AI voice agents for surveys phone your customers, ask a few questions as a short conversation, follow up with "why?" after a low score, record answers in fixed fields and flag unhappy people for a person to call back. They suit NPS, CSAT and research, if the survey is short, honest about being an AI and never turns into selling.


“ShiftMate is a voice agent specialist based in South Africa, and we think short feedback calls are exactly the high-volume work AI should take off people. This guide shows how to run phone surveys with AI voice agents: the questions, the rules, the measures, and getting unhappy customers to a person the same day.”
AI voice agents for surveys work like a trained phone interviewer who never gets tired: the agent phones a customer from your list, says who it is calling for and that it's an AI, asks permission, asks your questions one at a time, listens to the spoken answer, turns it into a score or a category, and asks a short follow-up when the answer needs a reason. At the end it thanks the person, writes the answers into fixed fields, and raises a flag if someone is unhappy or has a problem nobody has solved. If you're new to the technology, our AI voice agents guide explains how the parts fit together.
That is different from the keypad surveys many South Africans know from the end of a support call ("press 1 if you were satisfied"). A keypad survey can only record a number. A voice agent can hear "it was fine, I suppose, but I waited forty minutes" and record both the score and the reason, which is the part a business can act on. It's also different from a recorded robocall: the agent responds to what the person says, so it can repeat a question, slow down, or stop when asked.
Under the hood, every survey call is the same short loop, repeated for each question. Knowing the loop helps you see where things go wrong.
Something in your systems starts the call: a closed support ticket, a delivery, an installation, a policy renewal, or a monthly sample drawn for a relationship NPS survey.
The agent says its name, that it's an AI, who it's calling for, how long it will take, that the call is recorded, and asks if now is a good time. A "no" ends the call politely, and a "call me later" books a time.
One question per turn. The agent waits for the person to finish, copes with interruptions, and asks again in different words if the answer didn't fit.
The spoken answer becomes a value in a field (a score from 0 to 10, a yes or no, a category) plus the person's own words, kept as a quote.
Rules decide when to ask "why?": usually after a low score, a surprising answer, or a complaint. One follow-up, not an interrogation.
The agent thanks the person, says what happens next, and sends any problem to a person with the call summary, so the customer doesn't repeat themselves.
Hi, is that Anna? I'm Emma, an AI assistant calling for your internet provider about yesterday's installation. The call is recorded and takes about two minutes. Is now okay?
Name, AI, who for, recorded, how long, permission: all in the first breath.
Ja, quickly.
Thanks. On a scale from nought to ten, where ten is excellent, how would you rate the installation?
Yeah, no, it was fine. Maybe a six.
A six, thank you. What would have made it a nine or a ten?
Confirms the score back, then one open follow-up after a middling score.
The guy was late and nobody phoned to say so. And the router's still flashing orange.
I'm sorry about that. The flashing router sounds like it needs attention, so I'll ask someone from the support team to phone you today. Is this number the best one?
An unresolved fault becomes a call-back for a person, not just a data point.
Yes, this one.
Done. That's all my questions. Thanks for your time, Anna.
“"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.”
Anna's "Yeah, no, it was fine" is exactly the kind of answer that trips up AI phone surveys built on American speech. A survey agent that hears a no, or files "fine" as positive without the six that follows, records the wrong thing for every South African who answers that way. That's why the spoken answer has to be stored next to the score: so you can check it.
A voice survey beats a written one when you need the reasons behind the score, when the people you want to hear from don't read email, and when the experience was recent and personal enough that a short call feels like care rather than a nuisance. Voice surveys lose out when people expect to answer in their own time, when the questions are long or sensitive, and when you need thousands of answers cheaply from people who are happy to tap a link.
