Redact your Excel file: remove the data, don't just cover it

A workbook is the hardest file to redact by eye. Excel decides what to draw, while the .xlsx underneath keeps everything it was ever given: hidden sheets, the list behind a drop-down, the author of each note. Before it reaches ChatGPT or Copilot, which reads Microsoft 365 files on its own terms, that gap matters.

Everything below happens on a copy: first finding what the file hides, then coding the names, then cutting the sheet down to what the AI needs. That last step is the one I value most. If the routine should become something staff sign, a ready-to-edit AI policy has a place for it.

Does hiding a column or a sheet remove the data?

No. Hidden is a display setting saved beside the values. Excel undoes it in two clicks, and a program reading the file can ignore it altogether. The Library of Congress describes an .xlsx as a ZIP package of XML parts, so any ZIP tool opens one.

We ran the test on 28 September 2026. We built an invented workbook with openpyxl, an open-source Python library: a hidden row, a hidden column, a hidden sheet and a very hidden sheet, a state Excel’s menus can’t set. Read back with openpyxl 3.1.5 and no special options, all four returned, names included.

Formatting tricks fail the same way. Microsoft’s help on the ;;; number format says the values stay in the formula bar, and white font or black fill do even less. Nor would I rely on a filter: it hides rows too. Boxes drawn over a PDF are the same mistake in another format.

Check an .xlsx for hidden sheets and comments before it goes

Look before you change anything. Microsoft’s help warns that the Unhide command won’t list a sheet set to very hidden, so an empty Unhide list proves nothing. The box below opens a workbook on your own computer and names every hidden sheet, very hidden ones included.

It also counts hidden rows and columns, notes and comments, and lists the people the file records as author or commenter. It reads structure, not cells, so a visible column full of names won’t register. It takes a .pptx too, for when a slide deck goes out with the workbook.

If it lists a sheet nobody in the office recognises, my move is to stop there and ask who built the file and what that sheet feeds. Deleting it blind can break the totals you were about to send.

Replace names with consistent IDs and keep the lookup table apart

Deleting the name column is cleanest when the task never needs people. Often it does: you want the AI to find who logged the most overtime, then know who that is. A code gives the model something to point at while the names stay with you.

Copy the names to a new sheet and choose Remove Duplicates on the Data tab; Microsoft warns it deletes permanently, so work on the copy. Give each person a code, fill the main sheet with XLOOKUP (Microsoft 365, Excel 2021 or later), paste as values and delete the names.

Pick codes nobody can guess. The EDPB’s draft guidelines on pseudonymisation warn against pseudonyms predictable from a few observed values, and EMP-001 for the first surname alphabetically is exactly that. Sequential codes are the shortcut I’m most wary of. A code you can reverse is masked, not anonymous.

Is the coded sheet still personal data? For you, yes. The GDPR (EU and UK law) treats data as pseudonymised when it can’t be tied to anyone without extra information kept separately and protected, and your key file is that information. A key left as a hidden sheet in the upload undoes everything.

In EDPS v SRB (4 September 2025), a case under the rules for EU institutions, the EU Court of Justice said pseudonymisation may, depending on the circumstances, stop people other than the controller from identifying anyone. I doubt that ruling helps a sheet whose job title and start date still point to one person.

Whatever the recipient can work out, the file still sits with the AI provider afterwards, under the terms OpenAI applies to what you upload if it goes to ChatGPT.

Find and Replace for names: what it misses in notes and comments

Codes handle the name column. Names also sit inside other cells: a remarks column, a label typed beside a total, an email address. Press Ctrl+H, set Within to Workbook, search each surname alone, and step through the hits with Find Next instead of trusting Replace All.

Microsoft’s help holds the catch: on the Replace tab, Look in offers only Formulas. A surname inside a note or a comment survives Replace All untouched. Switch to the Find tab, choose Notes, then Comments, and edit or delete each hit by hand.

Mac and web differ again, as the table shows. Where an office has one Windows machine, the last search belongs there, I think. A sheet that was exported to plain text needs a column-by-column pass of its own.

Formulas, named ranges and PivotTables that still hold the names

A visible sheet can keep pulling from one you meant to remove. A formula such as =Staff!B4 puts a name on screen and shows where the rest live. Paste values over anything that refers to people, then delete the source sheet and look for #REF! errors.

Named ranges sit in Formulas, Name Manager, and a drop-down list built with Data Validation often draws on a range of names on a hidden sheet. Microsoft’s Inspector lists hidden names among the things it finds but does not remove, so delete them yourself.

A PivotTable can carry its own copy of the source rows; Microsoft’s file-size help points to Save source data with file, under PivotTable Analyze, Options, Data. On a copy for an AI, I’d paste the summary as values and delete the PivotTable. Caches are among the quieter ways data leaves an office.

Comments, notes, headers and footers: names outside the grid

Excel has two kinds of annotation. Notes are the older yellow boxes; threaded comments carry a Reply box. Microsoft’s comparison draws the line clearly: on a note you can remove or change the name, on a threaded comment you cannot.

So a thread keeps every participant’s name, whatever the cell says. Right-click and choose Delete Comment or Delete Note on the copy. A Word report travelling with the workbook has its own edits to accept or reject.

