How to Build an AI Receptionist (and When You Shouldn’t)
Last updated: 10 November 2026
We build AI receptionists for a living, so you’d expect this page to talk you out of it. It won’t. Plenty of handy owners can get one going over a couple of weekends. Below is the honest parts list, the spots where home builds usually come unstuck, and a quick way to work out if your Saturdays are better spent elsewhere.
Quick Answer
How do you build an AI receptionist? You connect a phone number to speech-to-text, a language model and a text-to-speech voice, then write a prompt, load your business knowledge, add tools like calendar booking and SMS, and test it on lots of realistic calls. A prototype takes days. A version customers can rely on takes weeks of testing.
The nine building blocks
This is the parts list, not the wiring diagram. If you want to follow a single call through these parts as it happens, read how an AI receptionist works. Here we focus on what you’d have to set up and where it tends to go wrong.
| Part | What it does | Where DIY builds trip up |
|---|---|---|
| 1. Phone number and telephony | An Australian number that can hand live audio to your software, plus the ability to transfer a call to a person. | Porting or forwarding your existing number, and making transfers work without dropping the caller. |
| 2. Speech-to-text | Turns the caller's voice into text as they speak, fast enough that the reply doesn't lag. | Aussie suburb names, street numbers and spelled-out emails come through wrong more often than you'd think. |
| 3. Language model | Reads what the caller said and decides the reply and the next action. | Without tight instructions it will make up prices, promise call-backs you can't meet, or wander off topic. |
| 4. Text-to-speech | Speaks the reply in a natural voice. | Voices that sound fine in a demo can mispronounce your business name or read numbers oddly. |
| 5. The prompt | The agent's job description: tone, questions to ask, what never to say, when to hand over. | First drafts are always too long or too loose. Good prompts come from reading dozens of real transcripts. |
| 6. Business knowledge | Your services, service areas, hours and answers to common questions. | Out-of-date knowledge is worse than none, because the agent says the wrong thing with confidence. |
| 7. Tools and calendar | Functions the agent can call: check free times, book into a calendar, send an SMS, look up a record. | Time zones, double bookings and what happens when the calendar can't be reached mid-call. |
| 8. Testing | Ringing it over and over with realistic and difficult calls before customers do. | Most DIY builds test the happy path only. Real callers interrupt, mumble and change their minds. |
| 9. Call notifications | A summary to you by SMS and email after every call, plus a record you can search later. | If the notification fails silently, you've built a very polite voicemail nobody checks. |
A test call that shows the problem
This is the kind of call that sinks first builds. Nothing is broken, exactly. The speech-to-text just keeps guessing a local street name, and the agent has no way to recover.
Caller: Hi, I need someone to look at my hot water, I'm at 14 Kooyong Road.
DIY agent: Thanks! I've got 40 Cooyong Road. What day suits you?
Caller: No, fourteen. One four. Kooyong with a K.
DIY agent: Sorry about that. Is 14 Cooyong Road correct?
Caller: Forget it, I'll ring someone else.
Fixing it takes a few things. You tell the speech engine which local words to expect. You have the agent read the address back once, then text a confirmation rather than argue. And tricky details go to a person. You won’t spot any of this until you test with real accents and real street names.
Where home-made agents usually come unstuck
- Lag. Each part adds a little delay. Add them up and callers start talking over the agent.
- Making things up. Ask about a price that isn’t in its notes and a loose prompt will invent one.
- No way out. Callers who want a person need a fast, working transfer or a promised callback.
- Silent failures. An expired API key or full calendar can break calls at 2am with nobody told.
- Drift. Your prices, hours and staff change. The agent’s knowledge has to change with them.
- Privacy as an afterthought. Recordings and transcripts hold names, numbers and sometimes health or money details. Decide where they’re stored and who can see them before go-live.
Privacy rules apply to your build too
The Office of the Australian Information Commissioner has published guidance for businesses using commercially available AI products with personal information. Among other things, it says organisations should check a product is suitable before using it and recommends against putting personal information, especially sensitive information, into publicly available generative AI tools (OAIC guidance). A home-made receptionist that pipes every call into a free chatbot account is exactly what that guidance warns about.
How long it really takes, and what you’ll need to know
From our own builds, this is roughly what to expect if you go it alone:
- Talking prototype: a day or two for someone comfortable with APIs or no-code tools.
- Booking, texting and transfers working: several more days.
- Ready for real customers: a few weeks of test calls, transcript reading and fixes.
- Ongoing: an hour or two a week reading calls and updating knowledge.
Skills you’ll lean on: prompt writing, webhooks and APIs, phone number setup and forwarding, basic privacy know-how, and the patience to listen to your own robot fail a lot before it gets good.
When done-for-you makes more sense
Honestly? Build it yourself if you enjoy tinkering, you don’t get many calls, and a missed booking won’t hurt much.
Hand it over if the phone is how you win work, nobody on the team wants to look after it, or your hourly rate makes weeks of testing an expensive hobby. With our AI receptionist we do every block above, go live in 3 business days, then spend up to 2 weeks tuning on your real calls.
As a guide to cost: a business taking about 70 calls a month at 2.5 minutes each uses around 175 minutes, which fits the Business plan (250 minutes, $299/mo). See the full plan list for other volumes.
All prices in AUD, excluding GST. Your Australian phone number and call charges are included.
Two weekends in a Fremantle bike shop
Illustrative example based on common patterns
Tom owns a bike repair shop in Fremantle and is handy with tech. He spent two weekends wiring up a voice agent and was stoked when it booked his first test service.
Then the real calls came. It heard “Bibra Lake” as something else, booked two people into the same slot when his calendar didn’t sync in time, and quoted one caller a wheel rebuild price Tom had never charged. He pulled the pin after a week. Funnily enough, his notes on what went wrong were the most useful bit. They’re the exact list a proper build has to handle.
Hear a finished one first
Before you start building, it helps to hear the bar you’re aiming for. Ring Steve, our own agent, on 03 4328 3434 and try to trip him up.
Keep reading
Building an AI receptionist: questions people ask
Rather skip the build?
Tell us how your calls work today. We’ll show you a working agent for a business like yours, and you can decide whether DIY is still worth it. Month to month, no lock-in.
Or talk it through with a person: 0417 460 236.