Redact in Word before it goes to AI

How to redact in Word comes down to one rule: change the words, not how they look. A client’s name coloured black, highlighted or hidden behind a shape is still in the file, and anything that reads the file rather than the screen can get it back. The safe method is to replace each detail with a marker, accept or reject every tracked change, delete the comments, and then check the parts of the document you do not normally see.

This guide takes that method step by step, with the menus as Microsoft names them in English. It is for the solicitor, accountant, HR adviser or office manager about to hand a Word file, or its text, to Claude, ChatGPT, Gemini or Copilot. The reason to do it first is the same as for any other outside provider: what goes in is personal data, as our guide on putting client data into ChatGPT sets out.

Three neighbouring questions are handled elsewhere. What the file says about its author, company and dates is a matter of metadata, which has a guide of its own. Redacting a PDF is a different job with its own traps. And the full list of identifiers to remove from an Irish document starts with the PPS number, covered in a guide of its own.

How to redact in Word, in the order that works

The order matters because some steps undo others. If you run Replace before accepting tracked changes, the old name can survive inside a deletion Word is still keeping for review. If you delete comments before reading them, you lose the one place a colleague wrote “check with Ms Ní Chíosóig”. The sequence below avoids both.

1. Save a copy. Work on the copy only, and give it a name that says so. The DPC’s guidance on redaction makes this its first principle, and Microsoft gives the same advice before Document Inspector, because some of what it removes cannot be restored.
2. Accept or reject every tracked change. On the Review tab, in the Accept list, choose Accept All Changes, or go through them one by one. This turns the history into ordinary text you can then replace.
3. Read the comments, then delete them all. Note anything that names a person, then use Delete All Comments in Document.
4. Replace the visible details. Home, Editing, Replace: each name, PPS number, Eircode, phone number, email and IBAN becomes a marker such as [CLIENT] or [PPSN].
5. Check the parts outside the body. Headers, footers, footnotes, text boxes, tables and anything formatted as hidden.
6. Run Document Inspector on Windows as a last check, knowing what it does not look at.
7. Send text, not the file, where the task allows it.

Each step has its own section below, starting with a test that shows how much a normal looking letter can carry. Step 7 also depends on the tool: the rules differ a good deal from one provider to the next, as our guide to Claude and your data shows for one of them.

One test letter, nine places the client’s name was hiding

To see what survives an ordinary redaction, we built a short solicitor’s letter as a .docx, following the Open XML format that Word files use. The client is invented: Siobhán Ní Chíosóig of Castlebar, with a PPS number whose check letter is impossible and an Eircode that cannot exist. In the body, her name had already been replaced with [CLIENT], so on screen the letter looked finished.

A real PPS number carries rules of its own, set out in our PPSN guide.

Then we opened the file as what it is. Microsoft’s own description of the format is that a .docx is a zip package, and that comments, headers, footers, footnotes and endnotes are separate parts inside it. Listing where the invented details sat, we found them in nine places after the visible redaction was done.

Two of the nine were in plain sight and were simply missed: the salutation and the IBAN in a table. Two were in text the page does not show by default: the hidden sentence and the tracked deletion. The other five sat in comments, header, footer, footnote and text box, which a reader scrolling through the body never meets.

What a text reader got out of it

We also ran the file through python-docx 1.2, a widely used open source library for reading Word files, and asked it for the text of each paragraph. It returned the hidden sentence, with the full name and the Eircode, as ordinary text. It left out the tracked insertion [PPSN] altogether, so the marker someone had typed in did not even appear.

That is one tool, and other tools behave differently. The lesson is not about python-docx. It is that what a program extracts from a .docx depends on the program, and you rarely know which one sits behind an upload button. The safe position is a file in which there is nothing left to extract.

Find and Replace, and the Irish names it walks past

Replace is the right tool for the body, and Microsoft’s page on it shows the route: Home, Editing, Replace, then Find what and Replace with. Microsoft also suggests choosing Replace rather than Replace All, so that you review each item before it changes. For redaction that advice is worth taking, because every stop is a chance to notice the forms the search did not match.

