Voice AI that answers the phone, books the job and hands over
Missed calls are the most expensive thing in a service business, and voicemail has never worked. A voice assistant is worth building when the calls you are missing are the simple ones.
What is voice AI?
Voice AI is a phone assistant that answers calls, understands speech and completes tasks such as taking a booking, capturing a job request or routing a caller to the right person. It suits Australian businesses losing calls after hours or during peaks, where the common requests are well-defined and staff remain available for everything else.
Get a fixed written quote- Typical timeline
- 8 to 16 weeks
- What drives cost
- How many call types the system handles, which systems it has to look into, and how much of the conversation has to be scripted and tested.
- Best for
- After hours and overflow calls with predictable, repeatable requests
- You own
- The phone numbers, the call data, the recordings and the accounts
- Built with
- Telephony integration, real time speech, booking and CRM handover
Your handover
Voice is the hardest modality, and it is worth saying why
Text gives a model time to think and a reader who can reread. Voice gives neither. The assistant has roughly a second to respond before the caller assumes the line has dropped, so the whole pipeline of listening, understanding, deciding and speaking has to fit inside that budget. It also has to cope with people interrupting mid sentence, which is normal human behaviour and technically awkward, and with everything happening in the background of a real Australian phone call: a ute on the highway, a busy salon, a worksite, a mobile with one bar in the Adelaide Hills.
- 01Call analysis with intents ranked by volume
- 02Voice assistant handling the agreed intents
- 03Telephony routing with unconditional failover to your existing line
- 04Disclosure and recording notice wording per jurisdiction
- 05Australian vocabulary hints for suburbs, streets and product names
- 06Booking or job capture integrated with your systems
- Warm transfer carrying transcript and context to staff
- Call transcripts and summaries logged into your CRM
- Reporting on containment, transfer rate and cost per call
More on voice is the hardest modality, and it is worth saying why
Then there is speech recognition itself. Broad Australian accents, callers whose first language is not English, place names that no general model has heard of and street names that sound like three other street names all produce errors. Suburb and product name lists get loaded in as vocabulary hints for exactly this reason, and the assistant confirms anything consequential by reading it back. Anyone claiming voice is a solved problem has not listened to a hundred real recordings from an Australian phone line.
Anyone claiming voice is a solved problem has not listened to a hundred real recordings from an Australian phone line.
Disclosure, recording and consent under Australian law
Callers should be told they are speaking with an automated assistant, in the first sentence, in plain words. It is the honest thing to do and it also improves the interaction, because people speak differently and more clearly once they know. Every deployment we build offers an immediate way to reach a person, and we do not build assistants designed to pass as human.
Recording is a legal question rather than a technical one
Recording is a legal question rather than a technical one. Rules for recording telephone conversations sit in state and territory surveillance and listening devices legislation and are not uniform across Australia, so the notice and the consent handling are configured for the jurisdictions you operate in and reviewed by your own adviser before go live. On top of that, a recording or transcript containing personal information falls under the Privacy Act 1988: collect only what you need, disclose the purpose at the start of the call, retain for a defined period, and secure it under APP 11. If audio is processed or stored overseas, that is a cross border disclosure under APP 8 and it needs to be documented. We default to Australian regions for the recording and transcript store.
How the engagement runs
How a voice AI deployment runs
We build against your recordings. Before writing a single flow we listen to real calls, transcribe a sample and count what people actually ring about, which is reliably different from what the business believes they ring about. Reception staff are the best source of truth in the whole project and are usually the last people consulted.
- 01Call analysisA sample of real recordings categorised by intent, duration and outcome
- 02ScopeWhich intents the assistant handles, which it transfers, and the wording of the transfer
- 03TelephonyNumbers, routing, business hours logic and an unconditional failover to your existing line
- 04VocabularySuburb, street, product and staff name lists loaded as recognition hints
- 05Conversation designShort turns, confirmation read backs, and a graceful path when the caller is unclear
- 06Disclosure and consentThe opening announcement, the recording notice and an always available option to reach a person
- 07IntegrationBookings into your calendar or booking system, call summaries and details into your CRM
- 08After hours pilotEvery call reviewed, transcripts scored, then business hours considered separately
Two decisions on your side that keep the project moving
Deployment starts in the least risky window, normally after hours, where the alternative is voicemail rather than a person. That sets a fair benchmark: the assistant only has to beat a recorded message, which it does easily. Extending into business hours is a separate decision made on the evidence from that first period.
Choose the right level
The calls worth automating and the ones to leave alone
Voice AI works when the request is common, the information needed is short, and the outcome is a defined action. It struggles when the caller is upset, when the request is unusual, or when the value of getting it right dwarfs the cost of a person answering. Triage is the deciding design question, and it happens in the first few seconds of the call rather than after a long menu.
Call type
01
Opening hours, location, parking, basic pricing questions
Automate
Yes
Why
Short, factual, and the caller wants speed over conversation
02
Standard bookings, reschedules and cancellations
Automate
Yes, with a confirmation read back
Why
Defined slots, and the outcome is verifiable in your calendar
03
Taking a job request or lead details after hours
Automate
Yes
Why
Beats voicemail comfortably, and the details land in your CRM
04
Complaints and account disputes
Automate
No, route straight through
Why
The caller wants a person with authority to fix it
05
Clinical, legal or financial advice
Automate
No
Why
Regulated advice, and a wrong answer spoken aloud is hard to retract
06
Emergencies and welfare concerns
Automate
Never
Why
Immediate human or emergency service, tested on every release
How we work this out during scoping
One rule we apply without exception: anything resembling an emergency reaches a human or an emergency number immediately, with no attempt at qualification. A gas smell, a burst main, a distressed patient, a caller who says they are not safe. That path is tested on every deployment and it never gets optimised for efficiency.
