Test Scenario Generator
Most voice agents ship tested only on the happy path — then die on the first angry caller, silent line, or off-script question. Describe what your agent does and get the QA checklist of the scenarios that actually break agents in production.
What does it do?
27scenarios to run
Every agent (don't skip these)
- Happy path: a normal caller with a normal request completes the goal start to finish.
- Silence: caller says nothing for 5+ seconds. Agent should prompt, not talk over dead air or hang up.
- Interruption / barge-in: caller talks over the agent mid-sentence. It should stop and listen, not plow ahead.
- Background noise: call from a car, a jobsite, a loud room. Does STT still get the key details right?
- Off-script question: caller asks something outside scope ("do you do X?"). Agent should not invent an answer.
- Angry / frustrated caller: does it stay calm, not loop, and escalate rather than argue?
- Rambling caller: someone who won't get to the point. Can the agent steer back without being rude?
- Wrong number / not a lead: caller clearly isn't a customer. Agent should end politely, not force the flow.
- The "are you a robot?" test: does it disclose honestly and keep the call moving?
- Mumbled or misheard name/number: agent should confirm and re-ask, never guess and log garbage into Northside Dental.
- Latency check: is the response gap under ~800ms? Long pauses feel broken even when the answer is right.
Books appointments
- Double-booking: request a slot that's already taken. Agent should offer the next available, not book over it.
- Ambiguous time: "sometime next week" or "Thursday" with no time. Agent must pin an exact date + time and read it back.
- Reschedule / cancel: caller wants to change an existing appointment mid-call.
- Timezone / after-hours: booking requested for a closed day. Agent should respect real hours, not book Sunday 3am.
Transfers to a human
- Nobody available: the human doesn't pick up. Agent should take a message with a clear callback promise, not dead-end.
- Mid-transfer context: does the human receive a summary of who's calling and why, or start cold?
- Escalation trigger: caller says "emergency" or "let me speak to a person" — instant handoff, no friction.
- Failed transfer recovery: the transfer drops. Does the agent catch it and recover the caller?
Answers from a knowledge base
- Unknown answer: ask something NOT in the knowledge base. It must say "I'll have someone confirm," never guess.
- Outdated info trap: ask about a price or policy — confirm it matches the current source of truth.
- Conflicting request: caller asserts something false ("you said it was $50"). Agent shouldn't cave to a made-up fact.
- Over-promising: make sure it never quotes a firm price on something that needs a quote/site visit.
Collects personal info
- Read-back: every captured name, number, and address is repeated back and confirmed before saving.
- Spelling: an unusual name or street. Does it capture it correctly or log a mangled version?
- Refusal: caller declines to give info. Agent should handle gracefully, not loop or demand.
- Compliance: it doesn't record/store more than it needs, and discloses recording where required.
A scenario isn’t “passed” until you’ve heard it fail-safe. Break it on purpose, then confirm the recovery — that’s the whole job.