AI voice agents in South Africa: what to ask before you buy, and what changes in the contact centre

An AI voice agent answers or places phone calls and holds the conversation itself. It works out what the caller needs, acts on it during the call, and hands over to a person when it should.

Most agents chain three models over a phone line:

• speech-to-text turns the caller's words into text;
• a language model decides what to say or do, and reaches your systems, such as a calendar or CRM;
• text-to-speech turns the reply back into a voice.

Some newer systems skip the text step. Either way, the agent depends on its instructions, the knowledge it can read, and its route to a person.

Fast enough that the caller never wonders whether the line dropped. A study of question and answer pairs in ten languages found that the most common gap before an answer was between 0 and 200 milliseconds, and the average was 208 milliseconds.

A voice agent can be slower than that. Each stage in the chain adds delay, and so does the phone network. A browser demo hides most of that, so ask for latency measured this way:

• from the caller's side, on real phone calls, from the moment the caller stops speaking to the first sound of the reply;
• on calls from South African numbers to the line you will use;
• as the typical turn and the slowest turns, because a good average can hide occasional turns that pause long enough to sound like a dropped call;
• across the whole chain, never one stage such as the language model alone.

Check the language list of the speech service the agent uses, then test it on your own callers.

• Microsoft's Azure speech service lists transcription for South African English, Afrikaans and isiZulu, with South African voices for English and Afrikaans.
• Google's Speech-to-Text lists Afrikaans, Sepedi, Swati, Southern Sotho, Tsonga, Tswana, Venda, Xhosa and Zulu, though which recognition model covers each language, and in which cloud region, varies.

A listing does not tell you how well the agent hears your callers on a weak mobile signal, switching between English and isiZulu in one sentence, or reading out surnames and account numbers. We found no independent accuracy figures for South African accents, so measure it yourself:

1. Take recordings of your own calls, used with your callers' consent, across the languages and provinces you serve.
2. Run them through the agent's speech-to-text and have a person mark every error that would change the outcome, such as a wrong date or a misheard reference number.
3. Place live test calls with speakers from each province you serve.

Yes, if you meet RICA's conditions and POPIA's duty to inform. RICA lets any person intercept a communication if he or she is a party to it, unless the purpose is an offence. In our reading, when your agent answers on your behalf, your business is a party to the call. A separate section lets a business intercept calls on a system provided for that business, for listed purposes such as keeping a record to establish the existence of facts, with the system controller's consent, and only if users were told in advance or consent.

POPIA adds a duty to inform. You must take reasonably practicable steps to make the caller aware of what is collected, the purpose, and any plan to send it to another country. So the agent's opening line should say three things: that the caller is speaking to an automated agent, that the call is recorded, and why. The sections we read do not require the first; we recommend it.

Only to people who have consented or, within limits, to your existing customers. Section 69 of POPIA prohibits direct marketing by any form of electronic communication, including automatic calling machines, without that consent or customer relationship. The Act defines an automatic calling machine as one able to do automated calls without human intervention. In our reading, an AI agent placing calls with no person on the line fits that wording.

You may approach a person only once to ask for consent, and every marketing communication must identify the sender and give a way to ask for the calls to stop.

For calls made by people, the Regulator's guidance goes further. Its guidance note on direct marketing takes the view that telephone calling is electronic communication, because calls now travel as digital packets. The note describes itself as guidance that POPIA overrides and that is not legal advice, and we found no court ruling on the point. If you ask for consent by phone, the note says to read out the prescribed Form 4 and record the call; an automated message must carry the same information and record the answer. The business carries the onus of proving consent.

A second register now applies. The National Consumer Commission opened registration on its national opt-out registry to direct marketers on 15 September 2026, and a marketer that fails to register may contravene section 11 of the Consumer Protection Act. Lists are cleansed from December 2026 to April 2027, and consumers can register to block direct marketers from May 2027. An outbound agent's call list has to pass that cleansing before every campaign.

A voice agent usually runs on several companies: the platform, the speech service, the language model and the phone carrier. Each that processes personal information for you needs a written operator contract, and any that sends it abroad needs a lawful basis for the transfer. Take more care if the system identifies callers by their voice: the Act's definition of biometrics includes voice recognition.

Ask each where the audio is processed and stored, and whether it may train on your calls.

Any decision that changes a caller's position in a way that matters. POPIA restricts decisions with legal or substantial effect that rest solely on automated processing profiling the person; the contract exception that relies on safeguards requires a chance to make representations.