Headers and footers show only in Page Layout view, Print Preview and on paper, so a name typed there years ago goes unseen. Tab names count too. Tabs are the piece I’d expect people to forget, because they look like structure. Author and company fields belong to the file’s metadata.

The Document Inspector in Excel, and the checks on a Mac or the web

On Windows, open File, Info, Check for Issues, Inspect Document, on the copy. Microsoft splits the Excel results in two. Comments, headers and footers, hidden rows, columns and sheets, custom XML and invisible content can be removed there and then.

The second group is found but left alone: external links, embedded files, macros, PivotTables and other features with cached data, scenarios, filters and hidden names. Microsoft adds two warnings. Undo may not restore what was removed, and deleting hidden rows or sheets that hold data can change formula results.

Microsoft’s Inspector help gives only the File, Info route and says nothing of Excel for Mac or the web. To me it is a second opinion, never the method. Word’s Inspector has blind spots of its own, white text among them.

Dates of birth, postcodes and job titles that point to one person

Removing names still leaves rows that describe someone: a birth date, a postcode or ZIP code and a job title can pick out one employee in a small office. NIST’s de-identification guide treats these quasi-identifiers as a problem of their own, and details that identify someone beyond a name reach well past spreadsheets.

Coarsen rather than delete: a birth year, a region, bands instead of job titles. Build the new column with a formula, paste it as values, then remove the original. The US HIPAA safe harbor, 45 CFR 164.514(b)(2), works the same way: dates to the year, ZIPs to three digits where that area tops 20,000 people.

Small groups give people away: if one team has a single member, its row is that person. NIST calls a table k-anonymous when every combination of these columns appears at least k times, and sets no k. Three is our own floor, not a NIST figure; below it, I’d merge categories.

Send only the columns the question needs

The quickest redaction is a column never copied. Write the question down, list the columns that answer it, and paste just those into a new workbook as values. Hidden sheets, PivotTables and links stay behind. The GDPR’s data minimisation principle, Article 5(1)(c), asks for the same.

An upload can outlive its chat. OpenAI’s help says ChatGPT saves uploaded spreadsheets to Library automatically, and deleting the chat doesn’t delete the file. Clearing a ChatGPT account properly covers both places.

I’d limit the columns more than feels comfortable: the AI can ask for one it lacks, and a column it never saw can’t leak. Running a model on your own hardware changes who receives the file, not how many columns the question needs.

Letting Nonimo find the names in a workbook

The tool at the top of this page takes an .xlsx, processes it on our server and returns the sheet with names, emails and numbers swapped for labels such as [PERSON_1]. In the preview you can redact a whole column by hand, or leave an item visible. What each plan allows covers heavier use.

The download is an Excel file again, without some parts, such as comments or PivotTables. Nonimo’s desktop app, on Mac or Windows, does the same swap on your own computer at a keypress. How the app handles your files is set out separately.

The parts missing from the download are mostly what the hand steps above delete anyway. What no detector judges well is a nickname typed into a remarks column, so that column still gets my own read. When the same workbook comes back every month, Nonimo’s app is on the home page.

Sources

• Microsoft Support, the Document Inspector article for Word, Excel and PowerPoint: the Excel inspectors, what is found but not removed, the Undo and formula warnings.
• Microsoft Support, Find or replace text and numbers on a worksheet: Look in on each tab, and the Mac and web differences.
• Microsoft Support, also used: Hide or unhide worksheets (the Unhide path; very hidden sheets missing from Unhide); Find and remove duplicates (permanent deletion); XLOOKUP function (versions); Hide or display cell values (;;;); The difference between threaded comments and notes; Insert comments and notes in Excel; Headers and footers in a worksheet; Reduce the file size of your Excel spreadsheets (PivotTable source data).
• EUR-Lex, Regulation (EU) 2016/679 (GDPR): Article 4(5) on pseudonymisation; Article 5(1)(c) on data minimisation.
• EDPB, Guidelines 01/2025 on Pseudonymisation: version for public consultation, adopted 16 January 2025; lookup tables and unpredictable pseudonyms, paragraphs 92 and 93.
• Court of Justice of the EU, Press release No 107/25, EDPS v SRB (C-413/23 P): pseudonymisation may, depending on the circumstances, prevent people other than the controller from identifying anyone.
• 45 CFR 164.514, via the Legal Information Institute: the safe harbor method: 18 identifier types removed, and no actual knowledge that what remains identifies anyone.
• NIST SP 800-188, De-Identifying Government Datasets (September 2023): transforming quasi-identifiers; k-anonymity defined, with no fixed k.
• Library of Congress, XLSX Transitional (Office Open XML): an .xlsx is a ZIP package of XML parts.
• OpenAI Help Center, Using Library to manage files in ChatGPT: uploaded spreadsheets saved to Library; deleting a chat leaves the file.
• openpyxl documentation, worksheet module: the visible, hidden and veryHidden sheet states.
• Nonimo test workbook, 28 September 2026: an invented .xlsx built with openpyxl and read with openpyxl 3.1.5.

Accessed 10 November 2026.