The honest comparison isn't "voice gets more answers". The voice completion rates quoted on vendor blogs that we checked had no source, and one widely read article gave two different ranges in two different sections. Run a pilot and measure it yourself, before you build a business case on someone else's number.
| Situation | Voice (AI survey calls) | Written (email, SMS, WhatsApp link) |
|---|---|---|
| You need the reason behind a low score | Strong: it can ask "why?" and a follow-up in the moment | Weak: open text boxes are often left empty |
| Customers who rarely read email or have little data | Strong: a phone call needs no data or app | Weak: links may never be opened |
| Long surveys (more than five questions) | Weak: people tire quickly on the phone | Better: people can pause and come back |
| Sensitive topics (health, money troubles) | Risky: people may not want to say it aloud | Better: more private to type |
| Matrix or ranking questions ("rank these six") | Poor: hard to hold in your head | Good: easy to see on a screen |
| Right after a service moment | Good: memories are fresh, the call feels like follow-up | Good: cheap and fast |
| Very large volumes at the lowest cost | Higher cost per answer | Lowest cost per answer |
| People who need help to answer (older customers, low literacy) | Strong: the agent can repeat and explain | Weak: no one to ask |
Consider mixing the two: a WhatsApp or SMS link for the quick score, and a voice follow-up for detractors and for customers who didn't reply. That keeps the cost of AI customer feedback calls where they add most value: the conversations where a reason, or an apology, matters.
Every channel has a sample bias, and AI survey calls are no exception. People who answer an unknown number in the middle of a weekday are not the same as people who answer email at night. Compare who answered against your customer base (region, age band, product) before you report a number, and keep the same channel month to month, so a change in score isn't really a change in method.
The calls that suit AI voice agents for surveys are short, frequent and tied to one moment: after a service call, a delivery, an installation or a claim. Relationship surveys such as NPS work too, as do short research interviews where the follow-up questions matter. Calls that need judgement, comfort or advice belong with people.
One satisfaction score, one "why", and a check that the problem was actually solved. Unsolved means a flag for a person, the same day.
"How likely are you to recommend us to a friend or colleague, from nought to ten?" then one reason. Ask it on a steady schedule to a fresh sample, so the trend means something.
"How easy was it to change your banking details?" Effort questions find broken processes faster than satisfaction questions, and voice catches where people got stuck.
Asked politely, with no win-back pitch on the same call. If you want to make an offer, ask permission for a separate call, and treat that one as direct marketing.
Five to eight structured questions where the follow-ups matter: testing a new idea, understanding why people chose a competitor, or hearing how a price change landed.
Did it arrive, does it work, was the person on site helpful? Insurance claims, fibre installs and furniture deliveries fit well.
Simple, non-clinical questions such as "did you get your appointment?" only. Anything about symptoms goes to a qualified person. Health information is special personal information under POPIA.
If you run a contact centre, a sensible place to start is post-call CSAT, because the client usually wants the number and team leaders want the reasons. Our guide to the 24/7 AI customer support phone agent covers the service calls that come before the survey.
You design AI phone surveys that people finish by writing for the ear, not the eye: few questions, one idea per question, scales read out in full the first time, the score confirmed back, and one open "why" where the reason matters. Every extra question costs you finishers, so cut anything you won't act on.
Most survey questions were written to be read. Read aloud by an agent, they become long, ambiguous and hard to answer. The table shows how common question types change on the phone.
| Question type | How it works by voice | Tip |
|---|---|---|
| 0 to 10 scale (NPS) | Works, but say both ends the first time: "nought means not at all likely, ten means extremely likely" | Confirm the number back; "a six, thank you" catches mishearing |
| 1 to 5 satisfaction (CSAT) | Works; words are often easier than numbers | Offer words ("very satisfied ... very dissatisfied") and map them to numbers |
| Yes or no | Easiest of all | Listen for "yes, but": the but is the data |
| Multiple choice | Hard beyond three options | Ask it open ("what was the main reason?") and let the agent sort the answer into your categories |
| Open "why" | Voice's real strength | Ask once, after the score; let the person finish; don't paraphrase their answer into something else |
| Ranking or grid | Very poor | Move to a written survey, or split into separate yes or no questions |
| Sensitive or personal | Risky | Ask only if you need it, say why, and offer to skip |
Say how long the call will take in the opening, then keep the promise. Three to five questions is a good default for after-service surveys; research interviews can run longer if you ask for that time up front. A good agent notices when someone is rushing (short answers, "I'm busy") and offers to finish with only the most important question.