The options under More matter too. Match case finds only the capitalisation you type, and Find whole words only avoids partial matches inside longer words. The Find all word forms option works on English verbs, which is no help with names. What the search finds is exactly what you type, and Irish names are typed many ways.

Fadas, the vocative and the English form

A search for Sean does not find Seán, because the letters are different. The same goes for Siobhan and Siobhán, Padraig and Pádraig, Maire and Máire. Irish adds a second layer. In a salutation or when addressing someone directly, the vocative lenites the name, so Siobhán becomes a Shiobhán and Pádraig becomes a Phádraig. A search for Siobhán will not stop at a Shiobhán, a chara.

Surnames change with the person as well. A woman whose father is Ó Cíosóig is Ní Chíosóig, a married woman may be Uí Chíosóig, and in English the family may write Kissane. In our test letter the body, the comment and the text box used three different forms of the same client.

The references no search will find

The DPC’s guidance on redacting documents makes the same point from the other side: a search finds only what it is told to look for, names get misspelled, and one person can be referred to by first name, by title, by relationship or by initials. Its own example list includes a neighbour and a grandfather. None of those contain the name at all.

The practical answer is to read the document once before you search it. Build a short list of every form the client takes, and every other person named, and search for each. Everyone else named in the file needs the same care. A client’s consent does not stretch to the other side, the witnesses or a spouse, a point our AI clause for engagement letters takes up.

Numbers have the same weakness. A PPS number typed with a space, a phone number split 087 or +353 87, an Eircode with or without its space: each is a different string. Search for the distinctive part, such as the last four digits, and read what comes up.

Changes and comments, before you cover anything

A name deleted with Track Changes on is not gone: Word keeps it as a deletion until someone accepts or rejects it, and in our test letter the PPS number survived that way. So in the copy, accept or reject every change and delete all the comments before you replace anything; No Markup only hides the marks. Our guide on how to remove track changes from a Word document takes each step on Windows, Mac and the web.

Saving the copy as Rich Text is no shortcut, since an RTF copy of a Word document has its own place for text deleted under tracked changes.

Headers, footers, footnotes and text boxes

Most redaction mistakes in Word are in parts of the file that do not scroll past with the body. Microsoft’s description of the format lists comments, headers, footers, footnotes and endnotes as separate parts of the package. Letterheads, file references and page footers are exactly where Irish firms put the client’s surname and the matter number.

Microsoft’s Replace page does not say which of these parts Replace All reaches. Rather than rely on it, open each part and read it, including the first page and any section with a header of its own. Firms whose letterhead carries the matter reference, as many Irish solicitors do, should look there first.

Footnotes, endnotes and text boxes

Footnotes and endnotes sit at the foot of the page or the end of the document, and they are easy to skip because they look like apparatus. In our test letter the footnote to the fee estimate carried the client’s full name and a date. Text boxes are worse, because they can float anywhere, sit behind other objects or be shrunk to almost nothing. Slides share the trap, so clearing a presentation looks past the edge of every slide.

Tables need a slower read than the body. They are where refund details, dates of birth and account numbers end up, and the eye tends to skim a grid. In the test letter, the IBAN for a refund sat in a two column table and survived an otherwise careful pass. A spreadsheet is all grid and no body, which is why the Excel method starts by making every hidden row and tab visible again.

Hidden text and white text

Word can format text as hidden, a font effect in the Font dialog box. Hidden text may not show on the page, but it is still in the file, and in our test it came straight out through a text reader. Microsoft’s Document Inspector looks for it. Its own page adds that it cannot detect text hidden by other methods, and gives white text on a white background as the example.

The same page says the Invisible Content check cannot find objects covered by other objects. So a shape drawn over a name, the Word equivalent of a black box on a PDF, is not something the Inspector promises to catch. The name has to go from the text itself.

Embedded objects and pictures

A Word file can carry an embedded spreadsheet, a chart built from a table of client data or a pasted image of a signed form. The DPC’s guidance points out that search functions cannot read text stored as an image, such as a photo of a certificate. Open each embedded object, or delete it from the redacted copy if the task does not need it. For Irish accountants, the embedded payroll sheet is the usual suspect. A photo of a driving licence is the same case: redact and watermark the licence copy before it goes into the file.

