There's no single best AI voice agent. Developer platforms such as Vapi and Retell AI suit teams that build; enterprise platforms such as PolyAI, Sierra and Parloa suit large contact centres; done-for-you specialists suit businesses that want the result. The best AI voice agents are the ones that pass a test on your own calls, at an all-in price.


“ShiftMate is a South African voice agent specialist, so we've looked at these platforms as builders, not just reviewers. This guide compares the main AI voice agent platforms with checked facts and prices, explains how to choose, and shows where a done-for-you South African option like ours fits.”
The best AI voice agents depend on who will build and run the agent: developer platforms (Vapi, Retell AI, Deepgram's Voice Agent API) suit teams with engineers; enterprise platforms (PolyAI, Sierra, Parloa, NiCE Cognigy) suit large contact centres; ElevenLabs Agents suits teams that care most about the voice; and done-for-you specialists suit businesses that want the result without a project. Below, ten of the best AI voice agent platforms are profiled the same way, followed by a test plan for your own calls and our guide to AI voice agents covers the basics.
Interest is real, and so is the noise. In a Gartner survey of 187 customer service leaders in 2024, 44% said they were exploring a customer-facing AI voicebot, 11% were piloting one and 5% had one live (Gartner).
We listed platforms that come up often in searches for the top AI voice agents and in widely read comparison articles. On 16 November 2026 we read each company's own website for what it does, who it's for, its published price and what that price includes. Funding comes from the company's announcement or a named news outlet.
The list is in no particular order. If a list publishes its own test results, ask who ran them and whether they partner with the platforms they rank. Prices change: check the linked page before you buy.
An AI voice agent platform lets software hold a spoken conversation: it hears the caller (speech recognition), decides what to say (an AI model with your information and rules), says it (a synthetic voice) and connects to the phone network and your systems. Some chain those parts; others use one speech-to-speech model. It isn't an old menu system (IVR), which routes callers by keypress, or a chatbot, which types. Lists of the top conversational AI platforms often mix in phone software that coaches human agents during calls; that's a different product, so it isn't here.
Most businesses start with one of four jobs: answering customer service calls around the clock and handing the hard ones to a person; outbound sales calls to leads who agreed to be contacted; booking and confirming appointments; and collections reminders within the conduct rules. Pick your first job before you pick a platform.
| Platform | Best for | Published price (checked 16 November 2026) | Price includes | Billed on top |
|---|---|---|---|---|
| Retell AI | Contact centres, no-code plus API | $0.07 to $0.31 a minute | Itemised: platform, voice, AI model, US phone line | Add-ons; numbers $2 a month |
| Vapi | Engineering teams | $0.05 a minute | Platform hosting | Speech recognition, AI model, voice, phone line |
| Bland AI | High-volume calling | $0.12 to $0.14 a minute | AI model, speech recognition, voice | Phone line |
| ElevenLabs Agents | Natural voices, many languages | Plans up to $990 a month, then $0.08 a minute | Speech recognition, voice | AI model, phone line |
| Synthflow | BPOs, white label | Annual enterprise contract | Phone lines, voices, transcription, integrations | None stated |
| PolyAI | Complex service calls | Custom | By contract | By contract |
| Sierra | Voice plus WhatsApp and chat | Outcome-based | By contract | By contract |
| Parloa | Large service teams | Custom | By contract | By contract |
| NiCE Cognigy | Existing contact centres | Custom | By contract | By contract |
| Deepgram Voice Agent API | One bill for speech parts | $0.075 a minute (Standard) | Speech recognition, AI model, voice | Phone line (not stated) |
Each of the best AI voice agents below is profiled in the same order. Strengths are the company's own claims unless we say otherwise. Where we say "we found none", the item wasn't on the pages we read, which doesn't prove it doesn't exist.
A newer South African voice AI company, based in Umhlanga, building for where the market is going: voice AI that is judged on what it produces, not on how many minutes it talks. What it offers: an AI outbound sales agent that calls, qualifies and books your leads (priced in rand), inbound voice agents built for each business, and AI practice calls that train and test call centre and sales teams. Why it's different: built for South African callers and rules (POPIA, the national opt-out registry), natural South African voices, AI sales agents that must pass difficult practice calls before they phone anyone, and one engine that sells, trains and interviews. Who it's for: 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. If you want to build agents yourself, the developer platforms below are built for that.
