What is an AI receptionist, and what can it actually do?

An AI receptionist is a phone answering service run by software instead of a person. It picks up when you cannot, tells the caller it is an AI, finds out what they need, and passes it to you in a form you can act on. It does not run your business. It does the first two minutes of the call, which is the part you were missing while you were under a sink or with a client.

The reason owners look at this at all is missed calls. Quo, a business phone provider, analysed 16.7 million missed calls across its customers over three months and reported that 69% of missed business calls did not get a callback within 48 hours. That is the provider’s own customers, not a South African sample, and it is a vendor’s own analysis. It does show how easily a missed call turns into a lost one.

What happens on a call, from start to finish?

The steps below are the same whichever product you use. Where a product skips one, ask why.

1. It answersThe call reaches it when you do not pick up, when the line is busy, or after hours, depending on how you set the forwarding with your phone provider.
2. It says it is an AIA good one opens with something like: "Good morning, you have reached Karoo Plumbing. I am an AI assistant and I can take your details for the team." Being upfront is the honest way to start the call. See do you have to tell customers they are talking to an AI.
3. It understands the requestTwo pieces of technology do this. Speech to text turns the caller’s voice into written words. A language model, the same kind of software behind chat assistants, reads those words and decides what the caller wants and what to say next.
4. It asks the questions you would askFor a plumber: what is wrong, where is the property, is water running now, what is the best number. For a salon: which service, which stylist, which day. You write these questions once and it asks them every time.
5. It checks the diaryIf it is connected to your calendar, it can see which slots are free. If it is not, it can only promise that someone will call back.
6. It takes the detailsName, number, address, the problem in the caller’s own words, and how urgent it sounds.
7. It hands over or lines up the bookingEither it transfers the call to a person, or it lines up a booking for you to confirm. Which one it does depends on how you set it up and on what the caller said.
8. It sends you a summaryA short message or email: who called, what they want, how urgent, what was agreed, and what you need to do next.

What is it good at?

• Routine calls with a pattern. Quote requests, appointment bookings, opening hours, directions, "are you in my area". Most small-business calls fall in this group.
• Never being busy. It can take several calls at the same time, which a person at a desk cannot.
• Asking every question, every time. A tired person skips the address. Software does not.
• Writing things down. You get the summary in one place, not on the back of an invoice.
• Calls outside working hours. Evenings and weekends are when many trades and guest houses lose enquiries.

What is it bad at?

• Angry or upset callers. Someone with a flooded kitchen or a complaint wants a person who can say sorry and mean it. A good setup hands these over quickly. A poor one keeps asking its questions.
• Unusual requests. Anything outside what you told it. It should say it does not know and pass it on. The risk is a confident wrong answer, which is why you test it.
• Strong accents on a bad line. Speech to text gets worse with noise, a weak mobile signal, several people talking or a caller in a car. It may mishear a street name or a surname, so confirm details back to the caller and check the summary.
• Anything needing judgement. Whether to take a job, whether a price can be negotiated, whether a symptom is an emergency. These stay with you or your staff.
• Prices and promises. It should only quote what you have given it, and it should not agree to discounts, refunds or times you have not approved.

Does it work in Afrikaans, isiZulu and isiXhosa?

It depends on the provider and on the voice model behind it, and we cannot give you one answer for the whole market. English is the safest. For the other languages, the evidence we could check shows support exists but quality differs.

ElevenLabs, a speech-technology vendor, says its Scribe speech to text models handle 90+ languages. For its Scribe v2 model it sorts them by word error rate, which is how many mistakes appear in a transcript compared with a correct one. On its page Afrikaans is in the "Good (>10% to ≤20% WER)" band, and Xhosa and Zulu are in the "Moderate (>25% to ≤50% WER)" band. Its published figure is a lab measure for that vendor’s model, not a measure of a phone line in Mossel Bay.

Other providers list the languages differently. Microsoft’s Azure Speech documentation lists af-ZA (Afrikaans), en-ZA (South African English) and zu-ZA (isiZulu) for speech to text, and has no isiXhosa row. Google Cloud’s Speech-to-Text language list shows Afrikaans, Zulu and Xhosa on its newer Chirp family of models. For Xhosa it also lists the older long and short models in some regions, with only the profanity filter as a feature. That list covers speech to text only; we did not check Google’s speech output. A language being listed tells you it is offered. It does not tell you how well it copes with your callers.

The talking side, the voice that speaks back, is a separate model from the listening side, and quality can differ between the two. On the same Azure page, the text to speech table lists voices for af-ZA, en-ZA and zu-ZA and none for isiXhosa. Mixed speech is also common here, such as a sentence that moves between English and Afrikaans, so ask the provider how it copes. Test it with real callers before you trust it with a language, and ask the provider which speech to text and voice models it uses for that language.

How does handover to a person work?

There are three ways, and you may want more than one.

Set the triggers before you start: the caller asks for a person, sounds upset, mentions an emergency word you chose, or the AI cannot answer twice in a row. When it hands over, it should say so plainly ("I will get a team member to call you") and send you the reason. If it cannot reach anyone, it should still take the details.

When we set up an agent, we start in draft mode: it prepares the summary or the booking and you approve it before anything is confirmed to the caller. Bookings and anything about money wait for the owner’s yes. That slows it down on day one and shows you what it gets wrong before a customer does. Our Never Miss a Lead agent is open to a waitlist and is not on sale yet (see pricing). Read about draft mode if you want the detail.

What should I check before buying one?

1. Does it say it is an AI? Ask to hear the opening line.
2. Which languages, and which models? Ask for a recording of a real call in each language you need. Then call it yourself with an accent and a poor line.
3. What does it do when it does not understand? It should ask again, then hand over or take a message, not guess.
4. How does handover work? Live transfer, call back, or both, and what triggers it.
5. Is it connected to your diary? Which calendar, and can it avoid double booking?
6. What do you get after each call? Ask to see a sample summary.
7. Calls and recordings. Ask whether calls are recorded, who can listen, how long they are kept and where. Read recording calls, RICA and AI receptionists and POPIA and AI agents.
8. What does it cost, and can you stop? See what an AI receptionist costs in South Africa and check for a long contract.

Is it worth it compared with a person?

It depends on your call volume and the kind of calls you get. A person can handle the angry caller, the odd request and the judgement call. Software can answer when nobody is free and ask the same questions every time. Many owners will want a mix. We compare the costs in AI receptionist or a human receptionist and cover what to do in the meantime in what to do when you cannot answer the phone on a job.

Yielde builds AI receptionists to order for owner-run businesses, and each build starts from one result you can measure, such as the share of calls that end with details taken. It is scoped and priced on a call before work starts. Start with the AI receptionist page if you want to see how it would work for you.