AI voice agents are software that answers and makes phone calls in natural conversation. An AI voice agent listens, works out what the caller wants, answers from your information, takes actions such as booking an appointment or updating a CRM, and hands over to a person when it should. Businesses use them as receptionists, support lines and sales callers.


“ShiftMate is a South African voice agent specialist, based in Umhlanga. We build AI voice agents that sell, answer calls and train the people who work alongside them, and we ran hundreds of test calls to get them to sound and reply like a South African on a South African line. This guide is everything we've learned.”
AI voice agents are software that holds a real phone conversation. A caller speaks the way they would to a person ("Hi, I need to move my Thursday booking, something's come up") and the agent understands, answers, does the job in your systems and confirms it. If the call needs a person, it hands over and passes on what was already said.
That makes an AI voice agent different from the phone technology most people know. An IVR menu ("press 1 for accounts") only understands key presses or a few fixed words. An old-style voice bot follows a script and gets lost as soon as the caller goes off it. A chatbot works in text on a website or WhatsApp. An AI voice agent understands everyday speech, follow-up questions, changes of mind and interruptions, and it can act: book, look up, update, send, transfer.
You'll see the same thing called an AI phone agent, an AI caller, an AI receptionist, conversational voice AI, a voice bot or AI phone call automation. Suppliers that sell the tools to build one often call themselves a conversational AI platform; large contact-centre suppliers talk about enterprise voice AI. They all describe software that talks on the phone. For a gentler start, read what is a voice bot and what is AI calling.
Inbound agents answer calls: the receptionist, the support line, the booking line, the after-hours line. The caller chose to phone, so the bar is simple: answer quickly, get it right, hand over well. Outbound agents make calls: following up a lead who filled in a form, confirming tomorrow's appointment, running a short survey, reminding someone about a payment, or making a sales call. Outbound calls carry more legal weight, because the person didn't ask to be called right now. Many businesses start with one direction and add the other once it works.
| IVR menu (press 1) | Human agent | AI voice agent | |
|---|---|---|---|
| Available | 24/7 | Working hours, unless you pay for shifts | 24/7 |
| Calls at once | Many | One | Many, up to the limits you set |
| Understands everyday speech | No, key presses or set words | Yes | Yes, if tested on your callers' accents |
| Finishes routine tasks | Rarely | Yes | Yes, through connections to your systems |
| Judgement, empathy, unusual cases | No | Yes | Hands over to a person |
| Consistency | Always the same | Varies with the person and the day | The same every call, including its mistakes |
| Usual price model | Low, mostly fixed | A salary and overheads | Per minute, per month or per result |
The last-but-one row matters more than people expect. A person who gets something wrong usually gets it wrong once. An AI voice agent with a gap in its information gets it wrong on every call until someone fixes it. That is why testing and call review sit at the centre of this guide.
Good evening, Ridgeside Physio. I'm the practice's AI assistant, and this call is recorded. How can I help?
It says it's an AI, and that the call is recorded, in its first line.
Yeah, no, I've got a booking on Thursday with Pieter but I need to move it, my son's got a match.
No problem, Anna. I can see Thursday at 10:00 with Pieter. He has Friday at 08:30 or Monday at 14:15. Which suits you?
"Yeah, no" isn't a refusal here. It listens to what comes next.
Monday, please. And is it still R650?
Let me move that first. Done: Monday at 14:15 with Pieter, and the booking system has confirmed it. On the price, I don't have your medical aid details on this call, so I'll ask reception to WhatsApp you the exact amount tomorrow morning. Is this number fine?
It only says "done" once the system confirms it, and it doesn't guess a price it can't see.
Ja, that's fine. Thanks.
AI voice agents work in a loop that repeats on every turn of the call: hear the caller, understand them, decide what to do, do it, and reply out loud, fast enough that the pause feels natural. Here is the loop in plain English.
The call arrives on a normal phone number (or goes out from one) and the audio streams to the agent. In South Africa that line may run over cellular networks, fibre or a noisy shop floor, and the agent has to cope with all of them.
