To use ChatGPT or Claude on client interviews and meeting transcripts without sending the client's details, put every file of the engagement into one case, replace the people, companies, emails and phone numbers with placeholders, review the matches, and send the safe copy. Person_1 stays Person_1 in every file, so the AI can follow who said what, and the real names go back into its answer on your PC.
Why a transcript leaks more than a document
A contract has its names in a few predictable places. A conversation scatters them. People say each other's names aloud, mention the client's company, a subsidiary and a rival, read out an email address or a phone number, and refer to a colleague by a nickname. Each speaker label at the start of a line repeats a name too.
Where the names hide in a typical engagement:
| Where | Example |
|---|---|
| Speaker labels | "Maria Lopez:" at the start of every turn |
| Names said in passing | "Ask Tom in finance, he signed that off" |
| Companies and projects | The client, its parent company, a competitor, an internal project name |
| Contact details read aloud | "Send it to m.lopez at the usual address" |
| The file itself | The file name, the title page, a header with the client's name |
| Companion files | The statement of work, the follow-up email, the notes you typed afterwards |
The list of what to remove before a document goes to AI covers the general case. This article is about the engagement: several files, one client, one set of people.
Decide what you want from the AI first
Write down the output before you touch the files: themes across ten interviews, a summary of one meeting, a list of decisions and owners, a slide outline. The output decides what the AI needs to see.
A theme analysis needs what each person said and their role, not the name of the client's biggest customer. If a detail is not needed for the output, delete it from the safe copy instead of only masking it. A passage that is not sent cannot leak.
Work from a transcript file. If your meeting tool or transcription service lets you export the transcript as a TXT, Word or PDF file, use that. Notes you took by hand and typed up work the same way, and so do scanned pages: text in pictures and scans is read too.
Put the whole engagement in one case
Create one case in ClientSafe for the engagement and add everything that belongs to it: the interview transcripts, the statement of work, the kick-off notes, the follow-up emails saved as files. ClientSafe reads Word, PDF, plain text and Outlook .msg files, among others.
Within a case, the same name gets the same placeholder in every file. A made-up example of what the three files look like after the safe copy:
| File | Original | Safe copy |
|---|---|---|
| Interview 3 (TXT) | "Maria Lopez: Our Hartwell team never saw the plan." | "Person_1: Our Company_1 team never saw the plan." |
| Statement of work (DOCX) | "Client contact: Maria Lopez, Hartwell Logistics" | "Client contact: Person_1, Company_1" |
| Follow-up email (MSG) | "Thanks, Maria. Tom Reyes will join the next call." | "Thanks, Person_1. Person_2 will join the next call." |
If each file were processed on its own, Person_1 in one transcript could be someone else in the next, and a question like "what did the client's sponsor say across all interviews" would get a wrong or invented answer. Working with a whole case explains the rule and its pitfalls.
An email thread is one more file in the same case. The saving steps are in summarizing a client email thread with AI.
Review the matches
Detection is a first pass, not a verdict. Open the list of matches and go through it, because a conversation produces the cases a contract does not:
- Speaker names. Check that every speaker label was found, including a first name used on its own.
- Names said in passing. Confirm the colleagues mentioned in the middle of a sentence are on the list, and that their placeholder matches the one used elsewhere in the case.
- Companies. The client, its parent, its competitors and its customers each need a placeholder.
- Emails and phone numbers spoken or typed in the text.
- False hits. Untick what is not personal data: a product called "Jordan", a street used as a metaphor, a team name that looks like a surname.
- Missed text. Mark what was not found, such as a nickname, a misspelled name that the transcription produced, or an internal code name for the client.
Transcription errors are the usual reason a name is missed. If the tool wrote "Lopes" in one place and "Lopez" in another, search the safe copy for the variant and mark it if it was missed.
What replacing names does not remove
A safe copy hides who is named. It does not hide who can be worked out. A job title that only one person holds, a town, a project everyone in the industry knows, "the year we closed the Berlin office": taken together, such details can point to one person or one company even with every name gone.
