AI

AI voice agents on agency phones: recording, disclosure and call rules

Software that answers or makes an agency's calls is covered by the same telemarketing, recording and privacy laws as a person. What each requires in Queensland.

· 15 min read

Kooky
Written by
Kooky

Builder of Shaka, the payment router that pays every agent their commission on closing date.

About Kooky and Shaka →

A caller rings a real estate office at 7.40 on a Saturday evening about a house they saw online. A voice answers, confirms the address, says the next open home is at 11 the following morning, offers to book a private inspection and asks for a name and mobile number. The voice is pleasant and quick. It is also software.

Tools of this kind, usually called AI voice agents, are now sold to agencies for two jobs. One is answering inbound calls when no one is at the desk. The other is making outbound calls: working through a database to revive old enquiries, confirm appointments or ask owners whether they are thinking of selling. The proptech industry's own awards this year featured one product whose maker claimed an enormous increase in the number of outbound calls an agency could place in a day.

No Australian statute is written specifically for a synthetic voice on a phone line. That does not leave a gap. A call made or answered by software is still a call, a recording is still a recording, and a name taken down is still personal information. This guide sets out the existing rules that apply, in the order an agency meets them. It is general information about the law as published by the regulators, not advice on any particular system.

2directions of call, with very different rules
9am to 8pmweekday window for marketing calls
$222,000infringement ceiling for each day of breaches

Sources: Telemarketing and Research Calls Industry Standard 2017; Do Not Call Register, compliance and breaches.

Inbound and outbound are different problems

The first thing to settle about any voice agent is which way the calls go, because the law treats the two directions very differently.

An inbound call is one the customer chose to make. The person rang the agency, about something they wanted. The questions are about what the software tells them, whether the call is recorded and what happens to the details they give.

Related readDeepfakes and cloned voices: how property identity checks respond

An outbound call is one the agency chose to make. If its purpose is to promote the agency's services or a property, it is telemarketing, and a body of rules built to restrain telemarketing applies in full: the Do Not Call Register, calling hours, and requirements about what the caller must say.

The rules by direction of call
IssueInbound: the customer ringsOutbound: the software rings
Do Not Call RegisterDoes not ariseApplies to marketing calls
Calling hoursNoneWeekdays 9am to 8pm, Saturdays 9am to 5pm
RecordingState and federal recording lawThe same
Privacy noticeRequired when details are collectedRequired, plus the rules on marketing
Accuracy of what is saidConsumer law and agents' conduct rulesThe same

Most of the legal risk in a voice agent sits in the right-hand column. An agency that uses the software only to answer its own phones has a modest list of things to get right. An agency that points it at a database has taken on the whole of telemarketing law, at a volume no human team could reach.

Outbound calls are telemarketing calls

The Do Not Call Register Act 2006 makes it unlawful to make an unsolicited telemarketing call to a number on the register unless the person has consented. The Act is concerned with the purpose of a call, not with who or what speaks. The Australian Communications and Media Authority, which runs the register, describes a telemarketing call as a voice call made to offer or promote goods, services or land.

The ACMA's information sheet for the real estate industry gives the trade's own examples: a call offering a free appraisal, a call following up after a property inspection, and a call soliciting a listing. Those are exactly the tasks outbound voice agents are sold to perform. A call does not stop being one of them because the appraisal is offered by software.

Consent is therefore the centre of the matter. The ACMA distinguishes express consent, where a person clearly agrees to be called, from inferred consent, which arises from an existing relationship. Its guidance for real estate is narrow on the second. Where the dealing was a single transaction, such as a sale or a purchase, consent can generally be inferred only while that transaction lasts, and calls months after settlement are unlikely to be reasonable. Consent may be inferred for calls about a property a person enquired about, but not about unrelated listings.

Related readGenerative AI on a conveyancing file: what Queensland's rules say

That guidance was written for human callers working through a list at human speed. A voice agent applies the same list in an afternoon. An old database of open-home visitors and past enquirers, most of whom agreed to nothing in particular, is the natural thing to feed it and the thing the law is least likely to permit. Numbers without recorded consent have to be checked against the register first, and a checked list is good for 30 days.

Liability rests with the agency. The register's guidance states that a breach is committed by the person who makes a call or causes it to be made, and that those who aid or are knowingly concerned in a contravention can be liable as well. An agency cannot hand the responsibility to the software vendor, and the vendor is not safe from it either. The penalties the register publishes run to $222,000 in infringement notices for each day on which contraventions occurred, and to $2.22 million a day in court.

