Removing the client's name is the obvious step. It is rarely enough. Documents identify people through account numbers, signatures, file names, comments and details that only work in combination. Use this list before you paste or upload a client document into an AI tool.

Direct identifiers

These point to one person or organization on their own. Remove or replace each of them, in every form it appears.

  • Names of people. Include the client, the other side, witnesses, family members, colleagues and anyone named in a signature block or a "cc" line. Check short forms, initials and misspellings.
  • Names of companies and institutions. An employer, a landlord, a school or a clinic can identify a person as well as a name can.
  • Addresses. Street addresses, postcodes and property descriptions.
  • Phone numbers and email addresses. Check letterheads, footers and signatures.
  • Account and card numbers. Bank accounts, IBANs, card numbers and policy numbers.
  • ID numbers. Tax numbers, national ID numbers, passport and driver's licence numbers, social security numbers.
  • Dates of birth. Often a pair of fields in a form, such as "Born" and "Age".
  • Reference numbers. Case, claim, invoice and contract numbers that someone could look up.

Details that identify when combined

Direct identifiers are the easy part, because you can search for them. The harder part is judging what the task needs. A request to fix the grammar of a letter needs no facts about the client at all. A request to check a settlement calculation needs the amounts and the dates, but not the names. Decide what the AI needs before you decide what to remove, and the list gets shorter and safer.

Some details are harmless alone and identifying together. A "plastic surgeon in a town of 4,000 who was sued in March" may be one person. When you read the document, ask whether a colleague or neighbor could work out who it is.

  • Rare job titles, unusual employers and small places.
  • Exact dates of events, admissions, hearings or transactions.
  • Unusual amounts that appear in public records.
  • Descriptions of a health condition, a family situation or a dispute that were reported in the news.

You do not always have to remove these. For many tasks, generalizing is enough: "a dispute in spring" instead of "a dispute on 14 March". Keep what the task needs and drop the rest.

What hides in the file

The visible page is only part of a document. Look in the places that are easy to forget.

Word documents

  • Comments and tracked changes, which carry reviewers' names.
  • Headers, footers and text boxes.
  • Document properties such as author and company.

Spreadsheets

  • Hidden sheets, hidden rows and columns.
  • Cell comments and named ranges.
  • Formulas that mention other file names.

Presentations

  • Speaker notes.
  • Text in images and charts.

Emails

  • The From, To and Cc lines, and the sender's signature.
  • Quoted earlier messages lower in the thread.
  • Names of attachments.

PDFs and scans

  • Scanned pages with names in a stamp, a handwritten note or a signature.
  • Bookmarks, form fields and document properties.

Every file

  • The file name. "Lopez_lease_final.docx" gives the name away before anyone opens the file.
  • Images and photographs. A photo of an ID card or a signature contains personal data even when the text around it has been cleaned.

A safe copy made from the document's text, not by editing the original, avoids many of these hiding places, because comments, metadata and embedded objects are not part of the text you paste. Images are the exception: leave them out unless you are sure they contain nothing identifying.

An example

Take a short email from a client to her accountant:

"Hi Tom, attached is my statement for account 4471-2290 at Harbor Credit Union. I moved to 12 Elm Street on 3 March, so please update my address. My husband Daniel is on the loan too. Thanks, Maria Lopez, 555-0142."

A thorough pass replaces the accountant's first name, the account number, the bank name, the street address, the husband's name, the client's name and the phone number. The date of the move is a judgment call: if the task is about the loan, you can generalize it to "in March" or drop it. The attachment's file name, the email's sender line and the signature need the same treatment.

After the pass, the text might read something like: "Hi Person_1, attached is my statement for account Account_1 at Company_1. I moved to Address_1 in March, so please update my address. My husband Person_3 is on the loan too. Thanks, Person_2, Phone_1." It is shorter on detail and just as useful for a question about the loan.

When in doubt, leave it out. You can always add a detail back to the next message if the AI needs it. You cannot take back a message that has already been sent.

A final check

  1. Search for the client's surname and every variant of it. Also search for the first name alone.
  2. Search for digits. Scan for long numbers and anything that looks like a date.
  3. Read the first and last page. Letterheads, signature blocks and footers are where identifiers cluster.
  4. Check the file name and any title or heading you typed into the chat.
  5. Ask whether a stranger could identify the client from what remains.

Automatic tools help with volume, and they miss things. Whatever you use, review the result yourself before anything is sent.

ClientSafe, a Windows app from GagarinSoft, finds names, companies, addresses, phone numbers, emails, account and ID numbers and dates of birth in a case's documents and shows you the matches to review. It leaves images out of the safe copy. Read more about ClientSafe.

General information, not legal advice. Check the rules that apply to you.

Sources