Speech recognition turns the caller's words into text as they speak, or a speech-to-speech model listens to the audio directly. It also has to decide when the caller has finished, which is harder than it sounds: people pause mid-sentence.
A large language model works out what the caller wants, using your instructions, your information and the call so far. Good set-ups also check each planned reply against your rules before it's spoken.
Through connections to your systems it books, looks up an order, updates the CRM, sends an SMS or WhatsApp, takes a payment through a secure flow, or transfers the call.
Text-to-speech turns the answer into a natural voice (or the speech-to-speech model speaks directly). If the caller interrupts, it stops talking and listens, the way a person would.
When the call needs judgement, empathy or a decision it isn't allowed to make, it hands over to a person, with a summary, so the caller doesn't have to start again.
A recording, a transcript, the outcome and the reason for any handover, so your team can review calls and improve the agent every week.
An agent that can only talk is a fancy answering machine. The value comes from what it can reach. The usual connections are a calendar or booking system (to see free slots and book them), a CRM (to recognise a caller, read their history and write up the call), an order, policy or account system (to give a real status), SMS or WhatsApp (to send a confirmation the caller can keep), a payment provider (through a secure flow, so card numbers never pass through the conversation), and the phone system itself (to transfer to the right person or queue). Each connection should be the narrowest one that does the job: an agent that only needs to read appointment slots shouldn't be able to delete patient records.
Ask every supplier which of your systems they already connect to, what it costs to connect the ones they don't, and what the agent does when a system is slow or down. A good answer is that it tells the caller honestly and takes a message; a bad answer is silence.
Most AI voice agents use one of two designs. The chained design passes the call through three parts in turn: speech recognition, a language model and a voice. You can choose and tune each part, see exactly what the agent heard, and check its words before they're spoken, but every hand-off adds delay. The speech-to-speech design uses one model that hears audio and speaks audio, which can sound more natural and reply faster, but gives you less control over each step and can cost more per minute. Neither is always better. Our guide on how to build an AI voice agent compares them in more depth.
Human conversation runs on very short gaps. Research across ten languages found that people answer each other within a fraction of a second, with a similar pattern in every language studied (Stivers and colleagues, PNAS). When an AI voice agent leaves a long silence before it replies, callers assume it didn't hear them. They repeat themselves, talk over it, or hang up. Suppliers advertise reply times, but those are their own claims, measured in their own conditions.
“We ran hundreds of test calls to get the delay right. A voice agent can say all the right things and still fail if the pause before each reply is too long: callers talk over it, repeat themselves or hang up. Speed isn't a setting you switch on; it's something you measure on every version, from where your callers really are.”
Two practical points follow for anyone buying. First, ask for the delay measured on calls that start where your callers are. A server in the United States or Europe adds a round trip on every turn for a caller in Durban or Polokwane. Second, ask exactly what the speed number measures.
“We learned to measure the real answer, not the filler. A quick "Sure" makes a speed number look good while the caller is still waiting for what they asked. And we measure the round trip from a South African caller, not from a server next door to the AI.”
Speech recognition that scores well on American test recordings can struggle on a South African line. A study of African-accented English found that off-the-shelf speech recognition makes several times more errors on accented speech, including South African accents, than on standard US benchmarks (AfriSpeech-200, 2023). South African callers also switch languages mid-sentence, use local words ("ja", "just now", "now now") and give answers whose meaning depends on how they start.
“"Yeah, no" is a South African way of starting an answer. It isn't a no. An agent built on American speech hears a refusal; ours has to listen to what comes after it, and if nothing clear follows, ask.”
The voice matters as much as the hearing. A Cape Town voice AI company told iAfrica that its data shows drop-off is "significantly higher when customers encounter robotic or foreign accents" (iAfrica, June 2026). The same article notes that nine of South Africa's eleven official written languages count as low-resource for AI, so support for isiZulu, isiXhosa and others is still uneven. Ask any supplier which languages work today, not which are on the roadmap, and test them on recordings of your own callers. Our guide to AI voice agents in South Africa covers languages in detail.
