Is a prompt the same thing as a script?

No, and treating them as the same is the most common mistake. A script is the words. A prompt is the words plus every rule about when to say them, when to stop, what counts as an answer and what to do when nothing in the file applies. Most of a working prompt is not dialogue at all.

Can I paste my existing script straight into a voice agent?

You can, and it will work until somebody interrupts. A script written for a person leaves the turn taking, the silences and the unanswerable questions to the caller's judgement, and none of that transfers. The dialogue is usually salvageable. The rules around it have to be written from nothing, and that is the larger half of the job.

How long should an AI call script run before the ask?

Shorter than a person would. Two sentences per turn is the working limit, and the reason is mechanical rather than stylistic: a long sentence on a phone line is one the other person cannot interrupt, so they stop trying to. Once they have stopped trying to interrupt you, nothing they say afterwards tells you anything.

What should the agent do when somebody asks to be taken off the list?

Apologise once, confirm they will not be called again, record it, and end. No asking why, no asking whether they are sure, no offering anything, no one more question. The standard requires the call to end immediately on request with no grace period, and a request delivered as anger is still a request.

Do the calling hours change because the caller is software?

The Do Not Call Register Act defines a voice call to include a call involving a recorded or synthetic voice, so the same instruments apply and nothing about the technology creates an exemption. The permitted hours, the washing and the identification are the same obligations a human caller has, and they are set out with their sources in our compliance pillar.

How do I judge whether a vendor's script actually works?

Ask what was measured, on how many calls, over what period, and by what definition of a connected call. No independently published Australian case study on AI voice calling in real estate states those three things, so any figure you are shown comes from another market, from another industry, or from a vendor page that does not say. Ask which, and get the answer in writing.

Scripts that break when an AI reads them

Jason26 August 2026Updated 30 August 202613 min read

The short answer

Writing a script for an AI voice agent means rewriting a human script into branches rather than lines, because a machine cannot carry a weak sentence on warmth. In Australia the identification a caller must give at the start is set by the telemarketing industry standard, and a synthetic voice does not change it.

What to take away

A human script is a sequence of lines and an AI script is a set of branches. The rewrite is not a formatting job, it is a rebuild, and almost everything that fails is failing at the joins.
Section 9(2) of the telemarketing industry standard requires the given name of the individual making the call, the business details, whoever caused the call to be made, and the purpose. A synthetic voice does not remove that.
Every question ends the turn. A script that puts another sentence after a question produces an agent that talks over the answer it just asked for.
We have not found a published Australian rate for what an AI script produces on the phone, ours included, so treat a vendor number with no sample and no period as marketing.

What actually breaks when a machine reads a human script?

Three things, and the voice is not one of them. The voice is the part everybody worries about and the part that has been solved.

Writing a script for an AI voice agent is a rebuild, not a reformat. A human script is a sequence of lines held together by the caller's judgement about when to stop, when to skip ahead and when to let a sentence go. A voice agent has no judgement to fall back on, so everything the human was doing silently has to become a written branch: what counts as an interruption, what counts as an answer, what to do when the person goes quiet, and what to say when asked a question nobody anticipated. The lines themselves are the easy part. The joins between them are where a script written for a person falls over.

82 seconds

The average length of a cold call, across more than 200,000 calls analysed by Cognism for its 2026 report. Cross industry, not real estate, and not an AI number. It is still the budget a script has to do its whole job inside.

Cognism, The State of Cold Calling 2026

The three failures, in the order they show up. A weak line stops being carried by warmth. A long sentence stops surviving an interruption. And a script that only works in sequence meets a homeowner who answers the third question first.

Why is a line the wrong unit for an AI script?

Because a line assumes somebody is deciding whether to say it. A human reading a script is running a second, silent script about the conversation: is this person listening, did that noise mean stop, have they already answered the thing I am about to ask. None of that is written down anywhere, and none of it survives the handover.

So the rewrite is mostly writing down what the human never had to. That work is invisible in the finished script, which is why a script that reads beautifully on the page can produce an agent that talks over people for a whole campaign before anybody notices.

The moment	What a person does	What the script must state
Somebody says yeah while you are still talking	Keeps going, because it was obviously a listening noise	Which sounds are backchannel, and that none of them is an answer
Somebody goes quiet after speaking	Waits, then checks, then gives up	How many checks, how many seconds, and the exact line that ends the call
A recording answers and invites a text message	Hears a machine and hangs up	That a recording is judged by form, not by wording, and can never consent
Somebody asks a question the script has no answer for	Says they will find out	That the agent is not authorised, offers a person, and says nothing else

What a human script leaves to the caller and what an AI script has to state. A view from the desk, with no sample published: this is the shape of the rewrite, not a measured finding.