We list ourselves first because it's our page. It isn't a ranking: the platforms below follow in no particular order.
| Retell AI, checked 16 November 2026 | |
|---|---|
| What it is: | A voice agent platform for contact centres (Retell AI). |
| Who it's for: | CX staff build in a no-code studio; developers extend it with APIs. |
| Price: | $0.07 to $0.31 a minute pay-as-you-go, built from parts: $0.055 voice infrastructure, $0.015 to $0.10 for the voice, the AI model (for example $0.032 a minute) and $0.015 for a US phone line. Numbers $2 a month (Retell pricing). |
| Includes: | Every part, itemised. |
| Strengths it states: | About 600 ms to respond (a vendor claim), simulation testing, SOC 2, ISO 27001. |
| Watch-outs: | A speech-to-speech model costs $0.38 a minute on its own. We couldn't confirm which plans include a BAA (the signed agreement for US health data); ask. |
| Funding: | In November 2026 it said it had passed $100 million in annualised revenue on a single $4.6 million seed round (source). |
| Best for: | Contact centres that want a no-code start with room for developers. |
| Vapi, checked 16 November 2026 | |
|---|---|
| What it is: | A developer platform for building, testing and running voice agents. |
| Who it's for: | Developers who want to choose each part. |
| Price: | $0.05 a minute for hosting. Its own estimate for 1,000 minutes is $82 to $129 a month. Plans start at $0, with Core at $29 a month; $5 free credit (Vapi pricing). |
| Includes: | The platform only: speech recognition, the AI model and the voice pass through at cost; the carrier bills the phone line. |
| Strengths it states: | "Sub-500ms latency" (a vendor claim), testing and simulation, SOC 2, HIPAA and PCI (Vapi). |
| Watch-outs: | The $0.05 is only the platform part; the HIPAA add-on costs $2,000 a month. |
| Funding: | A $50 million Series B led by Peak XV (source). |
| Best for: | Engineering teams that want control. |
| Bland AI, checked 16 November 2026 | |
|---|---|
| What it is: | An enterprise phone-agent platform that runs its own hosted models (Bland). |
| Who it's for: | Developers, teams and enterprises. |
| Price: | $0.14 a minute on Start, or $0.12 a minute plus $299 a month on Build (Bland pricing). |
| Includes: | The AI model, speech recognition and the voice ("no token charges"); phone lines are billed separately. |
| Strengths it states: | A 200 ms response (a vendor claim), "40+ languages natively", SOC 2, HIPAA with a signed agreement, GDPR, private-cloud set-ups on Enterprise. |
| Watch-outs: | Start is capped at 100 calls a day and 10 at once; the HIPAA agreement, single sign-on and data residency are Enterprise only. |
| Funding: | A $50 million Series C in June 2026, taking it past $100 million in total (source). |
| Best for: | High-volume calling at one all-in rate. |
| ElevenLabs Agents, checked 16 November 2026 | |
|---|---|
| What it is: | The voice company's platform for voice and chat agents (ElevenLabs Agents). |
| Who it's for: | No-code builders ("no engineering required") as well as developers. |
| Price: | Free (15 minutes a month), then plans from $6 to $990 a month (Business includes 12,375 minutes), then $0.08 for each extra minute (ElevenLabs pricing). |
| Includes: | Speech recognition and the voice; the AI model is billed separately, and your own carrier bills the phone line. |
| Strengths it states: | "90+ languages", SOC 2 Type II, ISO 27001, PCI DSS Level 1, HIPAA and GDPR. |
| Watch-outs: | We found no response time in milliseconds on its agents page, only "ultra-low latency". Signed HIPAA agreements are on Enterprise. |
| Funding: | A $500 million Series D at an $11 billion valuation in February 2026 (source). |
| Best for: | Teams that care most about how the voice sounds, in many languages. |
| Synthflow, checked 16 November 2026 | |
|---|---|
| What it is: | A voice agent platform for contact centres, BPOs and multi-branch brands, with a visual flow builder and APIs (Synthflow). |
| Who it's for: | Mid-market and enterprise buyers. |
| Price: | No per-minute rate: an annual enterprise agreement, with no free plan or trial (Synthflow pricing). |
| Includes: | Phone lines, CRM integrations, transcription, voices and SMS. |
