AI voice agents are real, they are improving quickly, and for most South African businesses they are still the wrong first purchase. Text handles the same enquiries at a fraction of the cost and complexity, and a voice agent that mishears an address or a policy number creates a worse experience than the queue it replaced. Voice earns its place when the caller genuinely cannot type — not because it sounds impressive.
What is an AI voice agent, exactly?
Three pieces stitched together: speech recognition that turns what the caller says into text, a language model that decides what to say back, and speech synthesis that says it. Each step adds latency and each step adds a place for the thing to go wrong. That is why voice is harder than chat, not because the underlying model is different — it is the same model with two lossy layers bolted to it.
Where does voice genuinely beat text?
- When the caller is driving, on a ladder, or holding something. Trades, logistics and roadside services get a lot of these.
- When the caller is older or not comfortable typing. This is a much larger group than most product teams assume.
- When the number is already the front door and changing customer behaviour would cost more than the automation saves.
- When calls arrive in a burst — a rates deadline, a load-shedding schedule change, a product recall — and the queue is the problem rather than the handling.
Where does it fall over?
| Situation | Text | Voice |
|---|---|---|
| Capturing an address or a reference number | Exact | Error-prone, and errors are expensive |
| South African place and surname pronunciation | Not an issue | A real one — Tshwane, Gqeberha, Mthatha, Nkosi |
| Code-switching mid-sentence | Handled well | Handled poorly |
| A caller on a bad line | Not applicable | Degrades badly |
| An angry caller | Escalate immediately | Escalate immediately — and faster |
| Cost per interaction | Low | Several times higher |
| Audit trail | The text is the record | Needs recording, storage and consent |
The pronunciation row is the South African specific that overseas write-ups never mention, and it is not cosmetic. A voice system that mangles the caller's own surname or suburb loses their confidence in the first ten seconds, and they ask for a person — which is the outcome the deployment was meant to avoid.
Does it actually replace a call centre?
It replaces the tier-one layer: hours, balances, order status, "have you received my document", "when is my appointment". That is a large share of volume and a small share of value. What it does not replace is the person who resolves a dispute, exercises discretion, or calms somebody down. A call centre that deflects those to a machine measures a better deflection rate and a worse business.
The honest model is the same as for chat: automation first for the routine, a person second, and a handover that is fast and obvious. Measuring deflection rewards trapping people. Measuring resolution rate and time-to-first-response rewards helping them.
What does POPIA require if you record calls?
Recording a call captures personal information and often special personal information, so the caller has to be told, the recording has to be kept securely and for a defined period, and it has to be deletable on request. An AI voice agent adds a second question: the audio is usually processed by a third-party service, which makes that service an operator under the Act, with the contractual obligations that follow. This is worth an hour with somebody who knows the Act before you switch anything on.
What should most South African businesses do instead?
Start with text, because it is cheaper, more accurate and easier to correct. A website chatbot at R299 a month answers the same tier-one questions with none of the transcription risk, and gives you a written record of exactly what was asked — which is the data you need before you can sensibly specify a voice deployment at all. If after three months your logs show a large share of enquiries coming from people who genuinely cannot type, that is the moment to look at voice.
The impressive part of a voice agent is that it sounds like a person. The expensive part is that it heard "Mthatha" as "Matata" and routed the job to the wrong branch.
What is the shortest version?
Voice is genuinely useful and genuinely harder. Automate text first, keep the escape hatch to a human obvious, and treat any vendor who does not want to talk about transcription accuracy on South African names and places as somebody who has not deployed here.
Where we do this
If you are reading this from Umhlanga and East London, those pages go through the local version of it: which languages the bot needs, what the busy hours look like, and which sectors ask what.