AI voice agents for sales: how they work, where they fit and the Australian rules

An AI voice agent for sales is software that answers or makes phone calls and holds a real conversation. It listens to the caller, works out what to say, and replies in a natural-sounding voice. It's good at fast, repetitive work: calling new leads straight away, qualifying them, booking meetings, following up and picking up inbound calls after hours. It struggles more with long, delicate negotiations. In Australia, a sales call made with an AI voice is still a telemarketing call, with the same obligations as a call made by a person.

This guide explains how these agents work, where they tend to fall down, and what to check before you put one on your phones.

How an AI voice agent works, in plain English

Every AI voice agent is made of a few parts joined together. Knowing what each part does makes it much easier to judge a demo and ask vendors better questions.

Speech recognition: turning sound into words

The first step is listening. A speech recognition model turns the caller's audio into text as they speak. It has to deal with phone-quality audio, background noise, accents and people talking over each other.

For Australian businesses, check how well it handles local accents, place names, suburb names and your industry's jargon. If the agent mishears a street name or a product, everything after that goes wrong. Ask whether you can add custom vocabulary, such as your product names and the suburbs you service.

The language model: deciding what to say

The transcribed text goes to a large language model, the "brain" of the agent. It reads the conversation so far, along with your instructions, and decides on the next reply. Your instructions set its goal (book a call, qualify the lead, answer a question), the facts it may use and the limits it must keep to.

This is where most of the business logic sits. A well-set-up agent knows your pricing rules, which questions to ask, when to hand over to a person and what it must never promise. A badly set-up one improvises. The language model can also call other tools during the conversation, such as checking a calendar, looking up a customer in your CRM or logging the outcome.

The voice: turning words back into speech

The reply is passed to a text-to-speech engine, which produces the voice the caller hears. Modern voices sound natural, with realistic pacing and tone. Pay attention to how the voice handles numbers, addresses, email spellings and Australian pronunciation, because these are where synthetic voices often sound odd.

The phone line: connecting to the real world

Finally, the agent needs a connection to the phone network. That usually means a telephony provider that handles phone numbers, calls in and out, call transfers and recordings. This layer decides practical things: which number shows on the recipient's phone, whether calls can be transferred to your team, and how call records reach your CRM.

Where latency and interruptions trip agents up

Each of those parts takes a little time. The audio has to be transcribed, the model has to think, the voice has to be generated and the sound has to travel down the line. Add those together and you get a pause between the caller finishing and the agent replying. In a normal conversation, people notice long pauses fast. A slow agent feels robotic, and callers start talking again or hang up.

Some things that affect latency:

• Where the servers are. Audio that travels overseas and back adds delay, which matters for Australian callers.
• The size of the language model. Bigger models can reason better but often respond more slowly.
• Streaming. Good systems start speaking as soon as the first part of the reply is ready instead of waiting for the full answer.
• Tool calls. Checking a calendar or CRM mid-call adds a pause, so the agent should fill it naturally ("Let me just check that for you").

Interruptions are the other big test. People cut in, say "yeah, yeah" while listening, or change their mind halfway through a sentence. The agent has to tell when the caller has actually finished speaking and when they're just pausing to think. It also has to stop talking when interrupted and respond to what was said, rather than carrying on with a script. If it's too eager, it talks over people. If it's too cautious, it leaves awkward silences.

The best way to test this is to call the agent yourself and behave like a real prospect: interrupt it, mumble, ask an off-script question, go quiet, say "hang on". A polished demo recording tells you very little.

Where AI voice agents fit in a sales team

AI voice agents work best on high-volume, time-sensitive tasks that follow a clear pattern:

• Speed to lead. Calling a new web enquiry straight away, while the prospect is still thinking about you.
• Qualification. Asking the same core questions every time and recording the answers in your CRM.
• Booking. Putting qualified prospects into a salesperson's calendar.
• Follow-up. Persistently and politely following up leads that people tend to forget.
• Inbound cover. Answering calls after hours, during busy periods or when the team is out on site.

