How to remove personal details from a spreadsheet before you use an AI assistant
A step-by-step method for finding and removing the values that identify people in a CSV or Excel file, including the ones that are easy to miss, and for checking what is left.
You have a spreadsheet of customers, staff, or orders, and you want an AI assistant to help you analyze it. The file contains real people's details. Once you paste or upload it, a copy sits with another company, under that company's terms, for as long as those terms say. Removing the personal details first keeps the analysis and leaves the people out of it.
Doing that well takes more than deleting the "Name" column. This guide gives a method you can follow by hand, and shows where each step goes wrong.
Key terms
- Direct identifier
- a value that points to one person on its own, such as a name, an email address, a phone number, or a customer number.
- Quasi-identifier
- a value that does not identify anyone alone but can do so in combination with others, such as postal code, date of birth, and job title.
- Free-text column
- a column holding sentences instead of a fixed kind of value, such as notes, comments, or descriptions. Names and phone numbers often appear inside them.
- Stand-in
- a replacement value used in place of a real one, such as
PERSON_001in place of a name. Also called a pseudonym. - Recipient
- whoever receives the file, here the AI service. What you remove depends on what the recipient needs.
What should you decide before touching the file?
Decide the question you want answered. The answer sets what you can remove. If you want the average order value per region, the assistant needs amounts and regions, and it does not need names, street addresses, or phone numbers.
This is the same idea data protection law calls data minimization: use only the data you need for the purpose. It also costs nothing, because a column you delete cannot leak.
Which columns hold personal details?
Go through the columns one by one. Do not rely on the header names, because "Ref" can be a customer number and "Description" can hold anything.
| Kind of column | Examples | What to do |
|---|---|---|
| Direct identifiers | Name, email, phone number, customer number, account number, tax number, employee number | Remove them, or replace each with a stand-in that is the same every time the same value appears (see replacing real names with consistent stand-ins) |
| Quasi-identifiers | Postal code, birth date, job title, department, rare product or event | Make them coarser: year instead of date, region instead of postal code, band instead of exact age |
| Free text | Notes, comments, ticket text, descriptions | Search them for names and numbers, or remove the column |
| Sensitive categories | Health, ethnic origin, religion, union membership, political opinion | Remove unless the question needs them, and then take advice first |
| Values the question needs | Amounts, quantities, dates, categories | Keep; round or shift only if the question allows |
Why is removing the names not enough?
Because a person can be recognized by a combination of ordinary-looking facts.
Why a combination of facts can identify a person
In 2000 the computer scientist Latanya Sweeney reported that, in 1990 US census data, 87 percent of the population could be picked out by three facts: a five-digit ZIP code, gender, and full date of birth. A later study by Philippe Golle found lower figures for the same three facts: 61 percent using the 1990 data and 63 percent using the 2000 data. But the point stands: three plain columns can identify most people in a population.
Another case shows the same with behavior instead of demographics. In 2006 Netflix released about half a million subscribers' film ratings with the names removed. Researchers Arvind Narayanan and Vitaly Shmatikov showed that a small amount of outside knowledge, such as a few ratings with approximate dates posted on the Internet Movie Database, was enough to find a person's record in the data.
What this means for your spreadsheet: a column does not need to be called "Name" to identify someone. Date of birth with postal code, or a rare job title in a small office, can point to one person. The more people share a combination of values, the harder it is to pick one out, and coarser values make larger groups.
What about names inside notes and comments?
Free-text columns are where most leaks happen. A "Notes" column that reads "Called Mrs Hale about the refund, her number is 555 0143" contains a name and a phone number inside a sentence. Deleting the "Customer" column does nothing for it.
Ways to deal with free text:
- Remove the column if the question does not need it.
- Search it for the names you know are in the file (customers, staff, suppliers) and for patterns such as email addresses and phone numbers, then replace each hit.
- Accept that you will miss some. Nicknames, misspellings, initials, and descriptions such as "the manager of the Brno office" do not match a search for a name.
What else is in an Excel file besides the cells?
An Excel workbook can hold comments, hidden sheets, hidden rows and columns, document properties such as author and company, links to other files, and the stored data behind pivot tables and charts. None of that is visible in the grid. The article on hidden data in Excel files lists where to look. If you start from a CSV file, these items are not present, but check that the CSV came from the sheet you meant.
How do you check what is left?
Check the file you are about to share, not the one you started from.
Should you keep a way back?
If you want to read the assistant's answer in terms of real people, keep a record of which stand-in replaced which real value, and keep that record separately from the shared file. Whoever holds it can undo the protection. See putting the original values back after an AI answers.
How does DataAttendant help with this?
DataAttendant is a browser tool that applies these steps on your own computer, to a CSV file (.csv, .tsv, or .txt) or an Excel workbook (.xlsx). It looks for names, email addresses, phone numbers, account numbers, and other values, proposes what to do with each kind of value, shows a preview, and lists what remains in the protected copy before you download it. You decide every treatment, and you can save your choices as a reusable plan for the next file. It does not decide for you what is safe to share, and it cannot find everything, so the checks above still apply. The file's contents are processed in your browser; they are not sent to DataAttendant. How DataAttendant works goes through the steps.
How do you stay in control?
By hand, the method costs your time, and each check above adds to it. In DataAttendant, you open the file, look at the preview and the check of what remains, and decide before anything is exported. The protected copy goes only where you send it yourself.
Sources
- Sweeney, L. (2000). Simple Demographics Often Identify People Uniquely. Carnegie Mellon University, Data Privacy Working Paper 3. The 87 percent figure is from her analysis of 1990 US census summary data.
- Golle, P. (2006). Revisiting the Uniqueness of Simple Demographics in the US Population. Workshop on Privacy in the Electronic Society. The 61 percent (1990 census) and 63 percent (2000 census) figures, checked against the paper on 20 October 2026. The paper does not explain the gap with Sweeney's 87 percent.
- Narayanan, A. and Shmatikov, V. (2008). Robust De-anonymization of Large Sparse Datasets. IEEE Symposium on Security and Privacy. Available at arxiv.org/abs/cs/0610105. Read 19 October 2026.
- General Data Protection Regulation, Recital 26 (the "means reasonably likely to be used" test) and Article 5(1)(c) (data minimization).