AI voice agents do best on calls that are frequent, follow a pattern and have a clear finish: a booking made, a question answered, a lead qualified, a payment arranged. They do worst on calls that need judgement, negotiation beyond set limits or real empathy. Here is how that plays out by role, then by industry, with a guide for each.
Each role below is a job an AI voice agent can take on in any business, with the guide that covers it in full.
Answers every call, gives hours, prices and directions from your information, books and moves appointments, takes proper messages and puts urgent callers through. It suits businesses that miss calls when they're busy or closed. Read the AI receptionist guide.
Handles order status, account questions, password resets and simple fixes around the clock, and hands harder calls to people with a summary. It's the modern replacement for a phone menu. Read about the AI customer support phone agent.
Books, confirms, reminds and reschedules by phone, which cuts no-shows and fills cancelled slots. See the AI appointment setter guide.
Calls agreed leads on an approved script, handles common objections and books the next step with a person. It carries the strictest consent rules. Read about AI cold calling software.
Phones a new enquiry within minutes, asks the qualifying questions and books the serious ones with a salesperson. Speed to lead is the main win. See AI voice agents for lead generation.
Runs short satisfaction and NPS calls, asks a follow-up question when an answer is vague, and flags unhappy customers for a call back. See AI voice agents for surveys.
Each industry adds its own rules and risks on top of the role, so the safe line sits in a different place for each one.
Bookings, repeat-script requests, directions and after-hours messages, with a clear safety line: the agent never diagnoses and sends emergencies straight to emergency services. See the AI medical receptionist guide.
Recall calls, confirmations, cancellations and filling the gaps they leave in the diary. See dental office AI phone answering.
Answers listing enquiries at night and on weekends, checks whether a buyer is pre-approved and books viewings. See the AI voice agent for real estate.
Takes phone orders and table bookings during the rush, reads back the order and sends it to the kitchen system. See AI voice ordering for restaurants.
Payment reminders and simple payment arrangements within the debt collection rules, with hardship and disputes going to people. See AI debt collection.
Policy questions, claim status and first notice of loss, with advice kept for licensed people under FAIS. See AI voice agents for insurance.
Call centres and BPOs sit across all of these. They use AI voice agents to take the routine share of calls, and increasingly use the same technology the other way round: an AI customer that new agents practise on before they take live calls. More on that in AI practice calls.
| Type of call | Good fit for AI? | Why |
|---|---|---|
| Booking, moving or confirming an appointment | Yes | Clear steps, clear finish, easy to check in your system. |
| Opening hours, prices, directions, what to bring | Yes | Answers come straight from your information. |
| Order or claim status | Yes, if connected | Needs a live link to the system that holds the status. |
| Following up a new lead within minutes | Yes, with consent | Speed matters most; the rules on consent apply. |
| Payment reminder or simple arrangement | With care | Strict collection rules; hardship goes to a person. |
| Complaint from an upset customer | Start only | It can listen and log the details; a person should resolve it. |
| Medical symptoms, legal or financial advice | No | Needs a qualified person; the agent should hand over. |
| Bereavement, abuse, vulnerable callers | No | Hand over at once, with care. |
One warning from our own work: a rule that is right for one kind of call can be wrong for another. Treat each use as its own project with its own tests, even if the same agent handles both.
“A rule we wrote for outbound calls broke inbound ones: it flagged a perfectly normal caller saying "Hi, I'm phoning about my fibre" as a mistake. The same sentence is right on one kind of call and wrong on another, so we test every call type on its own.”
AI voice agents are usually priced per minute of calls, per month, per result, or a mix, plus a set-up fee. Published per-minute prices depend on what they include: one developer platform lists US$0.07 to US$0.31 a minute, built from itemised parts (voice engine, voice, AI model and a US$0.015 phone line), while hosting-only building blocks start at about US$0.01 a minute and bill the AI models and phone line separately (vendors' own pricing pages, checked 16 November 2026; see AI voice agent pricing). Choosing a premium speech-to-speech model raises the per-minute price further. Large enterprise voice AI suppliers don't publish prices; some sell only annual contracts and one charges by outcome. The named platforms and their sources are on our pricing page.
