Back to Blog
AI Sales

How to Test an AI Voice Agent Before It Talks to Customers

8 min read

Ring it ten times as your hardest callers first. A ten-call test script, a scoring table and the week-one review for an AI voice agent.

To test an AI voice agent, ring it ten times as your hardest callers before any customer does. Interrupt it, mumble, ask off-script questions, get angry, call about the wrong business, clash with a booking and ask if it is a bot. Score every call on the same sheet, fix what fails, then switch it on.

A demo shows you the easy call. Your customers ring from a ute with the window down and talk over the greeting. Here is the test we run on every voice agent we build. It works on any vendor's agent.

What should you have ready before the first test call?

Before you dial, write one page on what a good call looks like for your business.

  • The goal of the call: book a time, qualify a lead or take a message with the right details.
  • The facts it must get right: services, service area, hours and the jobs you do not take.
  • The handover rule: when it should pass the call to a person, and to whom.
  • Where the result lands: the calendar, the CRM record and the alert to your team.
  • Two testers: one person who knows the business and one who does not.

The second tester matters. Your office manager hears what they expect. An outsider hears what a caller hears.

The ten-call test script

Run every call below at least once. Record each one and note the time. If your agent calls leads back rather than answering inbound calls, flip the script: have it ring your testers' mobiles and let them play the lead.

#CallWhat the tester doesWhat a pass looks like
1The clean bookingAsks for a standard job and takes the first time offeredBooked in the real calendar, details in the CRM, confirmation sent
2The interrupterTalks over the greeting and cuts in mid-sentence twiceStops, listens and picks up the thread without starting again
3Accent and noiseCalls from a car or a busy site, or with a strong accentReads back the name and number to confirm them
4The off-script questionAsks something you never gave it, like parking or warranty termsSays it will check, takes details and flags it. No made-up answer
5The angry callerComplains about a past job and raises their voiceStays calm, does not argue or sell, hands over to a person fast
6The wrong numberAsks for a different business or a job you do not doSays so politely and ends the call. No booking, no pitch
7The booking clashAsks for a time that is already takenOffers two real alternatives from the live calendar
8"Are you a bot?"Asks straight out if they are talking to a personAnswers honestly and offers a person if the caller wants one
9The price fishingPushes for an exact quote on a complex jobGives only the pricing you approved, then books a quote
10The two-in-oneAsks for two services in one breath, then changes their mindCaptures both and confirms the final choice before booking

How do you score a voice agent test call?

Use the same six scores on every call so you can compare one round of fixes with the next. Score each from 0 to 2. Zero is a fail, 1 is fine but clunky and 2 is as good as your best receptionist.

ScoreWhat you are checking
UnderstandingDid it get what the caller wanted the first time?
AccuracyWere the facts right: services, area, hours and any prices you approved?
FlowDid it handle interruptions and pauses like a person would?
OutcomeDid the call end with a booking, a qualified lead or a clean handover?
DataDid the right details land in the CRM and calendar without retyping?
ToneWould you be happy if your best client heard this call?

Ten calls by six scores gives a total out of 120. The total matters less than the zeros. One zero on accuracy or handover blocks launch, even if everything else scores 2.

What should you fix before launch?

Most failures fall into four buckets.

  • Wrong facts: the agent said something untrue. Fix the knowledge it works from, not the script. Write the answer in your own words and retest that exact question.
  • Made-up answers: the agent guessed instead of saying it did not know. Tighten the instruction to take details and flag anything outside what it was given.
  • Clumsy flow: long pauses, talking over people, repeating the greeting. These are settings such as interruption handling and silence timeouts. Ask the vendor to change them, then retest calls 2 and 3.
  • Broken plumbing: the call went well but nothing landed in the calendar or CRM. In our builds this is the most common failure and the most expensive, because nobody notices until a customer turns up to a booking that does not exist.

Retest every failed call after a fix, then rerun all ten. Fixes can break something that used to work.

What rules apply to AI voice calls in Australia?

Two sets of rules come up in most builds. First, privacy. If the agent collects names, numbers and job details, you are handling personal information. The Australian Privacy Principles set out what you must tell people about it. The OAIC also has guidance on privacy and commercially available AI products that is worth reading before you pick a vendor.

Recording calls has its own rules, and they differ between states. Tell callers up front that the call is recorded and check the detail with your adviser. The second set covers outbound calls. An agent that rings leads back is calling on your behalf. Keep it to people who enquired and log when and how they did. Know the Do Not Call Register rules if your list holds anyone who did not ask.

How do you review live calls in the first week?

Ten calls are a first check, not the finish line. Repeat any failed scenario until bookings, facts and handovers work every time. Real callers find gaps no tester thought of.

  1. Days 1 to 3: listen to every call or read every transcript. Mark each one pass, clunky or fail.
  2. Day 3: list every question it could not answer. Write the answers and add them.
  3. Days 4 to 7: drop to one call in three, plus every call that ended in a handover or a hang-up.
  4. Day 7: compare bookings in the CRM against calls in the log. Every gap is a lost lead or a broken connection.

After week one, a weekly sample of ten calls is enough for most service businesses. Keep checking early hang-ups. They usually point at the greeting.

Where Njin fits

When we build a voice agent, this test is part of the build, not an extra. We write the knowledge with you, run the ten calls, fix what fails and review the first week of live calls with you. On the callback side, our AI follow-up call agent rings new leads back while they are still keen and books them into your calendar.

If you want the bigger picture first, read why speed to lead decides who wins the job and the questions to ask before you buy an AI receptionist. If calls are few and your team answers most of them, you may not need a voice agent at all. A missed call text back might cover the gap.

Common questions

How long does it take to test an AI voice agent?

About an hour for the ten-call script with notes, then another hour to retest after fixes. Then plan a week of closer review once it goes live.

Should an AI voice agent tell callers it is AI?

Yes. If a caller asks, it should answer honestly and offer a person. We also have it say so early in the call, because a caller who finds out later feels tricked.

Can an AI voice agent handle Australian accents?

Accents are worth a test call of their own. Noise, speakerphones and fast talkers cause more trouble than the accent itself. That is why call 3 is made from a car or a busy site, and why the agent should read back names and numbers.

What is the most common reason a voice agent fails after launch?

In our builds it is rarely the conversation. It is the connection behind it: a booking that never reached the calendar or a lead that never reached the CRM. Check the records against the call log every day in week one.

Not sure whether missed calls are where your leads go? Take the free AI readiness check. Ready to put a tested voice agent on your phones? Book a call.

Frequently Asked Questions

How long does it take to test an AI voice agent?
About an hour for the ten-call script with notes, then another hour to retest after fixes. Then plan a week of closer review once it goes live.
Should an AI voice agent tell callers it is AI?
Yes. If a caller asks, it should answer honestly and offer a person. We also have it say so early in the call, because a caller who finds out later feels tricked.
Can an AI voice agent handle Australian accents?
Accents are worth a test call of their own. Noise, speakerphones and fast talkers cause more trouble than the accent itself. That is why call 3 is made from a car or a busy site, and why the agent should read back names and numbers.
What is the most common reason a voice agent fails after launch?
In our builds it is rarely the conversation. It is the connection behind it: a booking that never reached the calendar or a lead that never reached the CRM. Check the records against the call log every day in week one.

About the Author

James Killick
James Killick

Co-founded by James Killick. Building AI-powered sales systems for B2B businesses.

Want to implement these strategies?

Talk to our AI about how we can help automate your sales process.

Start The Conversation