Quick answer Advice on prompt redaction covers what you type, not what you upload. When a whole PDF, case file or contract is pasted into or attached to an AI tool, the names, dates of birth and account numbers inside it go along too, in one step, with no highlight to catch them. BeeSensible opens the document first, in the extension's side panel, in the dashboard or in the desktop app, highlights the sensitive parts on the page and produces an anonymised copy to share or paste in, before any of it reaches an AI tool.
A caseworker has a ten-page PDF and wants a two-line summary. She opens ChatGPT, uploads the file and asks for the main points. Done in ten seconds.
That file contained a client's full name, date of birth, national ID number and a paragraph about a family situation that was never meant to leave the case management system. None of it was typed. It went in as one upload, in one step, with nothing to catch it.
#The advice you have read is about typing, not uploading
Guides on safe AI prompts tell you to remove names, replace identifiers, mask account numbers. Good advice, and BeeSensible's own browser extension is built around exactly that moment: the text field, while you type, where you mask or remove a value before you send.
But a PDF upload does not go through a text field. The whole document goes in at once, formatting and all. Whatever prompt-hygiene habits you have built up for typed text do not apply, because there was no typing to catch.
#What is actually in that file
A single case file, contract or report routinely carries more personal data than a week of typed prompts combined: full names, dates of birth, national ID numbers, medical codes, account numbers, addresses, the names of people who are not even the subject of the file but show up in a paragraph about them.
None of that is unusual. It is what a normal working document looks like. The only problem is that none of it gets a second look before it goes into a tool that was never asked to keep it confidential.
#Redacting the document, not the prompt
BeeSensible is built for exactly this moment. Open a document in the extension's side panel, in the dashboard under Documents or in the desktop app (macOS and Windows), and the sensitive details on the page are highlighted, with the most sensitive ones (national ID numbers, financial details) picked out more strongly than the rest.
You untick anything that should stay, draw a box by hand over anything the detection missed and export an anonymised copy. Hidden document metadata is stripped in the same step, so nothing leaks through the file's properties either.
This works for PDFs (scanned ones included), scans and images, and Word, Excel, PowerPoint and text files; a scan is read with OCR, and the hidden parts of an Office file (comments, tracked changes, metadata) are checked and cleaned too. The desktop app handles PDF only. After redacting, BeeSensible reads the result back and warns you if anything is still there, but a photographed page still deserves a human eye. The document is processed on the same EU infrastructure as the rest of BeeSensible, in memory, and is not kept.
A redacted PDF is safer to share regardless of where it ends up. But which AI tool it ends up in is a separate question, and organisations increasingly cannot answer it for every tool their staff have opened.
A redacted document pasted into a tool your organisation has never assessed is a smaller problem than an unredacted one, but it is still a decision worth making on purpose rather than by default.
#Prompt hygiene and document hygiene are the same discipline
Typed text and uploaded files are two different moments with the same underlying risk: personal data reaching a tool before anyone looked at what was actually in it. Treating only one of those moments as the risk leaves the other wide open.
Redact the document. Watch the prompt. Neither one covers what the other does.
ChatGPT
