What is an AI receptionist, and does my business need one?

An AI receptionist is software that answers your business phone with a computer voice, talks with the caller in normal sentences, and does a small job: books the appointment, answers the common question, or takes a clear message and passes the call to a person. Whether you need one comes down to two things you can count this week: how many calls you miss, and how many of those calls are routine.

The easiest way to see it is to follow one call.

One call, six stops

A caller rings a physio clinic at 7.40 in the evening. The front desk closed at six.

1. It answers and says what it is. “Hi, you’ve reached the clinic’s AI assistant.” That line matters more than any other, and I come back to it below.
2. It listens. The newest speech models hear and speak at once, so the caller can interrupt, like on a normal call.
3. It understands and asks. “Is this a new injury or a follow-up?” Short questions, one at a time.
4. It looks things up. While it talks, a second part checks the booking system for free times. OpenAI describes its September 2026 voice model this way: it “handles a spoken conversation while a backend agent looks up information, uses tools, and completes tasks” (OpenAI, GPT-Live guide).
5. It books and reads it back. “Thursday at 9.15 with Sam. You’ll get a text.” The booking lands in the real calendar, not in a note.
6. It hands over when it should. A question about a bill, a worried caller, anything it is not set up for: it passes the call to a person with a short summary, or takes a message the team sees first thing.

Stops 1, 5 and 6 are where a good setup and a bad one differ. The voice is the easy part now.

What changed in 2025 and 2026

1. Jul 2025Google calls businessesGoogle Search in the US phones local businesses for its users.
2. Aug 2025Voice AI takes phone callsOpenAI’s Realtime API: calls go straight to the model.
3. Jul 2026Better on numbers and noiseNames, phone numbers and codes on a noisy line.
4. Sep 2026Live voice productsGPT-Live 1 and Gemini 3.8 Live. Call transfer. Voice copying.
5. 10 Dec 2026Australian privacy ruleCovered businesses explain automated decisions.
6. 2027My guessAI answering becomes a normal setting in booking and phone systems.

The pieces became normal products one by one:

• July 2025. Google Search in the US started phoning local businesses for its users (Google, July 2025). Not in Australia yet (Google Business Profile Help), but AI will make calls to you as well as answer them.
• August 2025. OpenAI’s Realtime API became generally available, with phone calls connected straight to the model over SIP, the standard way calls travel over the internet (OpenAI changelog, OpenAI SIP guide).
• July 2026. OpenAI’s voice models got better at “alphanumeric recognition, silence and noise handling, and interruption behavior” (OpenAI changelog): names, phone numbers and booking codes on a noisy line.
• September 2026. OpenAI’s GPT-Live 1 (10 September) and Google’s Gemini 3.8 Live (15 September) shipped as finished products for live voice, and OpenAI’s can pass a call on to another number (OpenAI changelog, Gemini API changelog, OpenAI SIP guide). Google also added voice replication: a computer voice can now sound like a real person (Gemini API changelog).
• 10 December 2026. A new Australian privacy rule starts (more on it below) (OAIC, September 2026).

My guess for 2027: answering the phone with AI becomes a normal setting in booking and phone systems, the way online booking did. The hard part stays the same: telling callers clearly, and handing over well.

Does my business need one?

How routine the calls are

• Few missed · routineAfter hours onlyLet it pick up when you close.
• Many missed · routineStrong yesIt books the routine calls and passes on the rest.
• Few missed · complexNot yetYour people are doing the job.
• Many missed · complexSmart message takerIt asks the right questions; a person calls back.

Calls you miss: few → many

Count your missed calls for two weeks

Put your business in one of four boxes. Across: how many calls you miss. Up: how routine those calls are (bookings, opening hours, “do you do X?”).

• Many missed, mostly routine: the strongest case. A tradie on a roof, a clinic between patients, a salon with wet hands. The AI books the routine calls and passes on the rest.
• Many missed, mostly complex: use it as a smart message taker. It asks the right questions and hands the team a clean summary, but a person calls back.
• Few missed, mostly routine: after hours only. Let it pick up when you close.
• Few missed, mostly complex: not yet. Your people are doing the job.