Ask the overall score before the detailed questions, so the details don't colour it. Avoid leading wording such as "how much did you enjoy...". Keep the wording exactly the same on every call: one advantage of AI survey calls over people is that the interviewer never drifts, never sounds bored on call 200 and never nudges a score with tone. The flip side is that every caller hears the same voice, so a voice that sounds foreign, robotic or salesy affects everyone's answers in the same direction. Voice choice matters as much as the questions.
“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.”
AI voice agents for surveys should tell people at the start that they're speaking to an AI and that the call is recorded. Some people will decline, and that's fine: a survey built on people who were misled is data you can't defend. Disclosure also changes how some people answer, which is another reason to compare AI-collected scores with human-collected ones before you merge them into one trend line.
“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.”
Many South Africans switch between English and Afrikaans, isiZulu or another language inside one sentence. AI voice agents for surveys that only cope with one language will misfile those answers, or ask the same question again until the person gives up. Off-the-shelf speech recognition also makes several times more errors on South African-accented English than on standard US benchmarks (AfriSpeech-200, 2023). Test with recordings of your own customers, and decide in advance what the agent does when it can't understand: ask once more in simpler words, then offer a call from a person.
Call soon after the moment you're asking about, while it's fresh, but not so soon that the problem hasn't had time to show (a router that fails on day two). Set a maximum number of attempts per person, space them out, and stop for good after a refusal.
AI survey calls in South Africa fall under POPIA because they collect personal information, and a genuine survey isn't direct marketing. The moment the call promotes a product, offers a deal or asks for a sales appointment, it becomes direct marketing, and POPIA section 69 and the Consumer Protection Act's direct-marketing rules apply. This is general information, not legal advice: take legal advice for your own survey.
| Rule | What it means for AI phone surveys | Applies when |
|---|---|---|
| POPIA (processing conditions) | Have a lawful reason, tell people why you're calling and what you'll do with the answers, collect only what you need, keep it secure, keep it only as long as needed, let people see or correct it | Always: names, numbers, recordings and opinions linked to a person are personal information |
| POPIA special personal information | Health, religion and some other categories need extra justification and care; keep survey questions away from them unless you truly need them | Health follow-ups, some research |
| POPIA section 69 | Direct marketing by electronic communication, including automatic calling machines, needs consent, with a narrow exception for existing customers (your own similar products, opt-out offered when the details were collected and on every message) | As soon as the call markets anything |
| CPA Amendment Regulations 2026 | In force since 15 April 2026; they set up a national opt-out registry that is being phased in from 15 September 2026 (details below the table) | Marketing calls; a survey that pitches is a marketing call |
| CPA calling times | For a consumer at home, unless they asked to be contacted: no Sundays or public holidays, Saturdays 09:00 to 13:00, other days 08:00 to 20:00 | Direct marketing; a sensible default for surveys too |
| Calls into other countries | US: the FCC ruled in February 2024 that AI voices are "artificial" under the TCPA, so calls need prior express consent, written for telemarketing. EU and UK: GDPR needs a lawful basis (Article 6), and health data needs an Article 9 condition | Offshore campaigns run from South African contact centres |
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). Never market to anyone who has registered a block.
See also the National Consumer Commission's announcement, NCC welcomes regulations to curb spam calls, and the Information Regulator's statement on the CPA amendment regulations. For a lawyer's summary of the registration and opt-out duties, see ENSafrica's newsflash on the CPA amendment regulations.