| Strengths it states: | "Under 0.5 s" response (a vendor claim); white-label workspaces per client; SOC 2 Type 2, ISO 27001, GDPR, PCI DSS; HIPAA self-attested. |
| Watch-outs: | You can't try it before talking to sales. |
| Funding: | A $20 million Series A led by Accel in June 2025 (source). |
| Best for: | BPOs running agents for several clients. |
| PolyAI, checked 16 November 2026 | |
|---|---|
| What it is: | Enterprise voice agents for customer service, built on PolyAI's own models in its "Dialog Studio" (PolyAI). |
| Who it's for: | Large companies with complex service calls. |
| Price: | Custom; we found no public price. |
| Includes: | Agreed per contract. |
| Strengths it states: | Its own models; SOC 2 Type 2, ISO 27001, PCI DSS and GDPR. |
| Watch-outs: | We found no free tier, and its homepage speed figures don't say clearly which applies to voice. |
| Funding: | An $86 million Series D in December 2025, taking its total over $200 million (source). |
| Best for: | Enterprises wanting a partner for complex service calls. |
| Sierra, checked 16 November 2026 | |
|---|---|
| What it is: | An enterprise customer-experience agent platform; one agent works across chat, SMS, WhatsApp, email and voice (Sierra). |
| Who it's for: | Large brands. |
| Price: | Outcome-based ("pay for a job well done"); no per-minute price on its site. |
| Includes: | Agreed per contract. |
| Strengths it states: | Builds agents from your procedures and call transcripts; certifications including SOC 2, ISO 27001, ISO 42001, HIPAA, GDPR and PCI DSS. |
| Watch-outs: | We found no free tier or response time on its pages. |
| Funding: | $950 million at a valuation above $15 billion in May 2026 (source). |
| Best for: | Big brands that want one agent on voice and WhatsApp, priced on results. |
| Parloa, checked 16 November 2026 | |
|---|---|
| What it is: | "The AI agent management platform for enterprise CX", to build, test and scale voice and chat agents (Parloa). |
| Who it's for: | Enterprise service teams. |
| Price: | Custom; we found no public price. |
| Includes: | Agreed per contract. |
| Strengths it states: | Testing built in; ISO 27001, SOC 2 Type II, HIPAA and PCI DSS; SAP endorsed-app certification. |
| Watch-outs: | We found no free tier, response time or language count on its pages. |
| Funding: | A $350 million Series D at a $3 billion valuation in January 2026 (source). |
| Best for: | Large service operations, especially ones already on SAP. |
| NiCE Cognigy, checked 16 November 2026 | |
|---|---|
| What it is: | AI agents for enterprise customer service across voice and chat, plus a copilot for human agents (Cognigy). |
| Who it's for: | Enterprise contact centres, with low-code design. |
| Price: | Custom; we found no public price. |
| Includes: | Agreed per contract. |
| Strengths it states: | Help for human agents as well as automation; ISO 27001, GDPR, PCI DSS and HIPAA badges. |
| Watch-outs: | We couldn't confirm a response time or language count on its pages. |
| Funding: | No longer independent: NiCE bought Cognigy for $955 million; the deal closed in September 2025 (source). |
| Best for: | Contact centres already on NiCE. |
| Deepgram Voice Agent API, checked 16 November 2026 | |
|---|---|
| What it is: | One API combining speech recognition, a managed AI model and a voice (Deepgram). |
| Who it's for: | Developers. |
| Price: | $0.075 a minute (Standard), $0.050 to $0.065 if you bring your own model or voice, $0.163 on Advanced; $200 free credit (Deepgram pricing). |
| Includes: | On Standard: speech recognition, the model and the voice. Phone lines aren't mentioned on the pricing page; ask. |
| Strengths it states: | "60+ languages", delay reported per turn, end-of-turn detection in about 260 ms (a vendor claim, not the full reply). |
| Watch-outs: | You still build the phone side and business logic. |
| Funding: | A $130 million Series C in January 2026 (source). In October 2026 it bought Daily, maker of the open-source Pipecat (source). |
| Best for: | Developers who want one bill for the speech parts. |
Picking among the best AI voice agents depends on three things more than any feature list: who will build and maintain the agent, how many minutes you'll run, and where your callers are. Cut the list to two or three, then test those.
