AI Receptionist: what it should answer, what it must escalate, and what it actually runs on

An AI receptionist is software that answers your business phone, understands speech, and either resolves the call or routes it. It should handle hours, directions, availability, quote status and message taking. It must escalate anything involving money owed, a complaint, or a caller in distress. Underneath it, four separate services each bill on their own meter.

Most articles on this stop at that paragraph. The three things that actually decide whether one works in a South African business are the escalation rules you set before it answers a call, the four layers it runs on, and the sentence RICA requires it to say. This post covers all three, with the statute and the vendor rate cards cited rather than summarised.

What should an AI receptionist answer?

The safe scope is anything where the correct answer already exists in writing and does not change based on who is asking. If a competent new employee could answer it from a one page briefing on day one, an AI receptionist can answer it.

• Operating hours, public holiday hours, and the address. Static facts, and the highest volume call type in most service businesses.
• Whether you cover a suburb or a service. A yes or no against a list you control.
• Where a job or quote currently sits, if it can read that status from your system rather than guess it.
• Booking or moving an appointment into a calendar with real availability, not a callback promise.
• Taking a structured message: name, number, reason, urgency, captured as fields rather than as a voice note somebody still has to listen to.
• Qualifying a new enquiry and handing it to the person who can quote it. This is where it earns the most, and it is the same argument as speed to lead: the first business to respond usually wins the job, and after hours the AI is the only one responding.

What must it escalate, every time?

Escalation rules are not a nice to have you add later. They are the specification. Write them before you choose a vendor, because a system that cannot hand over cleanly is worse than a voicemail: it holds the caller for two minutes and then loses them.

Trigger	Why it must reach a person
Money owed, refunds, disputed invoices	A wrong figure said out loud is heard by the caller as a commitment. Never let a bot negotiate a balance.
Any complaint, or repeat contact about the same issue	The second call about one problem is the retention moment. Handing it to software reads as avoidance.
Distress, medical urgency, safety, threat of harm	No scripted system should be the only thing between a caller and help.
Legal, contractual or regulatory questions	Judgement calls with liability attached. Route them, do not answer them.
The caller asks for a human	Absolute. One request, immediate handover, no retry loop and no second qualifying question.
Two failed attempts to understand the caller	Accents, line quality and code switching are real here. Cap the retries and hand over.

The last two matter more than the vendor demos suggest. Test any system on your actual callers, on a mobile line, in the languages your customers actually use, before it answers a real call. The failure mode that kills these projects is rarely a wrong answer. It is a loop the caller cannot get out of, which is the same pattern behind most of the reasons AI projects fail.

What does an AI receptionist actually run on?

There is no single product doing this. Every AI receptionist is four services chained together, and each one bills separately. Knowing the layers is what lets you read a quote and see where the markup sits.

1. Telephony. A real number, and the minutes carried over it.
2. Speech to text. Turning the caller audio into words.
3. The model. Deciding what to say and what to do.
4. Text to speech. Turning the reply back into a voice, fast enough that the pause does not feel wrong.

Published rates for those layers, taken from the vendors own pricing pages and checked on 3 September 2026:

Layer	Published rate	Source
Inbound minutes to a South African local number	USD 0.0107 per minute	Twilio voice pricing, South Africa
South African local number rental	USD 1.50 per month	Twilio voice pricing, South Africa
Transcription	USD 0.003 per minute (gpt-4o-mini-transcribe), USD 0.006 per minute (Whisper)	OpenAI API pricing
Speech to speech model, audio in and out	USD 10.00 and USD 20.00 per million tokens (gpt-realtime-2.1-mini)	OpenAI API pricing
Text to speech	USD 15.00 per million characters (tts-1)	OpenAI API pricing
Bundled voice agent minutes	USD 0.08 per minute, USD 0.16 for additional minutes	ElevenLabs API pricing

Two things fall out of that table. First, the telephony layer is close to free and the intelligence layer is not, which is the reverse of how phone systems have been priced for thirty years. Second, a bundled per minute price and a raw component price are not comparable until you know your average call length and your monthly call volume, so ask any vendor for both. That is the same arithmetic problem as costing a WhatsApp chatbot, and it is why the honest answer to what AI automation costs always starts with a question about volume.

What does South African law require before it answers?

This is the part the international guides leave out, and it is not optional. An AI receptionist records and processes the call by design. That puts it inside the Regulation of Interception of Communications and Provision of Communication-related Information Act 70 of 2002, which South Africans know as RICA.