In South Africa: Listed prices for South African AI voice agent services run from R849 a month plus calls to R7,500 a month, with several bundles around R1,900 to R2,400 a month including a set number of minutes and then about R0.97 to R13 a minute; some are ex VAT (WhichVoIP, a third-party comparison site, 31 October 2026, not confirmed on each supplier's own page).
A low per-minute rate often covers only one part of the call. Before you compare two quotes, ask which of these each one includes:
Most global platforms bill in US dollars, so a South African business carries the exchange rate on every minute. A local supplier that quotes in rand takes that risk off you, but check whether its rand price is fixed or moves with the dollar. Either way, ask for the all-in monthly figure at your real call volume, not a per-minute headline.
Take a normal month: how many calls of the type you want to automate, and how long they last on average.
Multiply minutes by the all-in rate (line, hearing, model, voice, platform), and add the monthly fees.
Spread the set-up fee over a year, and add the staff time to review calls and keep the information current.
From your pilot, how many calls ended with the job done (a booking made, a question answered) without a person? Not how many were "contained": a caller who gave up isn't a result.
Total monthly cost divided by finished results is your cost per result. Compare that with what the same result costs you today, including missed calls you currently lose.
Two outside numbers help frame this. Gartner predicts that by 2030 the cost per resolution for generative AI in customer service will exceed US$3, "higher than many B2C offshore human agents" (Gartner, January 2026). That's a forecast, not a measurement, but the point stands for South Africa, where people cost less than in the US: AI voice agents don't automatically save money here. The savings case is strongest where calls are missed, come at night, arrive in spikes, or tie up skilled people on routine questions.
Choose an AI voice agent by testing it on your own calls, not by reading feature lists. Shortlist by the kind of supplier that suits your team, run every candidate through the same test plan, then pick the one that performs on recordings of your callers at an all-in price you can live with. Names, prices and features of the main platforms are compared on our best AI voice agents page, and who's who in the industry is on voice AI companies.
| Kind of supplier | Who it suits | What you do yourself | Watch out for |
|---|---|---|---|
| Developer platform or framework | Teams with engineers whose product is voice | Almost everything: design, connections, testing, hosting choices | Low headline price, high build and running effort |
| No-code builder | Small businesses with simple, single-purpose calls | Write the instructions and load your information | Brittle once callers go off the expected path |
| Enterprise voice AI platform | Large contact centres with existing systems | Work with the supplier's team over months | Long contracts, custom pricing, slow changes |
| Done-for-you specialist | Businesses that want the result without an engineering project | Supply your information and rules, approve tests | Check who owns the agent, the data and the call recordings |
Any supplier worth considering will let you test before a long contract. Use the same plan for each one, so you compare like with like. Write down a pass mark before you start, not after.
| Test | How to run it | Pass when |
|---|---|---|
| Your accents | Play or role-play 20 calls using real wording from your recordings, with your callers' accents and language switching | It understands nearly all of them and asks again, politely, when it doesn't |
| Speed | Time the gap from the end of the caller's sentence to the start of the real answer, from a South African phone | No awkward silences; callers don't repeat themselves |
| Going off script | Change your mind halfway, ask two things at once, ask something it can't know | It keeps track and admits what it doesn't know |
| Interruptions (barge-in) | Talk over it mid-sentence | It stops and listens |
| The robot question | Ask "Am I talking to a real person?" | It says clearly that it's an AI |
| Opt-out | Say "Take me off your list" on a sales call | It stops selling, confirms, and the number is suppressed |
| Handover | Ask for a person, or get upset | A person picks up with a summary; the caller doesn't start again |
| Actions | Book, move and cancel in your real system | Your system shows exactly what the agent said it did |
| Repeat runs | Run every test again after any change | It passes each one more than once, not just once |
“Before a version of one of our AI sales agents may phone anyone, it has to pass every practice situation three times: the busy prospect, the one who only wants an email, the one who asks if it's a robot, the one who says take me off your list. Passing once proves nothing.”
If the supplier scores calls for you with AI, ask how you can check the scores. A mark that quotes the caller's and agent's actual words can be checked by a person in seconds; a bare number can't.