| If you... | Look first at | Why | Skip it if |
|---|---|---|---|
| Have developers and want to choose every part | Vapi, Deepgram Voice Agent API, or open-source frameworks | You pick the AI model, voice and phone carrier yourself | Nobody on your team will own it after launch |
| Want a no-code start with room to grow | Retell AI, ElevenLabs Agents | Business staff can build; developers can extend | You need a supplier to design the calls for you |
| Run a lot of outbound or high-volume calls | Bland AI, Retell AI | Bland: one all-in rate per minute; Retell: itemised parts you choose | You haven't sorted consent and the opt-out registry |
| Run a large contact centre with procurement | PolyAI, Parloa, NiCE Cognigy, Sierra | Designed with you, with enterprise security | You need a published price to start |
| Want results in South Africa without a project | A done-for-you specialist such as ShiftMate | Local callers, rules and rand pricing | You want a self-serve developer platform |
| Need your people trained as well | Tools that do both, such as ShiftMate's AI practice calls | The agent takes easy calls; people get the hard ones | You only want an agent |
Teams comparing Vapi alternatives usually look at four routes. Retell AI gives a no-code studio as well as an API, with every cost itemised. Bland AI bundles the AI model and voice into one rate and runs its own models. Deepgram's Voice Agent API puts the speech parts behind one API. Open-source LiveKit Agents and Pipecat give the most control for the most work. If nobody has time to run a platform, a done-for-you supplier may be the better alternative.
Both sell per minute to teams making many calls, but they're built differently. Bland AI runs its own hosted models and charges $0.12 to $0.14 a minute with the model and voice included and phone lines extra (Bland pricing, checked 16 November 2026). Retell AI lets you pick the voice and AI model, itemises each part from $0.07 to $0.31 a minute, and offers a no-code studio (Retell pricing, checked 16 November 2026). Bland claims a 200 ms response, Retell about 600 ms: both vendor claims. Run the same test calls on both, from South Africa.
Building gives you control and a lot of work. Read how to build an AI voice agent and how hard it really is before you commit.
The best AI voice agents cost more than their headline rate: the all-in cost is the platform fee plus the AI model, speech recognition, voice, phone line, numbers and any compliance add-on, plus the people who build, test and watch the agent. Standard pay-as-you-go agent platform prices run from $0.05 to $0.31 a minute, hosting-only building blocks start at $0.01, and speech-to-speech models can cost more (Retell lists one at $0.38). They include different things, so the cheapest number is often not the cheapest agent.
Most global platforms bill in US dollars. For a rand figure, add up the all-in dollar cost per minute, multiply by your expected minutes, and convert at the rate on the day you pay; then allow for the rand moving and your bank's charge on foreign payments. We don't quote an exchange rate because it would be out of date by the time you read this.
| Cost part | What it is | Who charges it | Example from a published price |
|---|---|---|---|
| Platform | Running and hosting the agent | The platform | Vapi $0.05 a minute; LiveKit Cloud $0.01 a minute |
| AI model | Deciding what to say | The platform or the AI lab | Retell lists models at $0.032 and $0.064 a minute |
| Speech recognition | Turning the caller's words into text | The platform or a speech company | Included by Bland AI, ElevenLabs and Deepgram (Standard) |
| Voice | Turning the reply into speech | The platform or a voice company | Retell lists voices at $0.015 to $0.10 a minute |
| Phone line | Connecting to the phone network | The carrier | Twilio US: $0.0140 a minute outbound, $0.0085 inbound |
| Compliance add-ons | Signed health-data agreements, single sign-on, data location | The platform | Vapi's HIPAA add-on is $2,000 a month |
| People | Building, testing, reviewing calls, fixing | You, or a supplier | Not on any pricing page; often the biggest cost |
Two worked examples from published prices: Vapi's own estimate for 1,000 minutes is $82 to $129 a month (Vapi pricing, checked 16 November 2026), and LiveKit's own example of a full phone call, including a US phone line, comes to $0.0479 a minute (LiveKit pricing, checked 16 November 2026). Both are vendor estimates; your AI model, voice and South African phone route will change them.
