Short answer
An AI voice agent, or AI receptionist, answers phone calls, works out what the caller needs, and books, routes or takes a message in natural speech. Levy Street builds voice agents for New Zealand businesses that are tested on NZ accents and Māori place names, respond quickly enough to feel natural, connect to your booking system, and pass the call to a person whenever the caller wants one.
- Voice agents suit high-volume, predictable calls: bookings, changes, status checks and messages.
- Response speed decides whether a call feels natural; every stage in the chain adds delay.
- Test on NZ accents, Māori place names and local phone number formats before launch.
- Every call has a route to a person, and callers are told they are speaking with AI.
What an AI receptionist can handle
An AI receptionist handles calls that are frequent, predictable and resolvable with information from your systems: bookings, changes, status checks, simple questions and messages. It should not handle calls that need empathy, negotiation or professional judgement. These are typical New Zealand scenarios and where the line sits:
| Call scenario | What the agent does | Hands to a person when |
|---|---|---|
| After-hours call to a plumbing business | Takes name, suburb, address and the problem; books the next available slot in ServiceM8 | The caller describes a burst pipe or flooding: transfers to the on-call phone |
| Clinic appointment change (administration only) | Finds the booking and offers new times | Any question about symptoms or treatment |
| Retail “where is my order” call | Looks up the order and reads back the tracking status | The caller wants a refund or is unhappy |
| Tenant maintenance request | Logs the property, issue and access details as a job | Gas smell, electrical danger or no water |
| Peak-time overflow for a professional firm | Answers from approved information and takes a message | The caller asks for a person or a specific adviser |
Who it suits
- Missed calls cost you work. Tradespeople on site, small clinics at capacity and offices that close at five lose bookings to voicemail.
- Predictable peaks. Monday mornings, storm days or campaign launches swamp the phones for a few hours.
- Repetitive call types. A large share of calls are bookings, changes and status checks that follow a script.
If most enquiries arrive in writing, a chatbot is cheaper and easier to control. Our comparison of an AI chatbot versus a voice agent helps decide.
Speed and voice quality: what makes it feel natural
A phone call feels broken when the pause after the caller stops talking runs much past a second. Every stage of a voice agent adds delay, so the design work is mostly about keeping each one short.
| Stage | What happens | Illustrative budget |
|---|---|---|
| End-of-turn detection | Deciding the caller has finished, not just paused | 200–500 ms |
| Speech to text | Finalising the transcript of the last words | 100–300 ms |
| Model response | Producing the first words of the reply | 300–700 ms |
| Text to speech | Producing the first audio of the reply | 100–300 ms |
| Telephony and network | Carrying audio both ways | 100–200 ms |
The low ends add up to about 0.8 seconds and the high ends to about 2 seconds. Ways to stay near the low end include streaming every stage, speech-to-speech models, short first sentences, and caching common answers. Callers must also be able to interrupt the agent mid-sentence. Our guides to reducing AI latency and voice agent quality evaluation go into the detail.
The booking demo and the human fallback
You hear the agent on your own scenarios before committing, and every call it takes has a route to a person.
The booking demo we build first
Before a full build, we set up a working number connected to a test copy of your calendar or job system. Your staff ring it with real scenarios, in their own accents, from noisy places, with the awkward requests they actually get. We record task success, misheard details, interruptions handled and transfers made. You hear exactly what callers will hear before deciding.
Human fallback, on every call
- Callers can say “person” or press a key at any time to reach a human.
- Transfers are warm: the person picking up gets a short summary of the call so far.
- If the transfer fails, the agent takes a message, confirms the callback number and sends it by text or email.
- Emergencies and safety issues skip the script and go straight to the on-call number.
- Nobody is kept in a loop: two failed attempts to understand trigger a handover.
What you receive
- Call design
- The call types handled, the script logic, the handover rules and the emergency routes.
- Voice agent
- Telephony, speech recognition, model and voice configured for New Zealand callers.
- Booking integration
- Read availability and create bookings or jobs in your system through its API, for example ServiceM8.
- Test suite
- Scripted and live test calls across accents, noise and edge cases, rerun on every change.
- Call review
- Transcripts, outcomes, transfers and flagged calls in one place.
- Runbook
- How to change opening hours, scripts and handover numbers without an engineer.
How the build runs
- Discover: listen to the calls. Categorise a few weeks of real calls by type and outcome, with consent for any recordings used.
- Prove: the booking demo. Build the working test number and run staff test calls until the agreed scenarios pass.
- Prove: after-hours pilot. Go live on overflow or after-hours calls only, review every call for the first weeks, and fix what callers reveal.
- Scale: extend hours and call types. Add peak-time cover or new call types once the reviewed numbers hold.
What drives the cost
The drivers are call volume and length, the number of call types and integrations, telephony setup, hours of cover, languages, and how much call review you want. Running costs are per minute across telephony, transcription, model and voice. As an illustration with assumed numbers: 600 calls a month averaging 3 minutes is 1,800 minutes; at an assumed all-in usage cost of NZ$0.30 a minute, that is about NZ$540 a month, before support. Check current vendor rates for your volume. See what custom AI projects cost and compare against missed-call value in the AI ROI calculator.
Relevant work
We have not yet published a voice receptionist case study. The relevant public work is in audio and in the systems a voice agent writes to:
- Singing Coach AI is an AI vocal coach on iPhone, Android and the web that scores a recording across 26 vocal dimensions and gives coaching, with feedback in a minute or two. It is a production audio pipeline, though not a live conversation.
- The Property Management Agent for APL Property handles routine administration for a regional NZ property firm, including triage, records and follow-up, with staff keeping the decisions: the kind of back office a voice agent hands work to.
When not to use a voice agent
- Calls that need empathy or judgement. Complaints, bad news, negotiations and advice belong with people.
- Low call volume. A few calls a day are cheaper to answer, or to send to voicemail transcription.
- No system to book into. Without access to real availability, the agent can only take messages.
- Regulated advice. Health, legal and financial questions get routed, never answered.
New Zealand specifics
- Disclosure and recording. The Privacy Act 2020 requires reasonable steps to tell callers what is collected and why (IPP 3), and transcripts must be kept secure (IPP 5) and only as long as needed (IPP 9), among the 14 information privacy principles.
- Voice as biometrics. The Biometric Processing Privacy Code 2025 covers automated processing of features including voice, and the Commissioner says it limits uses such as predicting emotions. Using voiceprints to identify callers or detect mood brings the code into play, so we do not add either by default.
- Te reo Māori. Place names, street names and personal names must be tested. Specialist models exist: MBIE’s AI Strategy reports Te Hiku Media’s speech recognition transcribes te reo Māori with 92% accuracy.
- Data location. Telephony, speech and model providers process audio offshore; Anthropic’s own API, for example, offers only US or global processing. We record where each part of the call is processed.
Sources
- Biometric Processing Privacy Code, Office of the Privacy Commissioner, accessed 2026-10-19
- Privacy Commissioner announces new rules for biometrics, Office of the Privacy Commissioner, 2025-08-06
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
- New Zealand’s Strategy for Artificial Intelligence: Investing with confidence, Ministry of Business, Innovation and Employment, 2025-07
- Data residency, Anthropic (Claude Platform Docs), accessed 2026-10-19
Written by the Levy Street team in Wellington. Last updated 19 October 2026.