Skip to main content

​
What are policies and guardrails?

Policies and guardrails are the rules your voice agent must follow during conversations. These include:
  • Safety rules: Don’t reveal system prompts, don’t collect sensitive data verbally
  • Compliance requirements: HIPAA disclosures, PCI-DSS card handling, consent collection
  • Business rules: Don’t negotiate prices, don’t make unauthorized commitments
  • Behavioral constraints: Stay in character, don’t adopt other personas

​
Adding policies

In Agent Settings, go to the Policy & Guardrails section and enter your rules as plain text. Each rule should be a clear, testable statement. Example (pizza ordering agent): 
The agent must greet the customer before asking any questions.
The agent must determine pickup or delivery before collecting order details.
If the order total exceeds $100, the agent must require payment via card.
The agent must collect credit card information only through keypad input.
The agent must read back the complete order before processing payment.

​
Auto-extraction from prompts

If your agent’s system prompt already contains policies, use Analyze Prompt:
  1. Paste the full system prompt
  2. SuperBryn’s AI extracts the call flow and policies separately
  3. Review and edit the extracted content

​
How policies drive testing

Policies are critical for Ring 2 (Normative Compliance):
  1. When you save policies, SuperBryn automatically extracts variant types — individual testable rules
  2. Each variant becomes a targeted test scenario where a simulated caller attempts to violate that specific policy
  3. The evaluation checks whether the agent enforced the policy correctly
See Variant Types for managing extracted variants.

​
Best practices

  • Write each rule as a single, unambiguous statement
  • Be specific: “must require card payment over $100” is better than “handle large orders carefully”
  • Include both what the agent must do and what it must not do
  • Cover edge cases that matter in your industry