Let the agent collect information, answer questions and book. Route the call to a person when it would end in a declined application, a rejected claim or a cancelled account.

This is not legal advice. Checked against the text of POPIA, RICA and the Information Regulator's guidance note on direct marketing on 9 October 2026. Have an attorney check your own arrangement.

1. Handover: what triggers a transfer to a person, whether the transcript travels with the call, and what the caller hears when nobody is free.
2. Latency and accuracy, measured on your own calls as described above.
3. A transcript and a summary of every call, which you can export.
4. An opening line that says the agent is automated, the call is recorded, and why.
5. A written operator contract, a list of every sub-processor, and the basis for any transfer abroad.
6. Where audio is processed and stored, and how long each kind of record is kept.
7. A clause that stops every provider in the chain from training on your calls.
8. How many calls the agent can take at once, and what callers hear when that limit is reached or a provider is down.
9. What sits outside the quote: telephony, speech services, the language model, numbers, and minutes over any allowance.
10. Whether you can keep your number. ICASA says an approved port takes a maximum of seven working days, and only mobile and geographic numbers can be ported.

Pick one narrow call type with a clear outcome, such as booking appointments or answering after hours, and give the agent a share of real traffic. Agree these in writing before it starts:

• the outcome you are measuring, and what counts as a resolved call;
• the calls the agent must hand over, and a person ready to take them;
• who reads a sample of transcripts every day and logs each failure;
• the result that would make you stop, extend or widen the pilot.

• Mishearing: accents, switching languages, names, numbers and noisy lines.
• Silence: a pause of a few seconds sounds like a dropped call, and the caller hangs up or talks over the agent.
• Wrong actions: the agent reaches the wrong system or books the wrong slot, with full confidence.
• Flattering numbers: containment counts calls the agent kept away from a person, so a caller who gave up and hung up counts as a success. Read it beside repeat calls and complaints.
• No way out: callers who cannot reach a person. In a Gartner survey of 5,728 customers in December 2023, reported by a trade site, countries not stated, 64% said they would prefer companies not to use AI in customer service, and 53% would consider switching to a competitor if a company planned to add AI. The main concern was that AI makes it harder to reach a person.

What reaches your people gets harder: exceptions, upset callers and handovers. Plan for these changes:

• A new role: someone owns the agent's instructions and knowledge, reads transcripts, and updates the agent when products or policies change.
• Quality assurance can review every automated call from its transcript, instead of sampling a few.
• Workforce planning shifts to fewer simple calls, longer complex ones, and staff for handovers in every hour the agent runs.

IAMX sets up voice agents that answer, qualify and route calls, built around your own calls, systems and words, tested in a simulator before going live, and handing over to a person when they should. The AI voice agents and automation page sets out what is included.

• Is it legal to record calls handled by an AI voice agent in South Africa?

  Yes, when your business is a party to the call, or records it on its own business system for a purpose RICA lists, with users told in advance or consenting. POPIA separately requires reasonably practicable steps to tell callers what is collected and why, so say at the start of each call that it is recorded and why.

• Can an AI voice agent make cold calls in South Africa?

  Not direct marketing calls to people who have not consented, unless they are existing customers and the rules for customers are met. POPIA's consent rule names automatic calling machines, and direct marketers must also register on the National Consumer Commission's opt-out registry and cleanse their lists against it.

• Does an AI voice agent understand isiZulu, Afrikaans and South African English?

  The major speech services list them. A listing does not prove accuracy for your callers, so test on your own recordings and live calls first.

• Do customers mind talking to an AI agent?

  Some do. In the Gartner survey above (countries not stated, reported by a trade site), 64% of customers said they would prefer companies not to use AI in customer service, mostly because it makes a person harder to reach. An agent that says what it is and hands over quickly when asked is the safer design.

• Will AI voice agents replace call centre jobs?

  Some of the work, not all of it. The agent takes predictable calls; people still handle exceptions, upset callers and handovers, and someone owns the agent. How much volume moves depends on how predictable your calls are, which a pilot will show.

• How long should a voice agent pilot run?

  Long enough to cover your normal mix of calls on one narrow call type, with transcripts read every day and the measure of success agreed before it starts.

