Aller au contenu principal

Coaching — Personal Identifiable Information (PII)

Why it's risky​

PII includes any data that could identify a natural person on its own or combined with other data: full name, email, phone number, postal address, national ID number (NIR, SSN), passport, date of birth, IP address, biometric data.

Sending PII to a public LLM means:

  1. GDPR exposure — under Article 6, you need a lawful basis to process personal data, and most public LLMs do not provide one for your use case.
  2. Data residency loss — the data may leave the EU and be retained in the model provider's infrastructure (US, often), breaching EU adequacy obligations.
  3. Model training contamination — unless explicitly disabled, the prompt may be used to train future models, embedding personal data into a third-party asset you can never recall.
  4. Re-identification risk — even partial PII (initials + zip code + DOB) can be cross-referenced with public records.

Examples​

"Write a follow-up email to John Smith (john.smith@acme.com, +33 6 12 34 56 78)
 about his contract renewal scheduled for 2026-06-15."

Even seemingly innocuous prompts leak PII when concatenated with the model's response and stored in a chat history.

Best practice​

  • Mask before sending — replace the name with [CLIENT_A], the email with [EMAIL], the phone with [PHONE].
  • Aggregate, don't individuate — ask the model for a template, not a personalised draft.
  • Use enterprise endpoints — if your tenant has a contract with the LLM provider that excludes training and includes EU residency, use that endpoint, not the consumer one.

Alternative phrasing​

❌ Risky:

"Draft a contract amendment for Marie Dupont, NIR 2 85 04 75 116 003 42, born 1985-04-12 in Paris."

✅ Safe:

"Draft a contract amendment template for an individual contractor. Include placeholders for full name, national ID, date of birth, and place of birth."

References​