This is documented, not hypothetical. Latanya Sweeney's study of the 1990 US census found that 87% of the US population had a combination of 5-digit ZIP code, gender and date of birth that was likely unique. More recently, a 2026 preprint by Tianshi Li took 125 public interview transcripts with scientists, found 24 that mentioned a publication, and showed that off-the-shelf LLMs with web search could link six of them to specific scientific works, in some cases uniquely identifying the interviewee. The author reports under $0.5 and about four minutes of run time per transcript.
So after the names are replaced, read the safe copy once as a stranger would, looking for details that narrow things down:
- Rare job titles, unusual qualifications, a named paper, patent or product.
- Small places, specific dates of events, team sizes.
- Unique incidents: a layoff, a failed launch, a dispute with a named kind of supplier.
Delete or generalize what the output does not need ("a senior engineer at a regional logistics company" instead of the exact title and city). In HR matters, where one person's story is the whole document, this step matters more than in a market-research interview; see AI in HR: what to strip first.
Send the safe copy and restore the answer
Copy the safe copy, or save it as a Markdown .md file, and paste or attach it in ChatGPT, Claude, Gemini or Copilot. Tell the AI that the placeholders are labels to keep as written:
These are transcripts of interviews for one client engagement.
Names like Person_1 and Company_1 are placeholders: keep them as written.
Group what the interviewees said into themes. For each theme, list
which Person_N said it, in which interview, with a short quote.
Mark anything that only one person said.Asking for the speaker and the interview next to each point lets you check the answer against the transcripts. Models blend interviews together when they summarize, and a quote is easy to verify.
Paste the AI's answer into the same case in ClientSafe, and the real names come back. The restored text is a client document: keep it with the client's files, not in the chat history. Restoring real names covers what to check in the restored answer.
What to check outside the tool
A safe copy does not answer three questions that belong to you and your client.
- The NDA and the contract. Some agreements restrict sharing information with third parties or require approval for AI tools, whether or not a name is attached. Read the wording; does pasting a document into ChatGPT breach an NDA goes through it.
- Whether you were allowed to record. In the United States, federal law requires the consent of at least one party to record a conversation, and about 11 states require all parties, according to the Reporters Committee for Freedom of the Press. Other countries have their own rules. If you recorded without the consent the law requires, anonymizing the transcript does not fix it.
- What the interviewees were told. If you promised research participants or employees that their answers stay within your team, or named the tools you would use, an AI service may not fit that promise. Ethics boards and clients can also set their own conditions.
If your account type affects what the AI provider keeps, check that as well: data settings in ChatGPT, Claude and Gemini. None of this is legal advice; for a specific engagement, ask your lawyer or compliance contact.
Frequently asked questions
Do I need to anonymize if the AI account is a paid one?
A paid plan is about the provider's terms for your data. Whether you may send the content at all is set by your NDA and by what the interviewees agreed to, and the plan does not change that. The safe copy keeps the identities out of the chat in the first place.
Can the AI still analyze interviews if everyone is Person_N?
Yes. For themes, disagreements and who said what, the placeholders keep the speakers apart. If the meaning depends on who someone is, such as a role in the client's hierarchy, add that to your prompt without the name: "Person_2 is the CFO".
What about audio and video recordings?
ClientSafe works on documents, not recordings. Use the transcript as a TXT, DOCX or PDF file, and keep the recording itself out of the AI service.
Is the anonymized transcript still personal data?
It can be. If someone can be identified from the remaining details, or if you keep the key that maps placeholders back to names, data protection rules may still apply to the safe copy. Ask your data protection contact before you treat it as anonymous.
Should I keep the originals?
Yes. You need them for the deliverable and for your own records, subject to the retention terms of your engagement.
ClientSafe replaces names, companies, emails and phone numbers in the files of one engagement with the same placeholders on your PC, and puts the real names back into the AI's answer. Detection is not perfect, so review the list before you copy.