxes it found; tick the ones that are genuinely sensitive and dismiss the false alarms. Dismiss by type when a whole category (say, every date) is fine to keep.
- Add anything it could not know. A codename, an internal project title, a handwritten note in a scanned page: draw those boxes yourself.
- Export, then run the five-second test. Open the downloaded file, drag your cursor across a redacted area and press Ctrl+C. Nothing should be selectable. If words come out, it was covered, not removed, and you are reading the wrong kind of black box.
The short version
The dangerous part of redaction is not applying the boxes. It is finding everything that needs one, across a document too long to read perfectly, in places you would never think to check. Hand that search to a scanner that reads the whole file for you, keep the final say over every box, and insist that the finding happens on your device rather than on someone’s server, because a tool that uploads your file to protect it has already lost the thing you were protecting. Try Redact PDF.
Frequently asked questions
Does the scanner upload my document to find the personal data?+
No. The pattern scan runs in your browser over the text already in the file. The name-and-address pass uses an AI model that runs on your device too, so the document never leaves your machine at any point. Most other "smart redaction" tools do the opposite: they send your file or its text to a server to analyse it, which means the unredacted original is uploaded before anything is removed.
What kinds of information can it find automatically?+
Two layers. A pattern scan catches structured data no human should have to hunt for by eye: email addresses, card numbers (checked with the same math a real card uses, so it does not flag random digits), ID and Social Security numbers, IBANs and bank codes, phone numbers and dates. When on-device AI is available, a second pass finds the things patterns cannot, mainly people’s names and full postal addresses.
Will it black out the wrong things or invent text?+
It only ever proposes a box over text that literally appears in your document. Everything is a suggestion you review, tick or dismiss before anything is applied, and nothing is redacted until you export. The AI cannot cause text to be removed that was not really on the page.
Once I apply the boxes, is the text really gone?+
Yes. On export the content under each box is removed from the file, not just covered, so it cannot be selected, searched or extracted afterwards. That is the difference between redaction and a drawn black rectangle.