Conversational AI, a plain-English guide
What conversational AI is, and what it is not
Conversational AI is software that holds a conversation in ordinary language: it reads what somebody actually wrote, works out what they want, and answers in words. The term is stretched to cover everything from a button menu to a phone robot, so this guide sets out the definitions, the real difference between a menu bot and one built on a language model, the ways it goes wrong, and what to ask before you choose a platform.
The bubble in the corner of this page is conversational AI. It is our own IndunAI assistant, answering for us. Ask it something awkward and judge it by the standards below. Written in Durban by the people who build it.
AI assistant, replies in seconds
hey do u guys do sat mornings and hw much for a gel overlay20:41
We do, Saturdays from 08:00 to 13:00. A gel overlay is R280 on our price list. Would you like to see Saturday's open times?20:41
An illustration with a made-up salon. Shorthand, no punctuation and two questions in one message: reading that is the whole point of conversational AI.
Definitions
Definitions.
Most of the confusion in this market comes from six terms used as if they meant the same thing. They do not, and the differences decide what you are actually buying.
Conversational AI
The umbrella term: software that understands people and answers them in ordinary language, in text or by voice. It covers the understanding (what does this message mean?), the deciding (what should happen next?) and the answering (which words go back?).
Chatbot
Any software that chats. The older kind follows a script of buttons and keywords; the newer kind uses a language model. Every conversational AI chatbot is a chatbot, but plenty of chatbots are not conversational AI.
Large language model
The engine inside most new conversational AI: a model trained on a very large amount of text to predict what comes next, which makes it very good at reading and writing. It knows nothing about your business unless it is told, on every reply.
Intents and entities
The older way of understanding a message: sort it into an intent, such as "make a booking", and pull out entities, such as a date and a time. Still useful inside systems, but brittle on its own, because people do not write in intents.
Conversational AI platform
The product around the model: the channels it answers on, the knowledge it answers from, the tools it can use, how it hands over to a person, and where you read what it said. Two platforms on the same model can behave completely differently.
Voice assistant
The same idea on a phone call: speech is turned into text, the conversation runs, and the reply is spoken back. It brings problems of its own, such as accents, background noise and people talking over the machine.
Generative AI is the wider family language models belong to: models that produce new text, images or sound. Conversational AI is one use of it. The next step along is agentic AI, where the model also takes actions, such as making the booking rather than describing how to make one; AI agent or chatbot covers that difference.
Being straight about it
Where it goes wrong.
Conversational AI fails in predictable ways. None of them is a reason not to use it. All of them are reasons to ask a platform how it deals with them.
It makes things up
Asked something it does not know, a language model will often produce a fluent, plausible answer anyway, which is usually called hallucination. The fix is to give it only facts you have written, have it say so when a fact is missing, and let software supply the facts that cost money, such as prices and free times.
It can be talked round
Models are trained to be agreeable. A customer who insists firmly enough that they were promised a time or a discount will sometimes get agreement. Anything that matters should be decided by a record, not by the conversation.
It wanders off the job
A general-purpose model will happily write an essay or discuss the news. A business assistant should stay on the business. On WhatsApp it is also a terms question: since 15 January 2026 Meta's WhatsApp Business terms have barred general-purpose AI assistants from the Business API, while assistants for support, bookings and sales remain permitted.
It is taken for a person
Customers are entitled to know they are talking to software. A good assistant says what it is, at the start and whenever it is asked.
It answers what it should not
Medical, legal and financial advice, or a customer in distress. These need hard limits the model cannot argue its way past, and a route to a person.
It has no way out
A conversation that needs a human must reach one, with the whole exchange attached, so the customer is not asked everything again by the person who picks it up.
The words go somewhere
Most capable language models run outside South Africa, so what a customer types is sent offshore to be answered. That can be handled properly under POPIA, but it has to be known and disclosed. POPIA and AI chatbots sets out what the Act asks.
Before you sign
Choosing a platform.
A conversational AI platform is judged by what happens around the model, not by the model's name. These questions separate one platform from another, and none of them needs you to be technical.
Which channels, on whose number?
WhatsApp, website chat, email, voice. On WhatsApp it should run through Meta's official WhatsApp Business Platform, on your own business number. Tools that drive the ordinary WhatsApp app from outside break WhatsApp's terms and put the number at risk.
