Check PDF Redaction Before AI Upload

Before uploading a partially redacted document, check the saved file itself. Covering a name on the page may leave the original text available for selection or extraction. A document can also contain comments, attachments or metadata that the ordinary page view does not display.

Do the preparation in an approved local or organisational workflow. Uploading the unredacted original to an external assistant and asking it to remove the sensitive parts has already disclosed the original to that service.

Decide what the assistant needs

Start with the task. If the assistant only needs three paragraphs from a policy, a carefully prepared extract may be easier to control than a full document. Preserve enough context for the answer to remain meaningful, including relevant dates and exceptions.

Write a clear removal list. Names alone may be insufficient if job titles, case numbers or combinations of details identify the same person. This is a judgement about the material and its intended recipient, not a pattern-matching task that a redaction button can settle.

Keep the original under its existing access controls and create a labelled working copy. The ICO's redaction guidance recommends working from a copy, recording the decision and checking the result. Its example of text covered by a simple rectangle shows why visual concealment can fail.

Apply removal, then inspect hidden content

Use a tool designed to redact the relevant PDF content. Marking an area for redaction is a preparation step; complete the tool's application and saving process before treating the copy as ready.

Adobe's redaction and sanitisation overview distinguishes removal of visible text or graphics from sanitising hidden data. Applied and saved redactions are permanent, so carry them out on the working copy.

Then review hidden content. Adobe's sanitisation instructions provide options to remove all hidden information or selectively remove identified items. Check the choices against the document's purpose. Removing comments may be required for disclosure, while losing a needed attachment may make the permitted task impossible.

Record any functional loss caused by the process. A form, digital signature or navigation feature may need separate consideration in the resulting document. Do not assume the sanitised copy remains suitable for its original operational role; it is being prepared for a defined use.

Reopen the exact output file

Close the working document and open the saved output from its final location. This avoids checking an unsaved view while accidentally uploading an earlier copy.

Inspect every intended redaction and the material around it. Search for the removed names and identifiers, including common variants. Try selecting and copying text near the redacted area into an approved plain-text viewer. Review document properties, comments and attachments using the available inspection tools.

For a scanned or mixed document, include the image view and any searchable text layer in the check. A successful text search is useful evidence when it finds a problem; an empty search result does not prove that an image or differently encoded item contains no sensitive information.

If the risk warrants it, have a second person review the final copy against the removal list. Keep the review focused on both missed content and unnecessary removal. The assistant cannot answer accurately if the redaction also removes the condition that changes the policy's meaning.

Keep the upload path unambiguous

Place the approved copy in a clearly named handoff location. Separate it from originals and intermediate files so the uploader does not have to choose between several nearly identical filenames.

Record a file digest or another reliable identity in the review record where the workflow supports it. This connects the review to the exact bytes uploaded. A later edit should create a new review requirement, rather than inheriting approval because the filename stayed the same.

Check whether the AI tool will receive only this file or also draw on a connected folder containing the original. The project knowledge-boundary note covers that broader source scope. A correctly prepared attachment does not narrow an independently authorised connector.

Rehearse with fake data

Before relying on a new process, create a synthetic PDF with recognisable test markers in visible text, comments and metadata. Apply the proposed workflow and check where each marker survives. This gives staff a concrete way to learn the difference between covering, marking, applying and sanitising.

If a marker survives unexpectedly, stop the handoff and repair the process. Return to the retained original to create a fresh copy rather than trying to reverse a permanent redaction. Keep the review evidence with the document decision, without copying the sensitive content into a widely accessible ticket or chat.

References

• Information Commissioner's Office, Avoiding accidental breaches when redacting information. Accessed 17 October 2026.
• Adobe, About redacting and sanitising PDFs in Acrobat Pro. Accessed 17 October 2026.
• Adobe, Sanitise PDFs in Acrobat Pro. Accessed 17 October 2026.