AI agents and chatbots, a guide
AI agent or chatbot? The difference is who does the work
A chatbot answers. An AI agent answers and then does something: checks the diary, makes the booking, sends the quote, alerts a person. That one difference decides what the software needs to be connected to, what can go wrong, and what has to check it. This guide explains the difference, what agentic and autonomous AI mean without the hype, why an agent needs a server checking everything it does, and which one a small business actually needs.
Useful whether or not you ever buy from us. Written in Durban by the people who build IndunAI, which sits on the agent side of this line. Where it fits, and where it does not, is one short section near the end.
The difference
A chatbot answers; an agent acts.
Both hold a conversation, and on a good day both read the customer's message properly. The difference is where the conversation ends. A chatbot ends it with words, usually a pointer to where the customer can go next. An agent ends it with something done, in a system that matters to your business.
A chatbot
chatbot
can i get a fade friday after 519:12
We are open Monday to Saturday, 08:00 to 18:00. To book, please call the shop during opening hours.19:12
An AI agent
AI assistant
can i get a fade friday after 519:12
Friday has 17:00 and 17:30 open. A fade is R150 on our price list. Which one suits you?19:12
17:3019:13
Booked: Friday 17:30, a fade. You will get a reminder the day before.19:13
Illustrations with a made-up barber. The chatbot's answer is true and polite, and the customer still has to phone during working hours, which is the thing they were trying not to do. The agent's answer needed the shop's real diary, its price list and the right to write a booking into that diary.
| Question | Chatbot | AI agent |
|---|---|---|
| What it produces | Words | Words, plus a change in a system: a booking, a document, a payment request, an alert |
| What it is connected to | Your answers to common questions | Your calendar, price list, payment provider or CRM, as well as your answers |
| How a conversation ends | "Someone will get back to you", or a link | With the job done, or with a person alerted and the details attached |
| What a mistake costs | A wrong answer, which somebody can correct | A wrong action: a double booking, a wrong price, a message sent in your name |
| What has to check it | You, reading the answers | Software, checking every action before it happens, and you, reading the record afterwards |
What "tools" means
In agent talk, a tool is something the model can ask to use: look up free times, write a booking, send a document, create a payment request. The model does not run the tool itself. It writes a request saying which tool it wants and with what details, and the software around it decides whether to run it. That gap between asking and doing is where an agent is made safe or unsafe, and it is the subject of the third section below.
It is a spectrum, not two boxes
Plenty of products sit in between. A chatbot that fills in a booking form and emails it to you has taken one small action. An agent that can only book, and hands everything else to a person, is a narrow agent. The useful question is not which label a product wears. It is what it can change without a person, and what checks that change.
The buzzwords
Agentic and autonomous AI, without the hype.
Agentic AI is software where a language model does not only write a reply but chooses the next step and uses a tool to take it. Autonomous describes how far it goes before a person or a rule gets a say. It helps to think of four levels.
1. It answers
Reads the question and replies from what it has been given. Nothing changes outside the conversation. Most chatbots live here.
2. It drafts
Prepares something, a reply, a quote, an email, and a person approves it before it goes anywhere. Useful, and the person is still the one acting.
3. It acts within limits
Takes defined actions itself, such as booking or sending a payment link, through tools that software checks against the record and your rules. This is where most of the everyday value is.
4. It pursues a goal
Given a goal rather than a task, it plans its own steps, picks its own tools and keeps going until it decides it is done. This is what "autonomous AI agent" usually means, and where most of the risk sits.
What agents are good at, and what they are not
Agents are good at work with clear steps and a result you can check: finding a free time and booking it, sending the right form, collecting the details a quote needs, and doing it at 21:00 as well as 09:00. They are weak at long chains of steps where each one depends on the last being right, because small mistakes add up, and at anything that needs judgement, such as a complaint, a refund or an exception to your rules. The more expensive a wrong action is, and the harder it is to undo, the more a person or a hard rule should stand in front of it.
Agentic AI in South Africa
The technology is the same here as anywhere, but three local details matter. POPIA applies to every agent that handles a customer's personal information, so the supplier should sign a written operator agreement with you, and because most capable models run outside South Africa, section 72, on sending personal information across the border, applies too. Section 71 protects people from decisions with a legal or similarly substantial effect made solely by automated means, which matters if an agent is ever allowed to decide rather than assist. POPIA and AI chatbots covers both. And the channel is usually WhatsApp, where Meta's terms since 15 January 2026 bar general-purpose AI assistants from the Business API while keeping business agents for bookings, support and sales permitted.
The part that matters
Why agents need a server that checks them.
A language model is a very good reader and writer and an unreliable clerk. It can misread, it can be talked round by a persistent customer, and it can be steered by instructions hidden in a web page or document it was asked to read. An agent that acts is only as safe as the software between the model and your systems. These are the six checks that software should make.
Availability is a record
Whether a time is free is read from your calendar, never taken from the conversation, and the slot is claimed once, so two people tapping the same time cannot both get it.