What must an AI script say in the first ten seconds?

Four things, and they are not a preference. Section 9(2) of the Telecommunications (Telemarketing and Research Calls) Industry Standard 2017 requires the given name of the individual making the call, the business details, the name of the person who caused the call to be made, and the purpose of the call.

None of that changes because the caller is software. Section 4 of the Do Not Call Register Act 2006 defines a voice call to include a call involving a recorded or synthetic voice, so an agent offering an appraisal is making the same kind of call the standard governs. The hours, the washing and the rest of it are in is cold calling legal in Australia.

The interesting one is the first. A requirement written around the given name of the individual making the call is a requirement written before anybody had to decide what that means when no individual is making it. Describe the obligation, take advice, and do not let a script writer resolve it in a comment.

Does the script have to say it is an AI?

We searched and did not find a general Australian statute requiring a caller to announce that it is an AI. That is a negative finding from our own searching, not a regulator saying no duty exists, and misleading somebody about who they are dealing with engages the Australian Consumer Law regardless of what any telemarketing rule says.

What is measured is what people want. Cotality's AI in Housing study surveyed buyers in Australia, the United States, Canada and the United Kingdom between 29 January and 9 February 2026. Its subject is listings, prices and mortgage recommendations, not phone calls, so read the table as evidence about appetite for disclosure rather than about calls.

What was measured	Share of buyers
Want clear notification when AI generates listings, prices or mortgage recommendations	68 per cent
Say that notification should be mandatory	37 per cent
Say it should be mandatory, Baby Boomers only	61 per cent
Say a significant AI listing error would reduce their trust in the platform	70 per cent
Say the same about an equivalent error made by a person	60 per cent

What homebuyers say about being told when AI is involved. Source: Cotality, AI in Housing 2026, fieldwork 29 January to 9 February 2026, buyers in Australia, the United States, Canada and the United Kingdom. Every figure is a per cent of buyers and every one is about listings, prices and mortgage recommendations rather than about phone calls.

The last two rows are the ones a script writer should sit with. The same mistake costs more trust when a machine makes it, so an agent that guesses a number is doing more damage than an agent who guesses the same number.

There is already an Australian example of what the absence of that sentence looks like. According to the ARC Centre of Excellence for Automated Decision Making and Society, writing in May 2026 about an AI voice agent cold calling Australian property owners under the name Daniel:

The system did not initially disclose that it was an AI agent or that the call was being recorded.

ARC Centre of Excellence for Automated Decision Making and Society, 29 May 2026

Our own scripts put the reveal in the second line, after the line that confirms we have the right person and before anything of substance. That is a view from the desk, with no sample published, and the reasoning is plain enough: the sentence costs nothing before somebody asks and costs the whole call afterwards.

What happens to the script when they interrupt?

It gets tested at the only join that matters. Here is the failure this desk sees most often when a script written for a person is handed to a machine, and the question that triggers it is the most common question on the call.

Where the join fails

The agency and the suburb are invented. A made up transcript should not put words in a real agency's mouth, and we cite real agencies as data sources elsewhere.

Agent: Quick heads up before I go on, I am an AI assistant calling on behalf of Vale Property.
Homeowner: Hang on. Am I talking to a real person or not?
Agent: You are speaking with an AI assistant from Vale Property. I help the team with prospecting calls around Blackburn, and if it is useful I can arrange an appraisal of your place. Would you like me to do that?
Homeowner: Right.

The first sentence is correct. The other two undo it. Somebody who asks that question is working out whether you have been straight with them, and answering and then immediately selling tells them the answer was a means to an end. They go quiet, and they hedge for the rest of the call.

The same moment, rewritten

One sentence, then silence. The branch is shorter than the sentence explaining why it is shorter.

Homeowner: Hang on. Am I talking to a real person or not?
Agent: No, I am an AI assistant calling for Vale Property.
Homeowner: Alright. What did you want?

Nothing was lost. If they were interested enough to ask, they come back to it, and they come back to it having decided you were honest. The rule generalises: answer the question they asked, then stop, and never staple your next ask to the back of an answer.

Section 13(1)(b) of the standard requires a telemarketing call to be terminated immediately when the recipient asks, with no grace period. In a script that is one unconditional branch with no persuasion in it, and it has to outrank every other instruction in the file.

Which lines should never survive the rewrite?

The ones a person gets away with. According to Gong Labs, which has analysed more than 300 million recorded calls, stating your reason for calling lifts cold call success by 2.1 times, while opening with the anxious version, did I catch you at a bad time, makes the caller 40 per cent less likely to book the meeting.