The main risk

Software makes it easy to break telemarketing law thousands of times before lunch

The rules on consent and the Do Not Call Register apply to each call. A system that dials an unwashed database does not commit one breach. It commits one for every registered number it reaches.

Hours and what the caller must say

A second set of rules governs how any telemarketing call is made, whether or not the number is registered. The Telecommunications (Telemarketing and Research Calls) Industry Standard 2017 applies to anyone who makes or arranges such calls to Australian numbers.

Three of its requirements matter for a voice agent.

Calls may be made only at permitted times: weekdays from 9am to 8pm and Saturdays from 9am to 5pm. Telemarketing calls are not permitted on Sundays. Software has no natural sense of this and will call at any hour it is told to, so the limits have to be built into its schedule, in the time zone of the person being called.

Related readA rent figure from software: what Queensland's Act asks before it runs

The caller must give certain information as soon as the call starts, including the name of the caller and the purpose of the call. A voice agent's opening script has to do that. An opening designed to sound like a friend ringing back does not.

The call must be ended when the person asks, or otherwise indicates that they do not want it to continue. This is a test of the software's comprehension as much as its script. A system that hears "not interested" and moves to its next persuasive line is in breach. And calling line identification must be switched on, so that the person sees a number they can ring back.

The register's guidance gives the penalty for a breach of the standard as up to $250,000 for each contravention.

Recording the call

Almost every voice agent records. The recording is how the system produces a transcript, a summary and a note in the agency's database. Two laws bear on it.

In Queensland, the Invasion of Privacy Act 1971 regulates the use of listening devices to record private conversations. The Queensland law firm Bennett & Philp summarises the rule: a person may record a private conversation to which they are a party, and it is an offence to record one to which they are not. Queensland is, in that sense, a state where a party to a conversation may record it.

The same summary points to the limit that matters more for a business. Even where the recording itself was lawful, the Act places strict controls on communicating or publishing it without the other party's consent, with exceptions that include use in legal proceedings and protection of the recorder's lawful interests. A recording that is transcribed, stored on a vendor's servers, reviewed by staff and used to improve a product has been communicated to a number of people.

Related readVirtual staging and AI-edited photos: what the law allows

The federal Telecommunications (Interception and Access) Act 1979 sits alongside. It makes it an offence to intercept a communication passing over a telecommunications system without the knowledge of the person making it. A system that captures the call within the phone network, as a software voice agent typically does, is closer to that law than a person holding a recorder near a handset.

The safe course under both is the familiar one. Tell the caller at the start that the call is being recorded and why, and give them a way to proceed without it. The announcement takes a few seconds and removes the question. Callers are not all in Queensland, either. Other states have their own listening device laws, some stricter, and a system that announces recording on every call does not have to work out where each caller is.

Does the software have to say it is software?

There is no provision in the telemarketing rules or the Privacy Act that says, in terms, that an automated voice must announce itself. Several existing rules lead to the same place.

The industry standard requires the caller to identify themselves at the start of a call. A system that gives a human name and lets the person believe a staff member is ringing has, at the least, made that identification misleading.

The Australian Consumer Law prohibits misleading or deceptive conduct in trade or commerce, and conduct can mislead by what it leaves out. Whether a person would have behaved differently had they known they were speaking to software is the kind of question that provision asks.

Related readAgency chatbots and tenants: what an automated reply cannot do

The privacy regulator expects it. In guidance published on 21 October 2024 on the use of commercially available AI products, the Office of the Australian Information Commissioner said that organisations should update their privacy policies and collection notices to explain clearly when and how AI will access, use and generate personal information, and that an AI system should be transparent about how it uses personal information. The law firm MinterEllison, summarising the guidance, records the regulator's advice to consider whether a chatbot is an appropriate way to collect information at all, particularly sensitive information.

Taken together, these make disclosure the only defensible design: an opening that says the caller is speaking with an automated assistant for the agency, that the call is recorded, and that a person can be reached on request.

What a compliant opening covers
  1. Who is callingThe agency's name, and that the voice is an automated assistant acting for it.
  2. WhyThe purpose of the call, stated at the start as the industry standard requires.
  3. What is keptThat the call is recorded, and how to speak to a person or end the call.