Document Inspector: what it checks and what it cannot see

Document Inspector is Microsoft’s own final check. On the File tab, select Info, then Check for Issues, then Inspect Document. Tick the checks you want, select Inspect, and use Remove All next to anything you want gone. Microsoft recommends running it on a copy, because what it removes cannot always be restored and Undo may not bring it back.

Its page lists what it looks for in a Word document, and just as usefully, what it says it cannot find.

Inspector check	what it covers	what the page says it misses
Comments, Revisions, Versions and Annotations	comments, revision marks from tracked changes, versions, ink
Document Properties and Personal Information	properties, username, template name and more
Headers, Footers and Watermarks	the content of headers, footers and watermarks	removes them whole
Hidden Text	text formatted as hidden	white text on white
Invisible Content	objects formatted as invisible	objects covered by others
Custom XML Data	XML stored in the file

Source: Microsoft Support, Document Inspector page, Word section. The page applies to Word for Microsoft 365 and Word 2016 to 2024; it does not list Word for Mac.

Three things follow. First, the Inspector does not redact anything: removing the header removes all of it, including the firm’s name, and it does not change a single word in the body. Second, footnotes, text boxes, tables and embedded objects are not on its list, so they stay your job. Third, its property check belongs to the question of metadata, which has its own guide on clearing a document’s properties.

Use it as the check at the end, not the method. Microsoft says the same thing in a quieter way, placing the Inspector as the final step after accepting or rejecting changes. And a clean file does not make a tool suitable for client work, which is a separate question, taken up in our comparison of AI assistants under the GDPR.

What the DPC and the Tusla decision mean for a file sent to AI

Ireland already has a GDPR fine that turns on documents that were not redacted. In April 2020 the Data Protection Commission decided an inquiry into Tusla, the Child and Family Agency, covering three breaches. In each, Tusla gave documents to third parties without redacting personal data: a foster carer’s address went to a child’s father, and a complainant’s address and her mother’s phone number went to the person she had accused.

The DPC found that Tusla had infringed Article 32(1) of the GDPR by failing to have appropriate organisational measures in place for sharing documents with third parties. It also found a late notification under Article 33(1) for the third breach, and fined Tusla €75,000, the first GDPR fine in Ireland after a statutory inquiry. The Dublin Circuit Court confirmed it in November 2020.

€75,000

That case had nothing to do with AI. It is relevant because an AI provider is also a third party, and handing it a document is sharing that document. The standard the DPC applied is organisational: whether the firm had measures in place, not whether one person made one slip. A written routine for redacting Word files before they go to a chatbot is that kind of measure.

What the DPC’s guidance adds

The DPC’s guidance on redacting documents and records, updated in August 2021, was written for subject access requests, but its advice on electronic files is general. It warns against highlighting or recolouring text, because the words stay intact and come back when copied as plain text. It recommends replacing the text with a marker such as [REDACTED], and it names comments, revision histories and editing mark ups among the content a file can conceal.

Its separate note on anonymisation and pseudonymisation sets the limit. Masking alone, it says, often allows “a very high risk of identification”, and will not normally count as anonymisation in itself. A redacted letter whose original sits in your own files is pseudonymised at best, and so remains within the GDPR, as our guide on pseudonymised data in Ireland explains.

Text, not the file

Once a Word file is uploaded, the provider receives the whole package, including every part this guide has discussed. What it then does with a file is a question for its own terms, which our guides to what ChatGPT keeps and Microsoft’s Copilot terms take provider by provider.

For most tasks, a model needs the words, not the file. Pasting the redacted text leaves comments, tracked changes, headers and properties behind by construction. Where you must upload, upload the checked copy. Solicitors preparing affidavits have the court’s own AI rules to follow as well, and Practice Direction HC 142 is where they start.

