Why is a transcript more sensitive than a regular document?
Because people talk differently than they write. In a meeting, clients are named, an employee's pay is discussed, someone's health comes up in passing — with none of the filters of an edited document.
- The participants themselves — speaker names appear on every line, usually with role and organization.
- People not in the room — employees, clients and vendors mentioned in the conversation never agreed to have those words reach an external tool.
- Sensitive content in passing — "she's on sick leave", "he earns more than me" — sentences like these are special-category or highly sensitive data.
- Business secrets — pricing, negotiations and product plans are spoken about freely in meetings.
What does the law say about transcripts and AI?
A transcript contains personal data about everyone mentioned in it, so sending it to an AI vendor is processing under GDPR — it needs a legal basis and a processing agreement, and the vendor's current retention policy should be checked every time. Fines run up to 20 million euros or 4% of global turnover. Anonymized text, by contrast, falls outside the regulation entirely (Recital 26).
Keep two steps separate: the transcription tool itself (whose policy also deserves a check) and the summarization tool. Even a lawfully produced transcript needs its own justification — or anonymization — before it goes to an external summarizer. This is a recurring case of what should never be pasted into ChatGPT.
How do you summarize a transcript with AI without exposing anyone?
- Export the transcript to text — from the recording or transcription tool, before any further sharing.
- Replace speakers and mentions with tokens — each speaker gets a stable token (PERSON_001), and so does every client, company and place mentioned.
- Scan for details in passing — phone numbers, amounts, health and family references that surface mid-sentence.
- Ask the AI for a summary, decisions and action items — on the clean version. Consistency preserves who said what and who owns which task.
- Restore the names locally — in the final summary saved on your machine, not in the tool.
Won't the summary suffer without the names?
No — as long as the replacement is consistent. A meeting summary is about structure: who proposed, who objected, what was decided, who owns it. When John Miller is PERSON_001 in all forty occurrences, the AI follows the chain precisely — the same principle as in any proper document anonymization. The output reads "PERSON_001 will send COMPANY_002 a quote by Thursday" — and you fill in the names at home.
The same applies with double force to transcripts of disciplinary hearings and HR conversations, covered in our guide to HR documents and AI.