“When we first had AI mark every practice call against a scorecard, about one review in three came back cut off and unusable. We changed it so every mark has to quote the words from the call, checked in code. After that, the same stored calls reviewed three more times got the same marks: 18 of 18 on one call. A score nobody can check is worse than no score.”
Choose the single most common, most routine call. Write down what the agent may say and do, what it must never do, when it hands over, and what counts as a finished call. Gather 50 real recordings or transcripts of that call.
Run the test plan above, plus your own worst calls: the angry caller, the confused one, the one who changes topic three times. Fix, then run every test again. Don't move on until it passes repeatedly.
Send a small share of real calls to the agent, during working hours, with someone from your team reviewing every call the same day. Brief the people who take handovers so they know what the caller has already been through.
Compare finished calls, handovers, repeat calls and complaints with the same call type handled by people. Then decide: widen it, fix and retest, or stop. Stopping is a fair result; a pilot exists to find out.
Building your own AI voice agent from parts looks cheap on a pricing page. The hard work is everything around the model: getting the delay down on real phone lines, deciding when the caller has finished speaking, handling interruptions, connecting your systems safely, writing the rules for handover, and testing every change. Teams usually underestimate how long the last ten percent takes.
“One of our earlier AI interviewers was good, but slow to reply, and on its first test run it dropped a turn. That taught us that a clever answer delivered late loses to a simple answer delivered on time, so we rebuilt for speed.”
| Route | Choose it when | Time to a working pilot | Main risk |
|---|---|---|---|
| Build it yourself | Voice is your product and you have engineers to own it for years | The longest | The agent works in the lab and struggles on real lines |
| Configure a platform | You have someone technical and fairly standard calls | Medium | You become the tester, the trainer and the support desk |
| Have a specialist build and run it | You want the result, not the project | Often the shortest | Lock-in: agree in writing who owns what if you leave |
For the full picture, read how to build an AI voice agent and the honest answer to how hard it is to build a voice agent.
Yes, AI voice agents are legal in South Africa and in most countries, but the rules on consent, calling times, opt-outs and personal information apply to them at least as strictly as to people, and sometimes more so. Outbound sales calls carry the most rules; inbound calls that a customer chose to make carry the fewest. What follows is general information as we understood it on 16 November 2026, not legal advice. Take legal advice on your own campaign before you launch.
Two habits help under almost every rulebook: say it's an AI at the start, and treat "take me off your list" as final.
“Our agents say they're an AI and that the call is recorded. People forgive a machine that's honest with them; they don't forgive one that pretends.”
“"Take me off your list" ends the sale. We made carrying on after that request an automatic fail, whatever the rest of the call scored, because some mistakes can't be made up for by a good opening.”
For campaigns, keep a consent record for every number you dial (where and when the person agreed), check the opt-out registry before every campaign, block calls outside the legal hours in the dialler itself rather than trusting a schedule, and keep recordings for as long as you need them and no longer. Our guides to AI cold calling software and AI voice agents in South Africa go into the detail.
AI voice agents don't fix a broken process, an empty knowledge base or a team that nobody trained. Put one in front of those and it repeats the same problem to every caller, quickly and politely. McKinsey's study of failed voice agents puts it plainly: most fail because they aren't reliable enough for real customers, they are given jobs they can't do, and they're judged on how many calls they keep away from people rather than how many problems they solve (McKinsey, June 2026).
The most damaging failure isn't a wrong answer that the caller notices. It's a confident "I've booked that for you" when nothing was booked. The caller arrives for an appointment that doesn't exist, and blames the business, not the software.
“A spoken promise isn't a completed action. Our agents only tell a caller something is done once the system confirms it is done, not because the conversation went well.”
Many suppliers report "containment": the share of calls that never reached a person. It's an easy number to push up. Make the handover hard to find and containment rises, while customers give up and phone back angrier. Measure resolution instead: the job finished, no repeat call within a week, and satisfaction (CSAT) for AI-handled callers compared with callers handled by people. Our guide to why AI voice agents fail sets out a full measurement plan.