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). Test any of them on your own calls before you sign; see our AI receptionist guide.
Many lists of the best AI voice agents compare AI against a US agent's cost per call. In South Africa the comparison is a local call centre seat, which changes the savings sum completely. And AI isn't automatically cheaper at scale: Gartner predicts that by 2030 the cost per resolution for generative AI in customer service will exceed $3, higher than many offshore human agents (Gartner). That's a forecast, but a good reason to price your own calls. If a supplier prices by outcome, agree in writing what counts as a result before you sign. For every component in detail, see AI voice agent pricing per minute.
Test your shortlist of the best AI voice agents on the same real call situations, run each several times, score every call against one written checklist, and measure from where your callers are. A demo shows a platform's best call; a test plan shows its worst.
The easy ones, the angry customer, and callers who give an ID number or rand amount, ask for a person, ask if it's a robot, or say take me off your list.
Write down what a good call does (books the slot, hands over with the details, says it's an AI, stops at an opt-out) and what is an automatic fail.
Same knowledge, script and handover rules. Note how much help each needed: it's part of the cost.
Have staff with your customers' accents play the callers, including people who switch between English and isiZulu or Sesotho mid-sentence. Use real names and amounts.
AI agents don't give the same answer twice. Count passes on all three runs, not the best run.
Time the gap to the agent's real answer (not a filler word), on a South African line. Note the slowest replies.
Test barge-in by talking over it, change the subject, go silent, ask something outside its knowledge. Check it stops speaking when interrupted, admits what it doesn't know and hands over cleanly.
Run your busiest hour's calls at the same time. Check delay and errors don't rise, and that the limit on calls at once covers you.
After each call, check your CRM: right contact updated, booking made, outcome written back? A good call that leaves no record is a failed call.
“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.”
When you compare platforms, the number that matters is how often each passes every run, not whether it can pass at all. Retest after every change to the script, knowledge or model: a fix in one place can break another.
| What to score | How to check it | Automatic fail? |
|---|---|---|
| Says it's an AI and that the call is recorded | Listen to the first 15 seconds | Yes, if it hides it |
| Stops after an opt-out request | Say "take me off your list" mid-call | Yes, if it carries on |
| Understands accents, names and amounts | Give names, ID numbers, rand amounts | If it books the wrong person or amount |
| Copes with interruptions | Talk over it twice | If it talks over the caller or loses track |
| Hands over well | Ask for a person | If the person has to start again |
| Only says what it knows | Ask something not in its information | If it makes something up |
| Only confirms what's done | Ask it to book, then check the system | If it says it's done when it isn't |
| Reply delay from South Africa | Time the real answer, slowest of each run | Set your own limit before testing |
| Sounds natural and local | Listen to the text-to-speech voice | If your testers can't follow it |
| CRM write-back | Check the contact, booking and notes after each call | If the record is missing or wrong |
| Holds up at peak volume | Run your busiest hour's calls at once | If delay or errors rise sharply |
| Containment and handover rate | Share resolved without a person; share handed over with details | Count containment only if the need was met |
After go-live, keep two numbers side by side: containment (calls the agent finished properly) and CSAT (how satisfied callers say they were), for AI calls and human calls. High containment with falling CSAT means callers are being blocked, not helped.
“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.”
If you score calls with AI, every mark should point to the words in the call. For what goes wrong in live calls, see why AI voice agents fail.
Not reliably out of the box. Speech recognition trained mostly on American and European speech makes far more mistakes on African-accented English, and speed figures are measured where the vendor chose. Test for both.
Research on the AfriSpeech-200 dataset found that off-the-shelf speech recognition makes several times more errors on African-accented English, including South African accents, than on standard US benchmarks (AfriSpeech-200, 2023). A language count ("40+", "60+", "90+") tells you what a platform can attempt, not how well it understands a caller from Chatsworth, Paarl or Soweto.
