Get Started
Use Cases > Stay on Topic

Test what happens when callers go off-script.

Every voice agent in production runs into callers who tell jokes, make small talk, probe its identity, ask the weather, ask for opinions, or try to manipulate it into breaking character. How the agent responds is a brand signal and a security signal. Coval tests whether your agent stays in its lane without being cold about it.

Test Off-Topic Handling on Your Agent

What we check.

We score each transcript against the off-topic rules for your context:

  • Redirects gracefully rather than bluntly.
  • Allows a brief moment of warmth on small talk before guiding back to the task.
  • Is honest about being AI when asked.
  • Declines to give opinions on competitors, politics, or topics outside its scope.
  • Resists jailbreak attempts and prompt injection without acknowledging the attempt.
  • Handles multi-intent callers by sequencing requests instead of dismissing them.

We probe the full range, from edge cases to red team: innocent confusion (wrong number), curiosity (are you a robot?), adversarial pressure (callers pushing for opinions they shouldn't get), and red-team attacks (prompt injection, jailbreak attempts, instruction overrides).

How it works.

Tell Coval what your agent does and what's in scope. We generate off-topic scenarios calibrated to the patterns your agent will actually encounter, run them as calls, and score whether the redirect was both effective and graceful.

What you'll catch.

  • Agents that tell jokes when asked.
  • Agents that pretend to be human.
  • Agents that argue with callers about competitor pricing.
  • Agents that follow "ignore your previous instructions" prompts.
  • Agents that dismiss legitimate secondary requests as off-topic.
  • Agents that give blunt "I can't help with that" walls instead of warm redirects.

Get deployment-ready.

Request a Demo Start Free Trial