A Word document run through Nonimo

The Nonimo app does the replacing step on your own computer, and the document never leaves it. Drop a Word file into the app and it hands back the text with the details covered, ready to paste into the AI, and a covered copy of the document: the same Word file, covered in its title and properties too. It pseudonymises: the details become markers, and once you paste the AI’s reply into the app, the real ones are put back. The page’s own tool, above this guide, gives back the same Word document, covered.

We ran the text of the invented letter through the Nonimo app.

in the letter	what the model would see
Client: Siobhán Ní Chíosóig	Client: [PERSON_1]
PPS number: 7654321X	PPS number: [REFERENCE_2]
Address: 0 Chapel Street, Castlebar, Co. Mayo F23 M0M0	Address: [RECORD_FIELD_1]
Email: siobhan.nichiosoig@example.com	Email: [EMAIL_1]
Ms Ní Chíosóig will proceed with her claim	Ms [PERSON_2] will proceed with her claim
IE29AIBK93115212345678	[IBAN_1]

Invented data; the IBAN is the public example from the IBAN registry. Output of the Nonimo app, 28 September 2026.

Accepting the changes and deleting the comments first is a good habit too, so no dates or editing history go with the file. As with any redaction, give the result a quick read before you paste.

Before you upload: checking the copy in ten minutes

Redacting and checking are different jobs, and checking is what people skip. For a letter or a short attendance note it takes about ten minutes, and done on the copy, with fresh eyes, it catches something more often than anyone expects. These checks are worth writing into the firm’s AI policy. A GP practice runs the same checks on a referral letter, under the extra rules in our guide to patient data and AI.

1. Reopen the redacted copy and use Next on the Review tab: there should be no tracked change left to accept.
2. Search the copy again for every form on your list, and for the last four digits of each number.
3. Copy the whole text into a plain text editor, the DPC’s own test for recoloured text, and read it there.
4. Look once more at the header and footer of each section, the footnotes and every text box.
5. On Windows, run Document Inspector and read what it reports before you remove anything.
6. Ask whether what is left, an employer, a town, a date, still points to one person.

The last check is yours. An employer in a small town, a rare job title and a date can identify someone as surely as a name. That judgement stays with whoever knows the file, and for health or union details the extra rules are in the Article 9 guide.

Firms that do this every day tend to build it into how the office works rather than leaving it to memory. Our pages for Irish solicitors and Irish accountants show how that looks in those two professions, where client files move between Word and AI tools most often.

Sources

• Microsoft Support, Find and replace text in Word. The route Home, Editing, Replace, the Find what and Replace with boxes, and the tip to choose Replace rather than Replace All. support.microsoft.com
• Microsoft Support, Replace Text. The options Match case, Find whole words only and Find all word forms (English). support.microsoft.com
• Microsoft Support, Accept or reject tracked changes in Word. No Markup only hides changes for now; Review, Tracking, Accept All Changes; Document Inspector as the final check. support.microsoft.com
• Microsoft Support, Insert, delete, or change a comment. Review tab, the arrow by Delete, Delete All Comments in Document. support.microsoft.com
• Microsoft Support, Remove hidden data and personal information by inspecting documents, presentations, or workbooks. The route File, Info, Check for Issues, Inspect Document; the Word checks; the limits on white text and covered objects; the versions it applies to. support.microsoft.com
• Microsoft Learn, Structure of a WordprocessingML document. A .docx is a zip package; comments, headers, footers, footnotes and endnotes are separate parts. learn.microsoft.com
• Data Protection Commission, Redacting Documents and Records (August 2021). Work on a copy; searches find only what they are told to; do not rely on highlighting or recolouring; replace text with a marker; concealed content in files. dataprotection.ie
• Data Protection Commission, Guidance on Anonymisation and Pseudonymisation (June 2019). Masking alone often leaves a very high risk of identification and will not normally be anonymisation. dataprotection.ie
• Data Protection Commission, Inquiry into Tusla Child and Family Agency, IN-19-10-1 (7 April 2020). Three breaches from documents shared without redaction; Articles 32(1) and 33(1); a fine of €75,000. dataprotection.ie
• Data Protection Commission, press release of 4 November 2020. The Tusla fine confirmed in the Dublin Circuit Court under section 143 of the Data Protection Act 2018. dataprotection.ie