Skip to main content

​
Why Testing Matters

The risks of skipping testing:
  • Agent gives incorrect information to real customers — damages trust
  • Edge cases cause confusion or repetitive loops
  • Functions fail silently — customers get no answer
  • You discover problems after launch, not before
Rule of thumb: Run at least 20 test conversations before going live.

​
Three Levels of Testing

​
Level 1: Browser Test (Fastest)

Test directly in the Studio — no phone number required.
1

Open the Studio

Go to studio.talkifai.dev/agents and open your agent.
2

Click Try Now

Find the Try Now button on the agent card. You’ll be redirected to the demo page at /demo.
3

Start Voice Test

Click Start Voice Test button. The browser will ask for microphone permission — click Allow.
4

Have a conversation

Behave like a real user. Watch the live transcription to see what the agent hears and understands. Take notes on anything that feels off.
Note: You’re testing YOUR agent, not a demo agent. All conversations use your configured agent settings.Text agents: If your agent uses Text architecture, use the Chat tab to test text-based conversations instead of voice.Session Duration: Demo sessions last up to 30 minutes. The timer will show remaining time.
Use the mute button to temporarily silence your microphone without ending the call.

​
Level 2: Phone Test

Call your agent on a real phone to verify carrier setup, audio quality, and latency.
  1. Assign a phone number to the agent — Telephony Guide →
  2. Call from your own phone
  3. Verify: audio quality, response latency, call routing

​
Level 3: Team Testing

Before going to production, have 2–3 team members test independently.
  • Give them specific scenarios to test
  • Collect structured feedback
  • Fix issues before real customers encounter them

​
Complete Testing Checklist

​
✅ Basic Functionality

Agent answers the call correctly?
Introduction is correct (name, company)?
Voice sounds right (speed, tone, quality)?
Language is correct?
Agent understands its role?

​
✅ Conversation Quality

Handles normal questions accurately?
Remembers context within the call (e.g., uses name after asking)?
Responses sound natural — not robotic?
Responses are concise — not overly long?
No unnecessary repetition?

​
✅ Functions (If Configured)

Functions are called at the right time?
Function results are used correctly in the response?
Function failures are handled gracefully?
User is informed when the agent is fetching data?

​
✅ Edge Cases

Handles unknown questions gracefully?
Redirects when user goes off-topic?
Stays calm with an angry or rude user?
Handles very fast speech?
Handles silence of several seconds?
Works with background noise?

​
✅ Escalation & End

Escalation triggers correctly when expected?
End call works properly?
Transfer works (if configured)?

​
Test Scenario Library

Copy and run these scenarios during testing:

​
Customer Support Scenarios

Scenario 1 — Happy Path:
"I'd like to track my order. The order ID is ORD-12345."

Scenario 2 — Information Not Available:
"Can I get an invoice from January 2022?"

Scenario 3 — Angry Customer:
"This is the third time I'm calling about the same issue and nothing has been fixed!"

Scenario 4 — Escalation Request:
"I need to speak with a manager immediately."

Scenario 5 — Off-Topic:
"By the way, what city are you based in?"

​
Appointment Booking Scenarios

Scenario 1 — Simple Booking:
"I'd like to book an appointment for tomorrow."

Scenario 2 — Fully Booked:
"Is there anything available today?" (test when slots are full)

Scenario 3 — Reschedule:
"I need to change my existing appointment."

Scenario 4 — Wrong Context:
"I'd like to book an appointment for my dog."
(For a human clinic — does it handle gracefully?)

​
Sales Scenarios

Scenario 1 — Qualified Lead:
"Yes, I'm very interested. My budget is around $500/month."

Scenario 2 — Not Interested:
"No thanks, I'm not looking for anything right now."

Scenario 3 — Price Negotiator:
"Can you just do it for $50?"

Scenario 4 — Competitor Mention:
"I'm already using [Competitor Name]."

​
Red Flags to Watch For

Symptom: Asks the customer’s name twice, or doesn’t use the name after learning it.Fix: Add to system prompt: “Use the customer’s name throughout the conversation once you know it.”
Symptom: Agent speaks 5–6 sentences per response.Fix: Add to prompt: “Every response must be a maximum of 2 sentences. Never exceed 30 words.”
Symptom: When asked something out of scope, the agent repeats itself or gives a strange answer.Fix: Add an explicit edge case handler to the prompt.
Symptom: Order status function fires before the user asks for it.Fix: Improve the function description — be very specific about when it should be triggered.
Symptom: “Order I.D. one two three four” sounds choppy.Fix: Add to prompt: “Spell out all numbers and abbreviations naturally — say ‘order number one-two-three-four’ not ‘order ID 1234’.”

​
Performance Benchmarks

After going live, monitor these metrics in Analytics →:
MetricTargetIf Below Target
Call completion rate> 80%Agent is frustrating — review the prompt
Function call success rate> 95%Check API/webhook connection
Escalation rate< 20%Too many escalations = prompt is unclear
Average call durationUse-case specificToo short = unhelpful; too long = inefficient

​
Production Ready Checklist

Only go live when all of these are checked:
20+ test conversations completed
All basic scenarios pass
All edge cases handled correctly
All functions working properly
Voice sounds natural and appropriate
Escalation path works correctly
At least 2 team members have tested independently
Agent is Activated in the Studio

​
Next Steps

Connect a Phone Number

Set up real calling after testing is complete.

Monitor with Analytics

Track live calls and continuously improve.