Developers paste customer PII into a third-party LLM API to draft support replies. What's the concern and action?

Support is slow, so developers start pasting real customer records — names, emails, account numbers, sometimes worse — into a third-party LLM API to draft replies faster. It works beautifully and it's a privacy incident waiting to be discovered.

The concern: sensitive-data disclosure

The moment PII leaves your boundary for an external API, you have disclosed sensitive data to a third party. That data may be retained, logged, reviewed by humans, or used to train the provider's models, depending on their terms. You typically remain the data controller with legal obligations under regulations like GDPR or CCPA, while the provider becomes a processor — but only if a proper agreement says so. Pasting ad hoc means none of those guarantees exist, and you've likely breached your own privacy commitments to customers.

"It's just text" is the failure mindset: that text is regulated personal data.

The action

• Minimize and redact. Strip or tokenize PII before it ever leaves your systems; send the model only what it needs to draft a reply.
• Check the provider's terms. Confirm data-use, retention, human-review, and especially a no-training guarantee, backed by a signed data-processing agreement (DPA) and the right deployment region.
• Use a private deployment for sensitive data. A self-hosted or contractually isolated, no-retention enterprise endpoint keeps regulated data inside your trust boundary.
• Govern it. Give developers an approved, redacting path so they don't improvise with the public consumer endpoint.

Why the distractors are wrong

• "No concern, it's just text": that text is regulated PII; this is the exact rationalization that causes breaches.
• "Longer API key": key length affects nothing about where the data goes or how it's used. It's a non sequitur.
• "Send more data": more PII out the door is more exposure and more regulatory risk, not better service.

What interviewers want to hear

That you frame it as sensitive-information disclosure with real privacy/legal consequences, apply data minimization and redaction, demand a DPA and no-training/retention terms, and provide a sanctioned private path for sensitive data.