The personal information it collects

A voice agent is a collection device. In a two-minute call it can gather a name, a phone number, an address, a budget, the fact that someone is selling because of a separation or a death, and a recording of their voice.

For an agency covered by the Privacy Act, the Australian Privacy Principles apply to all of it. The regulator's guidance on AI products works through them. Personal information should be collected only where reasonably necessary for the organisation's functions, and by lawful and fair means. The person must be told about the collection. The organisation must take reasonable steps to keep the information accurate.

Three points in that guidance are particular to AI.

What the system generates about a person is personal information too. A summary that says a caller is "motivated, likely to sell within three months" is information about that person, and if the system has it wrong, an inaccurate record has been created. The guidance states that incorrect output about an identifiable person, including what the industry calls hallucinations, must be handled as personal information.

Related readAI-drafted condition reports: what Queensland tenancy law still asks

Using recordings to train a product needs care. The regulator's view is that training AI on personal information will be hard to justify as something the person would reasonably expect, and that organisations should seek consent with a meaningful way to refuse. An agency should know whether its vendor uses call recordings for that purpose.

Due diligence comes first. The guidance expects an organisation to check, before adopting an AI product, that it suits the purpose, how it was tested, who can access the information and what the security risks are, and to do a privacy impact assessment. For a voice agent the practical questions are where the recordings are stored, for how long, and whether the vendor's staff can listen to them.

What it may and may not say

A voice agent speaks for the agency. Whatever it says about a property, the agency has said.

That matters because the things callers ask about are the things the law is strictest on. The Australian Consumer Law prohibits false or misleading representations about land, including its price, location and characteristics. Queensland's Property Occupations Act adds its own rules for agents, including the ban on giving a price guide for a property going to auction.

A system that generates its answers, instead of reading them from a fixed script, can be wrong with complete fluency. It may tell a caller that a house going to auction is "expected to sell around" a figure it found in the agency's notes. It may say a property is in a school catchment that it is not in. It may describe a unit's body corporate levies from last year's listing. Each of those is a representation by the agency, and the fact that software composed it is not a defence.

Related readAI suburb forecasts and growth scores: what an agent may pass on

The remedy is restraint in design. The system should answer from the listing's approved details and from nothing else, decline questions about price where the law forbids an answer, and pass anything it is unsure of to a person. An agency should be able to say, of any statement its system made, where the statement came from.

On the record

Every answer the system gives is the agency's answer

A wrong statement about price, size or zoning made by software carries the same consequences as one made by a salesperson. The recording that proves the caller was told also proves what they were told.

Work that needs a licensed person

Queensland regulates who may act as a real estate agent. Negotiating a sale, appraising a property for a listing and collecting money on a client's behalf are functions for people who hold a licence or a registration certificate under the Property Occupations Act, working for a licensed agency.

A voice agent can lawfully do the things a receptionist does: answer, take a message, give out published information, book an inspection. The further it moves towards the work of a salesperson, the less clear the footing becomes. A system that tells a caller what the seller "would take", relays offers back and forth, or advises an owner what their home is worth is doing more than answering the phone.

No published guidance from the Office of Fair Trading on where that line falls for automated systems was found in preparing this guide. The cautious reading is that a tool may gather and pass on information, and that anything resembling negotiation or an opinion of value should come from a licensed person. The agency's licensee remains responsible for the conduct of the business either way.

Records, and who checks the system

The useful side of a system that records everything is that an agency can see what it did. The regulator's guidance on AI expects exactly that: human oversight built into how a product is used, monitoring of its outputs, and staff trained in its limits.

For a voice agent, oversight has ordinary components. Someone reads a sample of transcripts each week. Complaints and requests not to be called again are logged and acted on, so that a person who said "do not ring me" on Tuesday is not rung by the same system on Thursday. The consent behind each outbound list is recorded, with the date the list was checked against the register. And recordings are kept only as long as there is a reason to keep them, since the same file that helps the agency in a dispute is a liability if it leaks.

The law gives an agency reason to start with the inbound side. An after-hours assistant that announces what it is, records with notice, answers from the listing and hands over to a person has few legal obstacles. An outbound campaign has all of them at once.

Kooky, from Shaka

Kooky edits Queensland Estate and builds Shaka, the payment router he made for Queensland property professionals. One payment comes in, and every agent, agency and party in the deal receives their signed share on closing date.