Assistants inevitably meet sensitive topics: health symptoms, legal trouble, money worries, mental distress, personal data. Planning for them protects users and organisations.
Define the Assistant's Scope
State what it should and shouldn't help with. A banking assistant can explain fees but shouldn't give personalised investment advice; a general assistant can share health information but not diagnose.
Provide Safe Responses
For out-of-scope or risky requests, specify how to respond: acknowledge the need, give general information where appropriate, and point to qualified help or official resources.
Crisis Situations
For signs of self-harm or danger, responses should be supportive and direct people to emergency or crisis services appropriate to their location. Test these scenarios carefully.
Avoid Overconfidence
Encourage appropriate caveats without burying useful information in disclaimers.
Personal Data
Instruct the assistant not to request unnecessary personal information, and handle what users share according to your privacy obligations.
Human Escalation
Provide routes to a person for complex, high-stakes or distressing situations.
Test Thoroughly
Build test cases covering sensitive requests, including indirect phrasing, and review responses with relevant experts.
Rely on More Than Prompts
Combine prompt instructions with moderation, output checks and escalation paths; prompts alone aren't a sufficient safeguard.