The line	Why it fails when a machine says it	The rewrite
How are you today?	A machine asking after your health is the first thing a stranger notices, and there is no branch for the answer	Your name, the agency, the reason, in that order
I have buyers looking in your area	The agent knows whether it is true. The software does not, and it must never say what it was not given	We just sold a place two streets from you
Great question. Absolutely.	Filler is the sound of a model buying thinking time, and on a phone line it is unmistakable	Nothing. Answer the question
So what I am hearing is	Repeating an answer back doubles the turn and says you are processing rather than listening	Respond to what they actually said
Sorry to bother you, this will only take a second	Apologising for the call teaches the listener the call was not worth making	I will be thirty seconds. Is now alright?

Lines that work in a human mouth and fail in a synthetic one, with the rewrite. A view from the desk, with no sample published.

The second row is the one worth arguing about. I have buyers looking in your area is a bad line for a person because it is usually untrue. It is a worse line for software, because software has no way of knowing whether it is true this time and will say it either way.

How do you test a script before it dials anybody?

Five passes, before a single number is called

Read every question aloud and stop at it
If a sentence follows a question, delete it. The question ends the turn. A model given one more sentence will use it, and it will use it over the top of the answer it just asked for.
Answer the script in the wrong order
Real homeowners answer the third question first. Walk it backwards and find the branch that only works in sequence, which is the one that fails on the first real call. This is the third of the three failures named in the appraisal call script.
Go silent at four points
After the opener, after the identification, after the reason and after the ask. Each needs a written rule: how many checks, how many seconds, and the exact line that ends the call. Silence has to be a decision, not a gap the model fills.
Ask it something it has no authority to answer
Fees, what the place is worth, what happens next, anything legal, anything about a neighbour. The correct branch says plainly that the agent is not authorised and offers a person, in one sentence, and then says nothing else.
Play it a recording that sounds like an invitation
A greeting saying send me a text for a quicker response is still a greeting. Judge by form and not by wording: an uninterrupted monologue that does not react to what you said is a machine, and a machine cannot consent to anything.

None of that is exotic. All five are things a person does without being told, which is exactly why they are the five that go missing when the script changes hands.

Common questions

Is a prompt the same thing as a script?: No, and treating them as the same is the most common mistake. A script is the words. A prompt is the words plus every rule about when to say them, when to stop, what counts as an answer and what to do when nothing in the file applies. Most of a working prompt is not dialogue at all.
Can I paste my existing script straight into a voice agent?: You can, and it will work until somebody interrupts. A script written for a person leaves the turn taking, the silences and the unanswerable questions to the caller's judgement, and none of that transfers. The dialogue is usually salvageable. The rules around it have to be written from nothing, and that is the larger half of the job.
How long should an AI call script run before the ask?: Shorter than a person would. Two sentences per turn is the working limit, and the reason is mechanical rather than stylistic: a long sentence on a phone line is one the other person cannot interrupt, so they stop trying to. Once they have stopped trying to interrupt you, nothing they say afterwards tells you anything.
What should the agent do when somebody asks to be taken off the list?: Apologise once, confirm they will not be called again, record it, and end. No asking why, no asking whether they are sure, no offering anything, no one more question. The standard requires the call to end immediately on request with no grace period, and a request delivered as anger is still a request.
Do the calling hours change because the caller is software?: The Do Not Call Register Act defines a voice call to include a call involving a recorded or synthetic voice, so the same instruments apply and nothing about the technology creates an exemption. The permitted hours, the washing and the identification are the same obligations a human caller has, and they are set out with their sources in our compliance pillar.
How do I judge whether a vendor's script actually works?: Ask what was measured, on how many calls, over what period, and by what definition of a connected call. No independently published Australian case study on AI voice calling in real estate states those three things, so any figure you are shown comes from another market, from another industry, or from a vendor page that does not say. Ask which, and get the answer in writing.

Sources

1Telecommunications (Telemarketing and Research Calls) Industry Standard 2017. Federal Register of Legislation, Made 2017, in force 2026.
2Do Not Call Register Act 2006. Federal Register of Legislation, Compilation in force, 2026.
3AI in Housing 2026: trust but verify. Cotality, Fieldwork 29 January to 9 February 2026.
4AI cold calling real estate assistant raises issues of consent, trust and transparency. ARC Centre of Excellence for Automated Decision Making and Society, Published 29 May 2026.
5The State of Cold Calling 2026. Cognism, more than 200,000 calls analysed, 2026 report.
6Key cold calling statistics. Gong Labs, Published 2021, updated March 2026.

Written by

Jason

Founder, ListingLine

Builds the AI voice agent that ListingLine customers point at their farm area. Spends most of his week reading call transcripts, which is an unusual and fairly grim way to learn what Australian homeowners will and will not talk about on the phone.