Safety and privacy¶
Safe operation
Agents can handle sensitive conversations. Keep prompts free of secrets, collect only necessary data, and define when the agent must refuse, escalate, or hand over.
The safest agent is useful without pretending to know more than it does.
Agents can handle sensitive conversations. Configure them so they are useful without exposing unnecessary data or making promises they cannot keep.
Do not put secrets in prompts¶
Never put API keys, passwords, private tokens, or internal credentials into prompts or knowledge-base content.
Use the platform's integration and configuration options for credentials.
Keep customer data minimal¶
Collect only the information needed for the workflow.
Examples:
- For callback: name, phone number, reason for callback.
- For appointment: name, contact details, preferred time.
- For support: customer identifier, issue description, urgency.
Avoid collecting sensitive information unless the workflow truly requires it.
Define boundaries¶
The prompt should say what the agent must not do.
Examples:
Do not provide legal, medical, or financial advice.
Do not promise discounts or delivery dates unless they are available in the knowledge base.
Do not ask for passwords or payment card details.
Human handover¶
Use human handover or forwarding when:
- the caller is upset,
- the request is urgent,
- the caller asks for a person,
- the agent is not confident,
- the topic is sensitive.
Review public-facing answers¶
Before going live, review answers for:
- accuracy,
- tone,
- privacy,
- legal or compliance risk,
- unnecessary internal details,
- hallucinated promises.
Safe fallback beats confident guessing
If the agent is unsure, it should say so and offer a useful next step.