The single most important rule for AI voice agents for surveys is to keep the calls free of selling. It's the line between a feedback call and a direct marketing call, and it's the fastest way to lose people's trust in your surveys for good. If customers learn that your "quick survey" ends with an offer, they stop answering the next one. If you want to follow up a happy customer with an offer, ask on the survey call whether they'd like a separate call about it, record the answer, and run that call under your direct marketing rules.
The same thinking applies to "stop calling me". The agent must treat it as final, mark the number so no other campaign calls it for that purpose, and end the call politely.
Survey recordings and transcripts are personal information. Decide how long you keep them (often far shorter than the scores), who can listen, and whether the supplier stores them outside South Africa, which brings in POPIA's rules on sending personal information across borders. Ask the supplier for this in writing; our guide to AI voice agents in South Africa covers POPIA, data and the opt-out registry in more depth.
You turn the spoken answers from AI voice agents for surveys into trustworthy data by storing three things for every answer (the value, the person's own words, and the agent's confidence), by checking a sample of calls against the recordings every week, and by sending anything the agent wasn't sure of to a person.
Every number in your results should link back to the words that produced it. When a manager asks why the branch scored 4.1, someone should be able to read the twenty answers behind it in two minutes. This is also how you catch systematic mistakes, such as an agent that records "not bad" as negative.
“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.”
The same rule works for survey data. If the agent must point to the words that support each field, you can check it in code, and a field with no supporting quote is flagged instead of quietly counted.
Be wary of any survey tool that claims to detect sarcasm reliably. A misread "great service, really" goes into your CSAT as a happy customer. Safer designs ask for the number directly, confirm it back, and treat the tone as a note rather than the score. Answers like "it was fine, I suppose" or "the delivery was great but the product broke" should be split into parts, not averaged into one middle value.
Open "why" answers are where the value is, and where the work is. Agree a short list of themes before you start (waiting time, staff attitude, price, product fault, communication), let the AI sort each answer into one or more themes, and have a person check a sample every week. Add a theme when "other" grows past a few percent. Report themes with real quotes next to them; a quote from a customer moves a meeting further than a percentage.
If you switch from people to AI voice agents for surveys, run both on random halves of the same sample for a few weeks before you retire the old method. Otherwise a jump in NPS may be a change of method, not of service.
A survey that hears a complaint and does nothing is worse than no survey: you've asked the customer to repeat their bad experience for your records. Decide up front what triggers a call-back (a low score, an unresolved fault, the words "cancel" or "complaint"), who receives it, and how fast. The agent should only tell the customer what will really happen.
Choose AI voice agents for surveys by testing them on your own survey and your own kind of callers, not by reading feature lists: ask each supplier to run the same test plan, check the captured data against the recordings, and measure the delay from where your customers really are. Published latency and accuracy figures are vendor claims until you've tested them on your lines.
There are three broad kinds. Developer platforms give your engineers the parts to build a survey agent, with the most control and the most work. Managed voice agents are set up and run for you by a supplier. Survey platforms with a voice option add phone calls to the questionnaires you already run. Which suits you depends on your team, not on a ranking; our overview of AI voice agent platforms explains the differences without naming a winner. Whichever you choose, check that each answer's fields and quotes flow into your CRM or helpdesk, where the people who act on them work.
| Criterion | What to ask | Why it matters for surveys |
|---|---|---|
| Data accuracy | Show us the stored fields next to the recording for 50 test calls | A wrong score is worse than a missing one |
| Disclosure and opt-out | Does it always say it's an AI? What happens after "stop calling me"? | Trust and the law |
| Accents and languages | Can we test it on recordings of our own customers? What happens when someone switches language? | South African answers are often mixed-language |
| Speed | How long between the end of a person's answer and the start of the next question, measured from a South African number? | Long pauses make people talk over the agent or hang up |
| Interruptions | What happens if the person answers before the question is finished? | People often answer early on short surveys |
| Hand-over | How does a complaint reach a person, and what do they see? | Closing the loop |
| Survey changes | Can our team change questions and rules without the supplier? | Surveys change often |
| Data location and retention | Where are recordings stored, for how long, who can access them? | POPIA and cross-border rules |
| All-in cost | Cost per minute including the phone line and AI, plus set-up, per completed survey | The per-minute price alone hides the real cost |
The best way to test AI voice agents for surveys is to call them with difficult pretend respondents before any real customer hears it. Write each one down, with a pass mark, and run the full set again every time the questions or the agent change.