Where do its answers come from?
Your services, prices, hours and written answers, given to the model on every reply, or whatever it can guess. Ask what it does when a fact is missing.
What can it do, not just say?
Book into your real calendar, send a quote or a payment link, file a lead, hand over to a person. "Someone will be in touch" is where many chatbots stop.
What does software check?
Whether a time is free, what something costs, whether a payment arrived. If the answer is "the AI handles it", ask again.
How does it hand over?
To whom, with what attached, and can you step into a conversation yourself and have the assistant stay out of it while you do?
Can you read everything?
Every conversation, as it happens, in a dashboard you control. An assistant you cannot read is an assistant you cannot correct.
Where is the data, and what is in writing?
Where conversations are stored, where the model runs, and whether there is a written operator agreement, which POPIA requires of anybody processing personal information on your behalf.
How is it priced?
Per message, per conversation, per user, or only after a call. On WhatsApp, Meta's own message fees come on top. Compare on the same unit, at your own volume.
What it costs
A conversational AI bill usually has three parts: the platform, the language model's usage (often bundled into the platform's price, sometimes passed through), and on WhatsApp, Meta's message fees. Setup can be a fourth. A low headline price with a tight message cap, or a price per user multiplied by a team of five, can cost more than it looks, so work it out on a real month of your own messages. WhatsApp Business costs shows how to work out the Meta part, and choosing an AI receptionist sets out how this market prices.
Where we fit
Where IndunAI fits, and where it does not.
IndunAI is a conversational AI platform for South African small businesses. By this page's definitions it is the language-model kind, with buttons where they help and software checking the facts.
What it is
One assistant on your WhatsApp Business number, your website chat and email. It answers from your own services, prices and written answers, says it will check with the team when a fact is missing, and books, quotes and hands over to your team. It is set to reply in the language the customer writes in.
What software checks
Free times are read from your real calendar and each slot is claimed once, prices come from your own list, and quote and invoice totals are calculated by the server. The website chat introduces itself as your business's AI assistant.
What it does not do
It answers messages, not phone calls: answering WhatsApp voice calls is built but not switched on for anyone. It is not a general-purpose assistant: somebody who asks it for essays or code is steered back to your business.
It costs from R1 450 a month including VAT for up to 300 conversations, plus a once-off R2 500 setup; fully managed, with setup included and build items within reason, is R4 995 a month. Meta's fees are separate. More on the product itself is on the AI chatbot for your business.
FAQ
Fair questions.
Is conversational AI the same as a chatbot?
Not quite. A chatbot is any software that chats, including the older kind that only follows a menu of buttons and keywords. Conversational AI is the kind that understands what somebody actually wrote and answers in ordinary language, which today usually means a large language model. Every conversational AI chatbot is a chatbot, but not every chatbot is conversational AI.
What is a conversational AI platform?
The product around the language model: the channels it answers on, the business knowledge it answers from, the tools it can use, such as a calendar, a payment link or a CRM, the hand-over to a person, and the dashboard where you read and control it. The model writes the words; the platform decides what it is allowed to do.
Can conversational AI answer phone calls?
Voice versions exist, which turn speech into text, run the same kind of conversation and speak the reply. IndunAI does not answer phone calls today: it answers WhatsApp messages, including voice notes, website chat and email. Answering WhatsApp voice calls is built but not switched on for any business, so we do not sell it.
Does conversational AI work in South African languages?
Quality generally follows how much of a language is written online, so English is strongest, Afrikaans is usually good, and isiZulu, isiXhosa and the other South African languages vary more. Test any platform in the languages your customers actually write in before you rely on it. IndunAI is set to reply in the language the customer writes in and to keep the whole message in that language.
What does conversational AI cost?
It depends on the platform's unit: some charge per message, some per conversation, some per user and some only after a call, and on WhatsApp Meta charges its own message fees on top. Compare platforms on the same unit, at your own volume. IndunAI is from R1 450 a month including VAT for up to 300 conversations, plus a once-off R2 500 setup; fully managed, with setup included and build items within reason, is R4 995 a month. Meta's fees are separate.
Chat to one.
The assistant on this page is conversational AI, set up for our own business. Ask it what your customers ask you, try to trip it up, and hold it to the questions above.
The rest of the plain-English guides, written the same way: what is documented, what is not, and how to check it yourself.