They're a weaker fit for complex enterprise deals, emotionally charged conversations, complaints that need judgement, and anything where a wrong answer has serious consequences. In practice, the strongest setup is usually a partnership: the agent handles first contact and routine work, and hands warm, qualified conversations to your people. Make sure the handover is smooth, with the person receiving a summary so the prospect doesn't have to repeat themselves.

Australian rules: an AI voice is still a telemarketing call

This is the part many business owners miss. Using AI doesn't put your calls in a separate category. The ACMA states that commercial calls made using an auto-dialler or that use pre-recorded or AI voices are all telemarketing calls. The Do Not Call Register Act 2006 also defines a voice call to include one that involves a recorded or synthetic voice.

A "check-in" call with a sales angle still counts. The ACMA says that where a call has more than one purpose and one of them is to sell or promote goods or services, the call is commercial and must comply with the telemarketing rules.

The telemarketing industry standard applies to anyone who makes or arranges telemarketing or research calls to Australian numbers, including numbers that aren't on the Do Not Call Register. Under the Telecommunications (Telemarketing and Research Calls) Industry Standard 2017, the obligations that matter most when you set up an AI agent include:

• Calling times. Unless express consent was given in advance, calls must not be made on a weekday before 9 am or after 8 pm, on a Saturday before 9 am or after 5 pm, or on a Sunday.
• Banned days. Unless express consent was given in advance, calls are banned on New Year's Day, Australia Day, Good Friday, Easter Monday, Anzac Day, Christmas Day and Boxing Day, and on a weekday holiday given in lieu of one of them.
• State holidays. The standard doesn't ban calls on a State or Territory public holiday unless it's already a prohibited day, but it doesn't exclude State or Territory laws that restrict calling further.
• Whose time zone. Calling times are measured at the usual residential address of the account holder for the number being called, not your office's time zone. Your dialler needs to know where the person lives.
• Identification. As soon as the call starts, the caller must give their given name (unless the call is made solely using a recorded or synthetic voice), the business name of their employer, the name of the person causing the call to be made, and the purpose of the call.
• Contact details on request. If asked, the caller must immediately give the contact details of their employer and of the person who caused the call, and who handles inquiries or complaints.
• A way to ask for those details. Where a call involves recorded or synthetic voices, the caller must provide a way during the call to request contact and complaint-handling details, for example by pressing a button to talk to an operator.
• Ending the call. The caller must immediately end the call if the recipient asks for it to be ended or otherwise indicates they don't want it to continue. Your agent has to recognise "not interested, bye" just as reliably as a person would.
• Caller ID. You must enable calling line identification, with a number that lets the recipient call back to find out who called and why, and that's reasonably likely to accept return calls for at least 30 days.

The Federal Register of Legislation lists the standard as in force and due to sunset on 1 April 2027, so watch for any replacement. Consent and Do Not Call Register washing are separate topics again. For anything specific to your situation, check with the ACMA or a lawyer.

Questions to ask before you choose

Use these when you speak to any vendor:

• Can I call the agent myself, unscripted, before signing?
• How fast does it respond on Australian phone lines, and where is the audio processed?
• How does it handle interruptions, background noise and Australian accents?
• Can I control exactly what it says, what it must never say and when it hands over to a person?
• Does it enforce calling hours based on the recipient's location, including national and state holidays?
• How does it identify itself and the business at the start of a call, and how can a recipient reach an operator?
• Does it end calls immediately when someone asks?
• What number shows on caller ID, and what happens when people call it back?
• Where are recordings and transcripts stored, and who can access them?
• How do call outcomes get into my CRM?

Getting started

Start small. Choose one clear job, such as calling new web leads or answering after-hours calls. Write the script and the rules, test it hard yourself, then review real call recordings every week and adjust. Expand only once the agent is reliably doing that one job well and within the rules.

If you want to see what this looks like in practice, Dialessa is an AI voice sales agent that calls new leads in seconds, follows up, and answers inbound calls. It's built so your team spends its time on the conversations that most need a person.