• Proceedings of the National Academy of Sciences (Stivers and others, 2009), copy held by Radboud University: A study of question and answer pairs in ten languages found the most common gap before an answer fell between 0 and 200 milliseconds, with a mean response offset of 208 milliseconds across the full dataset. Checked 9 October 2026.
• Microsoft: Microsoft's Azure speech service lists speech to text for South African English, Afrikaans and isiZulu, and text to speech voices for South African English and Afrikaans. Checked 9 October 2026.
• Google Cloud: Google Cloud Speech-to-Text lists South African locales for Afrikaans, Sepedi, Swati, Southern Sotho, Tsonga, Tswana, Venda, Xhosa and Zulu, with the recognition models and cloud regions varying by language. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 70 of 2002): RICA section 4(1): any person, other than a law enforcement officer, may intercept any communication if he or she is a party to it, unless it is intercepted for the purpose of committing an offence. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 70 of 2002): RICA section 6(2) lets a business intercept indirect communications in the course of business only with the express or implied consent of the system controller, for listed purposes such as keeping a record to establish the existence of facts, on a system provided for use in connection with that business, and only if the system controller has made all reasonable efforts to inform users in advance or the user consents. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 18(1): when personal information is collected, the responsible party must take reasonably practicable steps to make the data subject aware of the information collected, the purpose, and any intended transfer to a third country, among other things. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 69(1): processing personal information for direct marketing by any form of electronic communication, including automatic calling machines, is prohibited unless the data subject has consented or is a customer of the responsible party under section 69(3). Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 69(5) defines an automatic calling machine as a machine that is able to do automated calls without human intervention. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 69(2): a responsible party may approach a data subject whose consent is required, and who has not previously withheld it, only once to request that consent. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 69(4): any communication for direct marketing must contain the identity of the sender and an address or other contact details to which the recipient may send a request that such communications cease. Checked 9 October 2026.
• Information Regulator (South Africa): The Information Regulator's Guidance Note on Direct Marketing states its view that telephone calling is electronic communication, because telephone calls now predominantly use packet-switched VoIP. Checked 9 October 2026.
• Information Regulator (South Africa): The Guidance Note on Direct Marketing states that POPIA and the Regulations prevail over it in the event of any inconsistency, and that it does not constitute legal advice. Checked 9 October 2026.
• Information Regulator (South Africa): The Guidance Note says a responsible party that uses a telephone to obtain consent must read out the contents of Form 4 and record the call, and an automatic calling machine's message must contain all the information in Form 4 and record the data subject's response. Checked 9 October 2026.
• Information Regulator (South Africa): The Guidance Note says that under section 11(2)(a) of POPIA the responsible party bears the onus of proving that a data subject consented to direct marketing by unsolicited electronic communication. Checked 9 October 2026.
• National Consumer Commission: Registration of direct marketers on the national opt-out registry opened on 15 September 2026 and runs to December 2026; a direct marketer that fails to register may contravene section 11 of the Consumer Protection Act read with the 2026 Amendment Regulations. Checked 9 October 2026.
• National Consumer Commission: List cleansing against the national opt-out registry runs from December 2026 to April 2027, and direct marketers must cleanse their lists monthly before running any direct marketing campaign. Checked 9 October 2026.
• National Consumer Commission: Consumers can register on the national opt-out registry to block direct marketers from May 2027. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 21(1): a responsible party must, in terms of a written contract with the operator, ensure that the operator establishes and maintains the security measures in section 19. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 72(1): a responsible party may not transfer personal information to a third party in a foreign country unless one of the listed grounds applies, such as a recipient bound by law, binding corporate rules or agreement providing adequate protection, or the data subject's consent. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 1 defines biometrics as a technique of personal identification based on physical, physiological or behavioural characterisation, including voice recognition. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 71(1): a data subject may not be subject to a decision with legal consequences, or which affects them to a substantial degree, based solely on automated processing intended to profile them, subject to the exceptions in section 71(2). Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): The guide's legal sections were checked against the text of POPIA (sections 1, 18, 21, 69, 71 and 72), RICA (sections 4 and 6) and the Information Regulator's Guidance Note on Direct Marketing on 9 October 2026. Checked 9 October 2026.
• Independent Communications Authority of South Africa: ICASA says an approved number porting transaction takes a maximum of seven working days, and currently only mobile and geographic numbers can be ported. Checked 9 October 2026.
• Call Centre Helper, reporting a Gartner survey: A Gartner survey of 5,728 customers, conducted in December 2023, found 64% would prefer companies not to use AI in customer service and 53% would consider switching to a competitor if they discovered a company planned to implement AI; the top concern was greater difficulty reaching a person. The report does not say which countries respondents were in. Checked 9 October 2026.