South African callers also switch languages mid-sentence. Untapped AI in Cape Town supports English and Afrikaans, with isiZulu and isiXhosa in development; its CEO says drop-off is "significantly higher" when customers hear robotic or foreign accents (iAfrica).
“"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 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.”
Ask each platform: which text-to-speech voices sound South African, can it follow callers who switch languages, can you add your own words (suburbs, surnames), and does it read figures back before acting?
| Platform | Speed it claims | What the claim measures |
|---|---|---|
| Bland AI | 200 ms response time | Vendor claim, conditions not stated |
| Vapi | Sub-500 ms | Vendor claim on its homepage |
| Synthflow | Under 0.5 s | Vendor claim, agent response time |
| Retell AI | About 600 ms | Vendor claim, response time |
| Twilio ConversationRelay | p50 491 ms, p95 713 ms | Vendor's internal benchmarks |
| Deepgram | About 260 ms | End-of-turn detection only, not the full reply |
| LiveKit | Under 250 ms | Media transport worldwide, not the agent's reply |
| ElevenLabs, Sierra, Parloa, NiCE Cognigy | We found no voice response figure on their pages (checked 16 November 2026) | Measure it yourself |
| PolyAI | Under 300 ms, and 280 to 600 ms | Unclear which applies to voice |
These are the companies' own figures, checked 16 November 2026, measuring different things. People answer each other very quickly, across many languages (PNAS study of turn-taking), so a long pause sounds broken. Where the AI runs on servers in the US or Europe, a call from Durban adds a round trip that a figure measured next to the server doesn't show. Untapped AI routes its calls through South African data centres (iAfrica).
“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.”
So measure the real answer and the slowest replies, from a South African line. And plan for what happens when power or the network drops, and whether a missed call gets an SMS or WhatsApp follow-up.
AI voice agents follow the same rules as a person. In South Africa, an agent making marketing calls needs the person's consent under POPIA (with a narrow exception for your own customers), must respect the national opt-out registry being phased in under the 2026 Consumer Protection Act regulations, and must call within set hours. This is general information as reviewed on 16 November 2026, not legal advice: take legal advice for your own calls.
| Rule | What it means for an AI voice agent |
|---|---|
| POPIA section 69 | Direct marketing by electronic communication, including automatic calling machines, needs consent. A narrow exception covers your own customers, for your own similar products, if offered an opt-out when you collected their details and on every contact since. |
| CPA Amendment Regulations 2026 | In force since 15 April 2026, with the opt-out registry phased in (below). Direct marketers register with the National Consumer Commission, renew yearly (ENSafrica), cleanse lists monthly, and never market to anyone with a block. |
| CPA calling times | Unless the consumer asked to be contacted, no marketing calls to a consumer at home on Sundays or public holidays; Saturdays only 09:00 to 13:00; other days 08:00 to 20:00. |
| FAIS (financial advice) | Advice on financial products stays with licensed people. An agent can book, remind and collect details, not advise. |
| Debt Collectors Act and code; National Credit Act | Collections calls follow the same conduct rules, whoever makes them. |
| US: FCC ruling, February 2024 | AI voices count as "artificial" under the TCPA, so calls need prior express consent, written for telemarketing (FCC ruling). Debt collectors also follow the FDCPA and Regulation F. |
| US health data: HIPAA | Handling patient information needs a business associate agreement (BAA) with the platform; several sign one only on enterprise plans. |
| EU and UK: GDPR | You need a lawful basis (Article 6) and, for health data, an Article 9 condition. |
South Africa's national opt-out registry is being phased in: direct marketers register with the National Consumer Commission from 15 September 2026, cleanse their lists monthly (free cleansing December 2026 to April 2027), and consumers can block marketers from May 2027 (NCC).
For the regulator's own summary of the new rules, see the National Consumer Commission's statement on the regulations; for a law firm's reading, see ENSafrica on the CPA amendment regulations. Our guide to AI cold calling covers outbound consent in more depth.
“"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.”
“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.”
Questions to put to every platform on your shortlist: where are recordings stored, and can they stay in a region you choose? Will it sign a data processing agreement (and a BAA for US health data)? Does it say it's an AI and that the call is recorded, every time? Can you load your opt-out list before a call goes out? A badge on a homepage is a start, not an answer.