Money comes from the list
Prices are read from your price list and totals are worked out by ordinary arithmetic. The model never invents a figure, and anything not on the list goes to a person to price.
A payment is confirmed by the payment
A card payment is confirmed when the payment provider confirms it. A screenshot of a bank transfer is a claim for a person to check, not proof.
Knowing is not owning
Knowing somebody's name and phone number is not the same as being them. Changing or cancelling a booking should need proof that the person asking is the person who made it.
Words are data, not orders
Text from a customer, a website or an uploaded document is something to read, never an instruction to follow. A customer naming an internal function does not get to run it.
Every action is on the record
Each action is stored with the conversation that caused it, so a person can read what the agent did, and why, and correct it.
Questions to ask any supplier
Which actions can the agent take without a person? For each one, what does the software check before it happens? What happens when two customers want the same slot in the same second? Where do the prices it quotes come from? How does it know a payment arrived? What does it do with a customer who says they are somebody else? Can you read every action it took, and step into a conversation yourself? A supplier that answers "the AI handles it" to any of these has not answered. AI booking without double booking walks through the slot question in detail.
Deciding
Which one your business needs.
An agent is not automatically better. It is more useful where conversations should end in something being done, and more work to set up properly, because it has to be connected to the things it acts on.
A chatbot is enough
When customers ask the same handful of questions, the answers rarely change, and nothing needs doing afterwards: opening hours, directions, what you do and do not offer.
You need an agent
When a good conversation should end in a booking, a quote, a payment request or a lead in your CRM, and the job currently waits until somebody is free to do it by hand.
You need a person
For complaints, refunds, exceptions, anything clinical, legal or financial, and anybody who is upset. A good agent hands these over with the conversation attached rather than trying.
You may need neither yet
If you get a few messages a day and answer them quickly yourself, the best automation may be a written price list and a booking link. Revisit it when the messages start waiting.
Where we fit
Where IndunAI fits, and where it does not.
IndunAI is an AI assistant for South African small businesses, and by this page's levels it acts within limits: it holds the conversation, takes defined actions, and has software checking each one.
What it does
It answers on your WhatsApp Business number, your website chat and email; books into Google Calendar, a CalDAV calendar or GoHighLevel; sends reminders, forms and payment links; prepares quotes and invoices from your own prices; files leads in GoHighLevel; and hands conversations to your team with the details attached.
What checks it
It is built the way this page describes: the model proposes, the server verifies. Free times come from your calendar, each slot is claimed once, prices come from your list, a card payment counts only when the provider confirms it, and by default a proof of payment comes to you to check.
What it is not
It is not an autonomous agent that sets its own goals, and not a general-purpose assistant. It answers messages, not phone calls, and its one CRM connection is GoHighLevel.
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 and build items within reason included, is R4 995 a month. Bigger work on top of it is scoped after you sign up, and an agent that does something IndunAI does not do at all is priced per project. Meta's WhatsApp fees are separate. What the agents we build look like is on AI agents, and the answering side is on the AI chatbot.
FAQ
Fair questions.
Is ChatGPT an agent?
In its everyday form ChatGPT is a chat assistant: you ask, it answers. Its makers have added features that let it take some actions, such as searching the web or working through a task in a browser, and they describe some of those as agents. What it does not have is your diary, your price list or your payment account, and a business agent is defined by exactly those connections and by the software that checks what it does with them.
What is agentic AI?
Agentic AI is software in which a language model chooses the next step and uses tools to take it, such as reading a calendar, writing a booking or sending a document, rather than only writing a reply. It ranges from narrow agents that take a few defined actions under checks to autonomous agents that plan their own steps towards a goal.
Are autonomous AI agents safe for a small business?
An agent that plans its own steps with nothing checking them is a risk wherever a wrong action costs money or cannot be undone. For customer conversations the safer design is an agent that takes defined actions, with software checking each one against your calendar, prices and rules, a hand-over to a person for anything that needs judgement, and a record of everything it did.
Do I need an AI agent or a chatbot?
A chatbot is enough when customers ask a few common questions and nothing needs doing afterwards. You need an agent when good conversations should end in a booking, a quote, a payment request or a lead filed in your CRM. You need a person for complaints, exceptions and anything that needs judgement, and a good agent hands those over.
Is IndunAI an agent or a chatbot?
An agent, in the practical sense this guide uses: it holds the conversation and it acts, booking into your calendar, sending reminders, forms and payment links, preparing quotes and invoices and handing conversations to your team. Each action goes through software that checks it first. It is not an autonomous agent that sets its own goals, and it answers messages, not phone calls.
See an agent at work.
The assistant in the corner of this page is IndunAI, set up for our own business. Ask it what your customers ask you, try to talk it into something it should not do, and judge it by the six checks above.
The rest of the plain-English guides, written the same way: what is documented, what is not, and how to check it yourself.