Three sections decide the shape of your setup:

• Section 4(1) allows any person, other than a law enforcement officer, to intercept a communication if he or she is a party to the communication, unless it is intercepted for the purpose of committing an offence. You are a party to a call made to your own business.
• Section 5(1) allows interception where one of the parties has given prior consent in writing. Note the words: prior, and in writing. A verbal yes given during the call is not what that section describes.
• Section 6 is the business exception, and it is the one that actually governs a receptionist. It permits interception of an indirect communication in the course of carrying on a business, subject to conditions, including at section 6(2)(d) that the system controller has made all reasonable efforts to inform in advance a person who intends to use the system that communications may be intercepted.

Read section 6(2)(d) as a product requirement rather than a legal footnote. Informing the caller in advance means the notice arrives before the substance of the call, not in a policy document nobody opens. In practice that is one sentence, spoken first, every time, saying who they have reached and that the call is recorded. Build it into the opening turn and it is never missed. Bolt it on afterwards and it will be.

The personal information the call generates then falls under POPIA in the ordinary way: a lawful basis, a purpose, a retention period, and a security safeguard over the recordings and the transcripts. Our POPIA compliance checklist covers those obligations. The separate question of what a machine may decide on its own, rather than merely record, is dealt with in automating customer service, which reads section 71 and the Electronic Communications and Transactions Act on where a human has to stay in the loop.

Where the actual saving comes from

The saving is not the answered call. It is everything that stops being re-typed after it.

A call answered by a person usually produces a note somewhere: a sticky, a WhatsApp message to a colleague, a line in a diary. That note gets re-keyed into a quote, then into an invoice, then into a job card. The same five facts, typed four times, by people who each get a chance to typo them.

An AI receptionist captures those facts once, as fields, at the moment they are said. The mechanism is worth specifying explicitly when you brief a build:

• Capture once. Name, number, service, suburb, urgency and preferred time land as structured fields on the call record, not as a recording somebody has to listen back to.
• Reuse everywhere. That record creates the CRM contact, pre-fills the quote, and carries into the job card and the invoice without a second keystroke.
• Stop re-keying. The number the customer said is the number on the invoice, because nobody retyped it in between.
• Fire the reminder from the record, not from a person. Quote sent and not accepted after three days, follow up. Job booked for tomorrow, confirm today. The trigger is the field, so it does not depend on anyone remembering.
• Escalate on the same record. When the call goes to a person, they open a populated screen instead of asking the caller to start again. That is the difference between a handover and a transfer.

This is the pattern behind every useful agent build, and it is why the scoping question is which records already exist rather than which model to use. The wider version runs in AI agents for business and WhatsApp business AI, where the same capture once principle runs over chat instead of voice.

Questions to ask before you sign anything

Take these to any vendor, in this order. Our fuller AI vendor checklist covers contracting and exit; these are the voice specific ones.

1. Where are the recordings and transcripts stored, in which country, and for how long?
2. Is the RICA notification in the first turn of the call, and can I hear a recording of it?
3. What exactly happens when the caller asks for a person, and how long does that take?
4. Am I billed per minute, per component, or both, and what does a call of my average length cost?
5. What happens on a bad line or a strong accent, and how many retries before handover?
6. If I leave, do I get the call records and the structured fields, and in what format?

Frequently Asked Questions

Is it legal to use an AI receptionist in South Africa?

Yes, subject to RICA. Section 6 permits business interception of calls, but section 6(2)(d) requires all reasonable efforts to inform the caller in advance that communications may be intercepted. The recordings and transcripts then carry POPIA obligations for purpose, retention and security.

Does an AI receptionist have to tell callers it is not a person?

RICA governs the recording notice rather than disclosure of the agent nature. Disclosing it is still the right default: callers work it out quickly, and being told upfront reduces the frustration that drives the escalation requests you would rather avoid.

What does an AI receptionist cost to run?

It depends on call volume and average call length, because the layers meter differently. Published component rates run from USD 0.0107 per minute for South African inbound telephony to USD 0.08 per minute for a bundled voice agent, before any platform or support layer.

Can it handle South African accents and multiple languages?

Performance varies by provider and by language, and vendor demos are recorded on clean lines. Test on your own callers, on mobile, in the languages they actually use, and set a hard cap of two failed attempts before it hands over to a person.

Should a small business start with voice or with WhatsApp?

Usually WhatsApp, because the text channel is cheaper to run, easier to correct, and already where South African customers are. Voice earns its place when missed calls are demonstrably costing you work, particularly after hours.

Where to start

Do not start with the model. Start by counting the calls you miss, the calls you answer that nobody needed to answer, and the facts your team types more than once. Those three numbers decide whether a voice agent is worth building at all, and they decide what it should be allowed to say.

That is the work our vendor checklist assumes you have already done, and it is exactly what a Business Autopsy produces: the real call and admin picture in your business, before anyone sells you a system. If you want that mapped properly, book a Business Autopsy and we will start with your numbers rather than a demo.