| Pretend respondent | What they do | Pass mark |
|---|---|---|
| The rusher | "I've got one minute." | Agent offers to ask only the key question, and stores it correctly |
| The angry customer | Complains before the first question | Agent apologises, doesn't argue, offers a call from a person, flags it |
| The sarcastic one | "Oh, wonderful service, truly." | Agent asks for the number rather than guessing; stored score matches what they said |
| The "call me later" | Asks for another time | Agent books a time and doesn't carry on |
| The "take me off" | Asks not to be called again | Call ends politely; number marked; no further survey attempts |
| The language switcher | Answers half in Afrikaans or isiZulu | Answer understood or rephrased once, then a person offered; never misfiled |
| The rambler | Answers with a long story | Agent waits, then summarises the score and the reason correctly |
| The "is this a robot?" | Asks if it's a person | Agent says honestly that it's an AI |
| The vulnerable person | Sounds confused or distressed | Agent stops the survey and offers help from a person |
| The early answerer | Gives the score before the question ends | Agent stops talking and records the score |
“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.”
AI voice agents for surveys are their own call type. A rule that works for a sales call or a service call can be wrong on a feedback call, so test the survey on its own, even if the same supplier's agent already passed your other tests.
“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.”
One warning about testing with AI: pretend respondents written by AI tend to be too co-operative. They answer every question, accept every apology and never hang up. If your test callers are pushovers, your agent will look better than it is.
“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.”
People answering a survey expect a reply in about the time another person would take. Research across ten languages found that the gap between one speaker finishing and the next starting is typically very short, a fraction of a second (turn-taking study, PNAS, 2009). An agent that pauses noticeably after every answer makes people repeat themselves, talk over it, or hang up halfway through.
When you price AI voice agents for surveys, per-minute prices are only the start. The number that matters is the cost per completed survey: total minutes (connected calls plus voicemails, no-answers and declines) times the all-in rate, plus set-up and testing, divided by the number of completed surveys.
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).
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). These are general AI voice agent services, not quotes for survey work.
| Made-up example (not a quote) | Value |
|---|---|
| Customers on the list | 1,000 |
| Call attempts (up to three each) | 2,200 |
| Completed surveys | 300 |
| Minutes used across all attempts | 1,500 (5 minutes per completed survey) |
| All-in rate (phone line plus AI) | R4 a minute |
| Call cost | 1,500 x R4 = R6,000 |
| Cost per completed survey | R6,000 / 300 = R20, before set-up and testing |
AI voice agents for surveys don't fix a bad questionnaire, don't make an unrepresentative sample representative, and don't replace the person who has to call back an angry customer. Here's where voice AI isn't the right tool, or needs a person beside it.
Anything over about five questions, with grids or rankings, is better on a screen. Voice is for short surveys and the follow-up conversation.
Symptoms, mental health, debt troubles and similar topics need a person, or a private written form. A voice agent should stop and offer help, not keep asking.
The agent can hear a complaint and pass it on. It can't refund, repair or apologise on behalf of a manager. If nobody acts on the flags, the survey makes things worse.
Speech recognition still makes mistakes, especially on accents, noisy lines and mixed languages. Some answers will need a person to check them.
Phone surveys reach people who answer the phone. That's a sample bias you correct with method, not with technology.
A voice agent can't make a survey with a sales pitch at the end legal or trusted. Keep selling on separate, consented calls.
