How Law Firms Pseudonymize Case Files: Consistent Placeholders for AI, Experts and Co-Counsel
To pseudonymize a case file, replace each party, witness, company and address with a numbered placeholder that stays the same across every document in the file, black out the identifiers nobody downstream needs (account numbers, ID numbers), and keep the correspondence table in the matter file, never with the documents you send. The recipient, whether an AI, an expert or a translator, reads a coherent file; nobody outside the firm learns who the client is.
When firms pseudonymize rather than redact
| Recipient | What they need | Technique |
|---|---|---|
| Generative AI (summary, chronology, first draft) | Who did what, in which order | Pseudonymize |
| Expert witness, accountant, translator | The full reasoning, not the identities | Pseudonymize |
| Co-counsel in another jurisdiction before engagement | Enough to assess the case | Pseudonymize |
| Training, knowledge base, case study | The legal point | Pseudonymize, then remove descriptive details |
| Court, opposing party, public | Only what they are entitled to | Redact |
Several supreme courts publish their decisions pseudonymized. The Swiss Federal Supreme Court replaces parties with "A.", "B."; the French Cour de cassation built its own AI engine to pseudonymize the decisions it releases as open data. The convention below follows the same idea with clearer placeholders.
A placeholder convention that works
Parties first, by role. Number them in order of first appearance in the main document, and keep the numbering for the whole file:
[Person 1]claimant,[Person 2]defendant,[Person 3]… witnesses and third parties[Org 1],[Org 2]companies and public bodies[Address 1]addresses,[Date 1]dates when the exact date identifies the event
Write the legend at the top of your internal note, not in the document you send: "[Person 1] = our client".
Or label by role. Many firms prefer [Claimant], [Defendant], [Witness 1], [Employer]: an AI drafting a reply understands roles better than numbers. Use the same label set across the whole file, and avoid real initials or first names, which give part of the identity away.
Variants map to the same placeholder. "Anna Keller", "Ms Keller", "Keller", "A.K." and the signature "A. Keller" are all [Person 1]. This is where manual find-and-replace fails most: abbreviations, possessives, names split across lines, and scanned pages that find-and-replace cannot read at all.
Lawyers and judges: decide once. Counsel and judges are not usually the persons whose data you protect, but their names can identify the case. Most firms replace them too when sending to an AI.
Black out, do not tag, what has no narrative role. Bank details, card numbers, social security and passport numbers, case numbers that can be looked up. A placeholder for an IBAN adds nothing for the reader.
Document by document
Pleadings and briefs. Party names, addresses for service, counsel, case number. Keep the court name unless the court and date together identify a well-known case.
Contracts. Parties in the preamble, signatories, addresses, bank details in the payment clause, annex references that name people.
Judgments. Parties, witnesses, experts, sometimes the place of the events. Published decisions are often already pseudonymized; check whether the version you have is the published one or the notified original.
Exhibits. Emails (sender, recipients, signatures, quoted earlier messages), bank statements (account holder, counterparties), medical reports (patient, doctors, clinic), photos of IDs.
Correspondence. Letterheads and footers carry addresses, phone numbers and registration numbers.
Keeping it consistent across many files
Consistency within one document is easy; across a 30-exhibit file it is the hard part. Two practical rules:
- Process the main document first, fix the numbering, then process the exhibits against that legend.
- Check the outliers. After processing, search the output for the surname, the first name and the initials. A flattened PDF has no text layer to search, so check the version before flattening or rely on the tool's review screen.
Redact PDF AI's pseudonymize mode groups the variants of each name, numbers entities in order of first appearance across the whole document, shows every placeholder in the review editor so you can rename it ([Person 1] → [Claimant]), merge two entities or switch one to a black box, and exports a flattened PDF plus the correspondence table.
Where the key lives
The correspondence table is client-confidential information. Keep it in the matter file under the same access rules as the original documents, never attached to the pseudonymized version, never pasted into the AI chat. When the AI's answer comes back in placeholders, re-insert the names yourself.
Frequently asked questions
Does pseudonymizing satisfy professional secrecy when using AI? It removes the direct identifiers, which is the core of the risk. Check your bar's guidance and your engagement terms, and remove descriptive details when the facts alone would identify the client.
Should I pseudonymize the court and the judge? Only if, combined with the date, they make the case findable. For most AI tasks the court's name is useful context.
What if the AI invents a name in its answer? It sometimes does when placeholders look like gaps. Typed placeholders such as [Person 1] reduce this; say in your prompt that the placeholders stand for real people and must be kept as is.
Can the opposing party's names stay? Their data is protected too. In a document sent to a third party, pseudonymize all natural persons, not only your client.
How long does it take? Minutes per document with automatic detection and a review pass; hours by hand on a long scanned file.
Next step
Pseudonymize a case file with consistent placeholders, or read how to anonymize a document before ChatGPT for the AI-specific checklist.