Australia
Voice agents, built and run in Australia.
An Australian voice agent takes your inbound calls in the words you wrote, on an Australian number you buy or keep. A judge grades your test calls before a real caller reaches the agent.
Where it runs
Onshore, in four numbers
Data at rest
Sydney
AWS ap-southeast-2
Model inference
Sydney
AWS Bedrock
Voice pipeline
Sydney
Fly.io
Built in
Perth
Vertical AI Pty Ltd
What you are choosing
A platform, not a phone-answering service
You own it
The agent is yours to change
Nobody rewrites your prompt on your behalf, and there is no agency between you and the call. Open the editor, change a line, publish. The version that answers is the version you published.
Swap any part
Model, listener, voice
Change the language model, the speech recognition or the voice per agent, and calls keep running. You are not locked to one vendor's stack because a salesperson bundled it.
Read the calls
Every call, turn by turn
Transcripts, timings and scores for real calls, not a monthly summary. Most fixes are one line in the prompt, and reading the call is how you find which line.
Prove it first
Scored before it launches
Scenarios you write, graded by a judge, run on demand. The failing call shows up in your test results instead of in a complaint.
Hand it over
A person is one tool call away
Straight hand-off or a screened warm transfer that announces the caller. Configured per agent, and honest with the caller when no destination is set.
Describe, do not configure
Build it by talking to it
Everything in the editor is reachable from the AI assistant you already use, so you can set an agent up and change it in plain language instead of learning a console.
What it costs
The rate, on the page
You pay for the minutes a caller is on the line. There is no plan fee, no minimum term, and no call to book before you start. These are the same rates quoted on voice agent pricing in AUD, the Retell comparison and the Bland comparison.
| What you are billed for | Rate | Notes |
|---|---|---|
| Talk time, default model | A$0.25 a minute | The Fast model with a standard Cartesia voice |
| Talk time, premium voice | A$0.40 a minute | A premium ElevenLabs voice on the same model |
| Plan fee | A$0 | No monthly tier and no minimum term |
| Starting credit | A$20 | Minted when you add a card, lasts ninety days, draws down against usage |
| Inbound number | Rented monthly | An Australian landline or a 1300 line, priced in the workspace before you buy |
Where it runs, in detail
Including the parts that leave
Most of a call stays in Sydney. Two parts of it can leave: a model you select yourself, and speech synthesis on three of the four voice providers. The table names both. The full component-by-component list, with every provider named, is at docs.meddle.sh/docs/trust/data-residency.
| Component | Where it runs | The catch |
|---|---|---|
| Database and file storage | Sydney | Transcripts, agent configuration, embeddings and audit logs, on AWS ap-southeast-2 |
| Voice runtime and web app | Sydney | No selectable alternative, so nothing here can be misconfigured out of the country |
| Language model, default | Sydney and Melbourne | The Australian Bedrock inference profile. 4 of the 32 selectable models run in Australia, and the picker labels each one |
| Speech recognition | Australia | Our provider's Australian endpoint, on every environment that serves calls |
| Speech synthesis | Follows the voice | 1 of the 4 synthesis providers has an Australian endpoint. On the other three the reply text is synthesised in the United States and the audio comes back |
How you build one
Five steps, in the order you would guess
01
Describe the job
Write what the agent says, what it knows, and the rules it follows. Plain sentences, not a flowchart. Do it in the editor, or by describing it to the AI assistant you already use.
02
Give it something to look up
Point it at your prices, your hours, your policies. It reads from that instead of inventing an answer, and says so when the answer is not there.
03
Give it something to do
Book the job, send the text, check the order, pass the call to a person. Each one is a tool the agent can reach for mid-call, and you choose which.
04
Test it until it stops surprising you
Write the calls you dread as scenarios and play them at the agent. A judge grades each run, so you find the one that fails before a caller does.
05
Put it on a number
Buy an Australian inbound number in the workspace and route it to the published version, or keep the number you have and forward it. Every call after that is scored and readable turn by turn.
Where teams start
Three jobs worth giving it first
After five
The night line
Out of hours is the first job most teams hand an agent, because after five it only has to beat a voicemail box. Give it a small set of queries it finishes end to end and leave the rest on voicemail: the out-of-hours pattern.
Overflow
When the queue breaks
Losing one of four agents in a centre taking 30 calls an hour moves the average wait from 21 seconds to 107. Losing two makes it infinite. The Erlang C working is on the queue maths page.
Opt-in
Press 2 for the agent
Add the agent as a choice in the wait message rather than routing callers to it. A caller who presses 2 has opted in, which makes them the best early adopter you have: the queue opt-in pattern.
Questions
What Australian buyers ask first
Where does an Australian call actually get processed?
Do I need a new phone number?
Can the agent put someone through to a human?
How do I know it works before I point real callers at it?
What does an Australian voice agent cost here?
Who am I buying from?
Does this work for outbound calling too?
Build one and call it.
The fastest way to judge a voice agent is to ring it. Build yours, place a test call, and read the transcript.