Handover, failover and what happens at two in the morning
Handover is where most voice deployments are judged. A caller who has already explained the problem and is then asked to repeat it from the beginning will remember that far longer than they remember the assistant. So transfers carry context: the transcript so far, the intent, any details captured, delivered to the person picking up. Outside staffed hours the equivalent is a structured summary in your CRM with a callback flagged, not a voicemail nobody plays.
Failover matters just as much
Failover matters just as much. Speech services and model providers have outages, and a phone line that fails silently is worse than never having automated it. Every build has an unconditional fallback to your existing routing if the assistant does not respond within a set time, so the worst case is the situation you had before. We also monitor cost per minute, since voice bills across telephony, speech recognition, the model and speech synthesis at once, and a chatty assistant on a busy line adds up faster than a text channel ever does.
When voice AI is the wrong fit
If you take fifteen calls a day and someone answers them all, this is not your problem. If your callers are elderly, distressed or ringing about something sensitive, a person is the right answer and the goodwill lost is not worth the saving. If your booking rules are genuinely complicated, involving multiple practitioners, room availability and travel time, the assistant will stumble and the underlying fix is a proper booking system before any voice layer goes near it.
The rest of the answer
It is also worth checking whether the calls need to happen at all. A large share of inbound calls to service businesses are status enquiries, and they exist because customers have no other way to find out. A simple job status page or a text message when a technician is on the way removes those calls entirely, at a fraction of the cost and with none of the risk. When the question is really about deflecting routine contact rather than answering the phone, an AI chatbot or SMS automation frequently gets there sooner. Practices and clinics weighing this up should read how we approach healthcare projects, since patient contact carries obligations a general assistant does not account for.
How we scope it
Four ways to scope your Voice AI project
We do not publish package prices, because the same brief can be a short build or a long one. These are the shapes the work usually takes. Tell us which one sounds like you and you will get a fixed written quote that spells out exactly what it covers.
Proof of value
One use case, evaluated honestly before it goes near a customer
Fixed written quote, agreed before work starts
- Call analysis with intents ranked by volume
- Voice assistant handling the agreed intents
- Telephony routing with unconditional failover to your existing line
Production build
In production, with a human approval step and an evaluation set
Fixed written quote, agreed before work starts
- Everything in Proof of value
- Disclosure and recording notice wording per jurisdiction
- Australian vocabulary hints for suburbs, streets and product names
- Booking or job capture integrated with your systems
Embedded platform
Built into the product rather than bolted onto it
Fixed written quote, agreed before work starts
- Everything in Production build
- Warm transfer carrying transcript and context to staff
- Call transcripts and summaries logged into your CRM
- Reporting on containment, transfer rate and cost per call
Voice AI Model care
Monitoring, evaluation and retraining as the inputs drift
Rolling monthly, quoted in writing
- Evaluation set rerun as the model and the inputs change
- Cost and quality reported monthly, not assumed
- Prompt, tool and guardrail changes as the work shifts
- Rolling, cancel with 30 days notice
These are shapes, not menus. Most quotes end up somewhere between two of them, and we will say so when the honest answer is the smallest one. Describe the problem and we will tell you which it is.
Questions buyers usually ask
Frequently asked questions
How long does a voice AI deployment take?
Around 8 to 16 weeks. Call analysis and conversation design take longer than people expect, and the after hours pilot needs several weeks of real traffic before anyone should judge it. Integrating with a booking system or CRM adds time depending on how open those systems are. Simple call capture with no scheduling can land at the shorter end.
What does voice AI cost to operate?
Voice bills on several meters at once: telephony minutes, speech recognition, the model and speech synthesis. The practical measure is cost per call, which we report from the pilot. Longer conversations cost more, so conversation design has a direct commercial effect. The build itself is quoted in writing after we have listened to a sample of your calls.
Will callers know they are talking to a computer?
Yes, because we tell them in the first sentence. Voice quality is convincing enough now that disclosure is a deliberate design choice rather than something the technology forces, and we think it is the right one. Callers can also ask for a person at any point and the assistant transfers without argument or a retention script.
Who owns the phone numbers, the recordings and the call data?
You do. Numbers stay on your account or port to a carrier account in your name, recordings and transcripts sit in storage you control, and CRM records are in your CRM. We work under access you control and can pull at any time. If you stop using the assistant, calls route back to your existing arrangement with nothing to unpick.
What happens if the assistant cannot understand the caller?
It tries once to clarify, then transfers or takes a callback rather than looping. Repeated failed attempts are the fastest way to lose a customer, so the design deliberately gives up early. Every unrecognised call is transcribed and reviewed, which is how vocabulary and intents improve over the first months of operation.
Can it take bookings straight into our calendar?
Yes, where your booking system has a usable API. The assistant checks live availability, holds a slot, confirms the details back to the caller and writes the booking, then sends a confirmation by text. Where availability rules are complex, we usually recommend fixing the booking logic first, since a voice layer over ambiguous availability produces double bookings rather than efficiency.
Related services
Start with the calls you are already missing
Tell us how many calls go unanswered after hours and what most of them are about. Inside a business day you will have a pilot scope and a fixed price in writing.