Once AI voice agents take the routine calls, the calls left for people are the difficult ones: the upset customer, the tangled account, the caller the AI couldn't help. Gartner predicts that by 2028 none of the Fortune 500 will have fully eliminated human customer service, and that by 2027 half of the organisations that planned to cut their service workforce because of AI will drop those plans (Gartner, September 2025). In South Africa, one large BPO told The National that AI adoption had pushed its average training period from three weeks to eight (The National, via BPESA, November 2026). Plan for that: the people taking handovers need practice on hard calls, and supervisors need to know how to review AI calls.
The same applies to AI used for training. An AI customer that gives in too easily teaches agents the wrong habits, so it has to be tested just as hard as an agent that talks to real customers.
“We tested whether our AI customers gave in too easily. In 21 practice collections calls, customers who could pay nothing agreed to pay anyway in 9 of 12 calls, as often with a good agent as a weak one. Practice against a pushover teaches the wrong thing, so we measure it before we trust it.”
The AI voice agents market is small today and forecast to grow fast, though research firms disagree on how fast because each defines the market differently. Grand View Research valued it at US$2.5 billion in 2025 and forecasts US$35.2 billion by 2033 (Grand View Research). Market.us forecasts US$47.5 billion by 2034, from US$2.4 billion in 2024 (Market.us). A 2026 report from Meticulous Research forecasts US$43.83 billion by 2036 and lists South Africa among the countries it covers (Meticulous Research). All three are forecasts, not measurements. The details and more sources are on our AI voice agents market size page.
Investors moved first. CB Insights counted US$2.1 billion raised by voice AI companies in 2024, up from US$264 million in 2023 (CB Insights). Buyers are following more carefully: in a Gartner survey of 187 customer service leaders in mid-2024, 44% were exploring a customer-facing voicebot, 11% were piloting one and 5% had one deployed (Gartner, December 2024). In other words, most businesses were still trying it out.
South Africa matters to this market because it runs a large, growing call centre industry serving its own consumers and the UK, US and Australia. BPESA reports that global business services headcount grew from 65,000 in 2019 to an estimated 150,000 in 2024 (BPESA, March 2025), and that the sector created 26,346 new jobs serving international markets in 2025 (BPESA). AI voice agents are already in use: one South African BPO says its AI collections agent deals with more than 100,000 people a day (The National, via BPESA, November 2026). South Africa's draft National AI Policy was gazetted on 10 April 2026 (Bureau for Economic Research). We found no reliable survey yet of how many South African contact centres use AI voice agents.
“Voice AI isn't replacing call centres. What it's doing is making voice available and affordable to businesses that could never afford a call centre, and supporting the call centres that exist: training their people, generating leads, and making the low-value calls their operators don't want to make. This space is changing and opening up.”
Speech-to-speech models and better turn-taking are cutting the pause before replies. Expect speed to become a basic requirement rather than a selling point.
Buyers in South Africa, India and elsewhere want agents that sound local and handle code-switching. Low-resource languages are where the most work remains.
Alongside per-minute prices, more enterprise suppliers charge per conversation or per outcome. That shifts the risk, but read what counts as an outcome.
Large contact centre and software companies are buying voice AI companies. Ask what happens to your agent if your supplier is bought.
Simulated callers, repeatable test suites and call review are becoming a standard part of running an AI voice agent, not an afterthought.
South Africa's opt-out registry (being phased in from September 2026 to May 2027), the FCC's AI voice ruling and new AI laws in Europe mean consent and disclosure will be checked more, not less. Gartner predicts AI-related rule changes will raise assisted service volume by 30% by 2028 (Gartner, January 2026).
Our voice AI trends page follows these in more detail, and voice AI companies maps who's who, including funding and deals.
ShiftMate is a South African voice AI company that builds AI outbound sales agents, inbound voice agents made for each business, and AI practice calls that train call centre and sales teams, all on one engine. We're based in Umhlanga, KwaZulu-Natal, and we build for South African callers first.