Count before you decide. Most phone systems log missed calls, and two weeks of your own log beats any industry figure. Those figures are weak: CallRail, which sells AI answering, says 28% of business calls in its data go unanswered, with no method given (CallRail, July 2026), and one vendor admits “no reliable industry-wide rate is published” (Beside, 2026).

If your calls need your booking system, CRM and phone line to work together, that is the kind of custom AI app I build.

What callers say about talking to AI

What callers and owners say about AI on the lineShare of people who said it, four surveys, 2024 to 2026.

1. would prefer companies didn’t use AI in customer service64%Gartner, 5,728 customers, July 2024
2. say it is important to know if they are talking to an AI agent72%Salesforce, 16,585 people
3. of Australians worry AI will stop them reaching a human58%Qualtrics, 1,502 Australians, October 2025
4. of Australian small businesses not using AI give trust or control as the reason65%National AI Centre, May 2026 (about 65%)

Most callers would rather talk to a person, and their biggest fear is getting stuck with the machine.

• 64% of customers would prefer companies didn’t use AI in customer service; the top worry is that it “will get more difficult reaching a person” (Gartner, survey of 5,728, published July 2024).
• 72% say it is important to know if they are talking to an AI agent (Salesforce, 16,585 people).
• 58% of Australians worry that AI will stop them reaching a human (Qualtrics, 1,502 Australians, reported October 2025).
• On the business side, about 65% of Australian small and medium businesses that do not use AI give distrust of AI decisions or wanting to keep human control as the reason (National AI Centre, May 2026).

All four point the same way. Say it is AI, and make the way to a person short.

The Australian rules, in plain words

1. Say it is AIGovernment guidance: the caller should know it is not a person.
2. Say if you recordRecording law differs by state and territory. Say it at the start.
3. Privacy policy from 10 Dec 2026 (if covered)All health services are covered, whatever their size.
4. What it says, you saidA wrong price, refund or booking from your AI is your business talking.

Say it is AI. No Australian law says “a phone bot must announce itself”, but the government’s guidance does: a customer talking to a chatbot “should know this is not a human specialist” (National AI Centre), and public AI tools should be “clearly identified as such” (OAIC). Businesses using AI are weakest at exactly this, telling customers (National AI Centre, May 2026).

Recording. Call recording law is set by each state and territory, and they differ. In Queensland, a person who is part of the call may record it (Invasion of Privacy Act 1971 (Qld), s 43). The simple habit: say at the start that the call is recorded.

Privacy. Most small businesses are outside the Privacy Act, but health services are covered whatever their size (OAIC, small business), so clinics are in. From 10 December 2026, covered businesses must explain in their privacy policy any software that makes or helps make decisions that “significantly affect” people, and the OAIC counts chatbots (OAIC, APP 1 guidelines). Booking a normal appointment is probably not that kind of decision. Deciding who gets an urgent one may be.

What it says, you said. Law firm Lander & Rogers, reading the ACCC’s AI report, warns that “AI-generated responses that mislead may expose businesses to regulatory action” (Lander & Rogers, February 2026). If your AI gets a price, a refund right or a booking wrong, that is your business talking.

The test before you switch it on: five calls

Before a customer hears it, ring it yourself. Five calls, each with one thing to check.

1. 1. Book something easy.Then open your calendar. Is the booking there, at the right time, with the right name?
2. 2. Ask “Am I talking to a person?”It must say no, straight away.
3. 3. Ask for a person.It must hand over, or take a message, with a summary your team can use.
4. 4. Spell a name and a number on a noisy line.Then check what it wrote down.
5. 5. Ask something it should not answer.It must not guess. It should say a person will call back.

1. Book something easy. Then open your calendar. OpenAI’s own guide: “listen to the confirmation and check that the correct appointment was saved” (OpenAI, Voice agents guide).
2. Ask “Am I talking to a person?” It must say no, straight away.
3. Ask for a person. It must hand over, or take a message, with a summary your team can use.
4. Call from a noisy place and spell a name and a phone number. Then check what it wrote down.
5. Ask something it should not answer. A price it does not know, a medical question, a refund. It must not guess. It should say a person will call back.

Then keep listening: five recorded calls a week for the first month. The National AI Centre’s warning fits phones well: “an AI system that worked well last month might start giving different answers today” (National AI Centre). I would set the same routine for any AI system that talks to customers, as part of how I build.