Design an AI-assisted service that uses the minimum necessary customer data and has an accountable lifecycle for every copy.
AI support can collect free-form messages, recordings, transcripts, account data, inferred intent, tool outputs, feedback, logs, and reviewer notes. Privacy begins by redesigning the task so sensitive data is not requested or copied unless the outcome truly requires it.
Map every data copy and purpose
Follow information from the user interface through routing, model, retrieval, integrations, analytics, quality review, support access, backups, exports, and deletion. Identify controller and provider responsibilities with qualified advice. A statement that the model does not train on data does not answer logging, retention, access, or subprocessor questions.
Define privacy controls in the workflow
- Notice appropriate to channel, purpose, and user expectation
- Minimum fields and redaction before model or log exposure
- Authentication separated from unnecessary conversation disclosure
- Purpose, access role, retention, deletion, and export for each data class
- Vendor and model settings aligned with contracts and actual configuration
- Safe fallback for users who cannot or do not want to use AI
Reduce exposure by task design
| Need | Lower-data approach | Escalate when |
|---|---|---|
| General answer | Use public approved knowledge | Account facts are required |
| Case routing | Collect category and contact path | Sensitive narrative appears |
| Account action | Authenticate in dedicated flow | Identity or authorization is uncertain |
Run a privacy design review
- Map normal and exception data flows.
- Challenge every collected field and retained copy.
- Configure redaction, access, and retention.
- Test prompts that solicit or reveal sensitive data.
- Verify vendor deletion and export behavior.
- Review sampled interactions for unexpected disclosure.
Watch free-form data expansion
- Users sharing credentials, health, financial, or identity information unprompted
- Staff copying full transcripts into unrelated systems
- Knowledge retrieval exposing internal or customer-specific documents
- Long retention justified only by possible future analytics
Maintain a data-control register
Record data class, source, purpose, legal or policy basis, system, provider, location, access role, retention, deletion method, user request path, owner, and last test. Update it when a model, prompt, channel, integration, review process, or vendor changes.
Continue with the next decision
Sources and further reading
Primary and contextual sources used to verify definitions or give readers a relevant next resource.
- NIST Privacy Framework Primary voluntary framework for identifying and managing privacy risk.