This guide covers the whole picture: what AI voice agents are, how they work, what they cost, why they fail, the rules and the market. Our own place in it is simple. The headline forecasts measure software sold; what matters to a business is what the calls achieve. So we build for results: leads qualified, meetings booked, calls answered, people trained, and we think buyers should judge every supplier, us included, on those numbers rather than on minutes.
What that looks like today: our AI outbound sales agent phones the leads you agree on and books the qualified ones with your team, quoted in rand for your volume rather than from a price list. Inbound, we build each voice agent around your callers, your systems and your rules, rather than handing you a generic receptionist, medical, dental or restaurant product. And AI practice calls in Sim Studio, with plans on the Sim Studio pricing page, let your people practise hard calls on AI customers and get a score and what to fix after every call. If you want a chat agent on your website rather than a phone agent, see AI Customer Service Chat.
“We're open about who we want to work with: call centres and sales teams that want their people trained, their leads worked and the low-value calls taken off their operators, and businesses that could never afford a call centre but still need their phones answered and their leads called.”
“The market responds better to natural, authentic South African voices. An American or British voice on a South African line sounds like a scam call. So our sales agents always sound South African, from the first word to the last, even on UK or US campaigns.”
“Everything we build runs on one engine: the same voice AI that sells also trains people and interviews candidates, so every lesson from one product makes the others better. That engine, built in South Africa and tested on South African calls, is how we intend to lead voice AI here, and what we intend to take to international markets.”
For a call centre or sales team, we think the right place to start is one job: practice calls for new agents, or an AI sales agent on the leads nobody gets to. For a smaller business, it's the phone itself: calls answered after hours and enquiries called back while they're still warm. Describe a normal week of calls to us, and we'll tell you honestly whether an AI voice agent will help, what it would cost in rand, and whether you'd be better off starting with practice calls for your team.
An AI voice agent is software that answers and makes phone calls in natural conversation. It understands everyday speech, answers from your information, takes actions such as booking an appointment, and hands over to a person when a call needs judgement. Unlike an IVR menu, which only understands key presses and routes calls, an AI voice agent resolves them.
An AI voice agent works in a loop on every turn of the call. Speech recognition hears the caller, a language model decides what to do using your rules and information, connections to your systems carry out the task, text-to-speech replies in a natural voice, and the agent hands over to a person, with a summary, when it should.
It depends on what the price includes. Hosting-only building blocks start at about US$0.01 a minute; one developer platform lists US$0.07 to US$0.31 a minute, built from itemised parts (voice engine, voice, AI model and a US$0.015 phone line) (checked 16 November 2026). Listed South African services run from R849 a month plus calls to R7,500 a month. Add set-up, testing and call review, and compare suppliers on cost per finished result rather than per minute.
Yes, within the rules. Automated sales calls need consent under POPIA section 69, the national opt-out registry is being phased in (marketers register from 15 September 2026, consumer blocks from May 2027), and CPA calling hours apply. The agent should say it's an AI. This is general information, not legal advice.
Some can, but not all, and you must test it. Research on African-accented English found off-the-shelf speech recognition makes several times more errors than on US benchmarks. Ask which languages work today, then test candidates on recordings of your own callers, including language switching and local phrases such as "yeah, no".
The most common mistakes are judging the agent on containment instead of problems solved, hiding the handover to a person, testing once instead of after every change, and letting the agent say a task is done before the system confirms it. McKinsey's 2026 study adds giving agents jobs they can't do and leaving nobody to own them after launch.
No, not fully. They take a growing share of routine calls, but people still handle judgement, empathy and complex cases. Gartner predicts that by 2028 none of the Fortune 500 will have fully eliminated human customer service. The calls left for people get harder, so training and practice matter more, not less.
Build it only if voice is your product and you have engineers to own it for years. Otherwise use a platform if you have technical staff and standard calls, or have a specialist build and run it if you want the result without the project. Whichever route, agree in writing who owns the agent, the data and the recordings.
Tell us about your calls. We'll show you where an AI voice agent fits, and what it would cost in rand.
ShiftMate's AI sales agent: calls, qualifies and books your leads.
Train agents on AI customers before live calls.
The main platforms compared, with sources and prices.
Accents, languages, POPIA, the opt-out registry and rand prices.