Not even the best AI voice agents will decide what your agent should say, test it on your callers, train the people who take the hard calls, or keep it working when a supplier changes something. Those jobs stay with you, or whoever you pay. Here are the honest limits, and when not to use one at all.
None of the big comparison lists covers this. Concentrix said AI adoption had forced it to extend training from an average of three weeks in 2024/25 to eight weeks in 2025/26 (reported by The National and republished by BPESA). a16z wrote that realistic voice agents "can essentially act as 'simulators'" for coaching and training (a16z). Plan for agents who take handovers, supervisors who review AI calls, and staff who now get only the difficult calls. AI practice calls let people take those hard calls many times before a real customer hears them.
Don't use one if you get few calls and each is long and personal, if you can't get consent for the calls you want to make, if nobody will own the agent after launch, or if your callers would be harmed by a mistake that a person would have caught. A better-trained team or a website chat agent may serve you better. And be wary of any supplier that promises returns before the agent has been tested on your own calls.
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.
On a shortlist like the one above, we're the done-for-you option for South African callers: we build the agent for you, run it with you, and prove it before launch. A version of our AI outbound sales agent may phone nobody until it has passed every practice situation three times, inbound agents are built around your own callers and handover rules, and AI practice calls train the people who take what the agent hands over.
“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.”
When you choose among the best AI voice agents, we think the feature list matters least. Decide first who will write, test and run the agent after launch: your developers, a platform's team or a specialist. Then test the shortlist on recordings of your own callers and price the finished job, all in. We build for buyers who choose that way.
“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.”
If you're comparing the best AI voice agents for South African callers, list the five calls you'd hand to an agent first: we'll tell you whether an agent will help, the rand cost, and whether a platform above suits you better. For the wider market, see voice AI companies to know and our South African buyer's guide.
The best AI voice agents depend on who builds and runs them. Vapi, Retell AI and Deepgram's Voice Agent API suit developers; PolyAI, Sierra, Parloa and NiCE Cognigy suit large contact centres; done-for-you specialists suit businesses that want the result. Test any shortlist on your own calls at an all-in price.
An AI voice agent platform runs software that holds a spoken conversation: it understands what the caller says, decides what to answer from your information and rules, replies in a synthetic voice and connects to your phone line and systems. An IVR only routes callers through a menu and can't answer questions itself.
Common Vapi alternatives are Retell AI, Bland AI, Deepgram's Voice Agent API and the open-source LiveKit Agents and Pipecat. Bland AI published $0.12 to $0.14 a minute with the model and voice included; Retell AI published itemised prices from $0.07 to $0.31 a minute and a no-code studio (checked 16 November 2026). Neither suits everyone: test both on your own calls.
Standard pay-as-you-go agent platform prices ran from $0.05 to $0.31 a minute on 16 November 2026; hosting-only building blocks start at $0.01 and speech-to-speech models can cost more. Add the AI model, voice, phone line and add-ons, multiply by your minutes and convert at the day's exchange rate.
The common mistakes are choosing on a demo instead of your own calls, comparing headline per-minute prices that include different parts, trusting a language count instead of testing your callers' accents, accepting vendor speed claims without measuring from South Africa, ignoring consent and the opt-out registry, and forgetting to train the people who take the handed-over calls.
Not reliably out of the box. Research on the AfriSpeech-200 dataset found off-the-shelf speech recognition makes several times more errors on African-accented English than on US benchmarks. Test any platform with your own callers' accents, names, amounts and language-switching, at least three times per situation, before you buy.
Yes, within the rules. POPIA section 69 requires consent for direct marketing by automatic calling, with a narrow exception for existing customers. The 2026 Consumer Protection Act regulations phase in a national opt-out registry: marketers register with the National Consumer Commission from 15 September 2026 and must cleanse lists monthly. Calling hours apply. Take legal advice.
Pick about ten real call situations from your recordings, write a pass mark for each, and have staff with your customers' accents play the callers. Run each situation at least three times, measure reply delay from South Africa, test at peak volume, check CRM write-back, and include an opt-out request and a request for a person.
Send us the five calls you'd hand to an agent first. We'll tell you whether ShiftMate, one of the platforms above, or neither suits them.