There's a wider point. Analysts expect AI and people to keep working side by side in customer service: Gartner predicts that by 2028 none of the Fortune 500 will have fully eliminated human customer service (Gartner, September 2025). AI voice agents for surveys are a good example of why. The survey can be automated; listening, deciding and putting things right is still people's work. If you want to understand the common ways voice agents go wrong more broadly, see why AI voice agents fail.
We think the value of a feedback call is in what happens after it. An AI can ask every customer the same question the same way; the survey only pays off when a person phones the unhappy ones back that day and fixes what went wrong. That's why we'd rather design a call around the one answer you need than read out a long questionnaire, and why we'd like to hear from customer experience teams and call centres that already know who will make those call-backs.
For them, the first step is a conversation: talk to us about whether your feedback calls suit a voice agent built for your business. AI practice calls let your people rehearse what a survey uncovers (the angry detractor, the fault that was never fixed, the customer who wants to cancel) before a real customer hears them, with plans on the Sim Studio pricing page.
Tell us the one thing you most want to learn from your customers. We'll say whether a voice agent, a written survey or a person on the phone is the right tool for it.
“Companies are shifting towards products and models that produce results. Voice AI should be judged on outcomes, the lead qualified, the meeting booked, the agent ready for the floor, not on how many minutes it talks. That's what we build for.”
Yes. AI voice agents for surveys can phone customers, ask survey questions as a short conversation, follow up with "why?" after a low score, store answers in fixed fields, and flag unhappy people for a call-back. They work best for short surveys of three to five questions, with clear disclosure that the caller is an AI.
No reliable general figure exists. The voice completion rates quoted on vendor blogs we checked had no source, and some contradicted themselves. Response depends on your customers, timing, opening and survey length. Run a pilot of a few hundred calls and measure completion, refusals and drop-offs yourself before you commit.
No, not if it's a genuine survey. A call that only asks for feedback isn't direct marketing, but adding an offer, upgrade or sales appointment makes it marketing, bringing in POPIA section 69 consent and the national opt-out registry, being phased in from 15 September 2026. This is general information; take legal advice for your situation.
Yes. Say it's an AI, who it's calling for and that the call is recorded, in the first few seconds. Some people decline, which is fine. Hiding it costs trust, and answers from people who were misled are data you can't defend. Calling US numbers brings a separate rule: under a 2024 FCC ruling, AI-voice calls need the person's prior express consent.
Work it out per completed survey, not per minute: total minutes across all attempts times the all-in rate (phone line plus AI), plus set-up, divided by completed surveys. For scale, a third-party comparison site listed South African AI voice agent services at about R0.97 to R13 a minute beyond their bundled minutes (October 2026); those are general services, not survey quotes.
Four common ones are: adding a sales offer, which turns a survey into direct marketing; asking too many questions, so people drop off; having no named owner for call-backs, so complaints go nowhere; and trusting automatic sarcasm detection instead of asking for the score and confirming it back. Testing on difficult pretend respondents before launch catches most of them.
Not always out of the box. Research on African-accented English (AfriSpeech-200, 2023) found off-the-shelf speech recognition makes several times more errors than on standard US benchmarks. Test AI voice agents for surveys on recordings of your own customers, including people switching between English and Afrikaans or isiZulu, before it goes live.
For feedback calls, talk to ShiftMate about whether yours suit a voice agent built for your business, and whether a voice call is the right tool at all. It also sells AI practice calls that train your team for the hard conversations a survey uncovers, and an AI outbound sales agent that calls, qualifies and books leads.
Talk to us about whether your feedback calls suit a voice agent built for your business, and train the people who make the call-backs. Tell us which customers you'd ask first.
The complete guide: what they are and how to choose.
Answering service calls day and night, and handing over well.
And how to test them before customers hear them.
Accents, languages, POPIA, the opt-out registry and rand costs.