1 · What it actually is
Four different things get sold under the one name, and confusing them is how money gets wasted. A chat assistant helps you write and think while you are at the keyboard, and is priced per person. A website chatbot answers your customers, trained on your business, at any hour, priced per business. Workflow automation moves data between systems and often contains no AI at all. A custom build is something specific to how your business works, quoted from scope.
The first makes you faster at work you were going to do anyway. The second means work happens when you are not there. Most of the confusion about whether "AI works" is really confusion between those two, and there is a longer version in what AI automation actually means.
2 · What to automate first
Whichever job you can count. In order of how easily you can check whether it worked:
The enquiry nobody answers. Countable, immediate, cheapest to fix, and it requires no change to how anybody works. This is first for almost every business.
The quote that sits. If a quote takes three days to leave the office, that is a queue rather than a workload, and queues are what software is for.
The shared inbox nobody owns. info@ becomes a backlog in every business that has one.
The diary somebody manages by hand. Bookings and the reminders that protect them.
Notice what is not on that list: content generation, social posting, and anything described as a growth engine. They may be worth doing, but you cannot tell whether they worked — and an unmeasurable subscription is how a business ends up paying for four tools it has forgotten about.
3 · How to decide whether it is worth it
Three numbers you already have. How many enquiries you received last month. How many you never replied to — the honest count, not the estimate. And what an average job or order is worth.
The second times the third, times whatever conversion rate you believe, is the ceiling on what fixing it can be worth to you. Compare that with the monthly fee. If it does not clear it comfortably, do not buy anything. There are businesses whose phone is always answered and whose inbox is empty by close, and they should keep their money.
4 · What is genuinely different here
Language. Eleven official languages and real code-switching in ordinary conversation. A model trained mostly on American English handles this worse than its benchmark scores suggest, and a customer notices immediately.
Load-shedding. Your phone line, your card machine and your Wi-Fi go down with the power; a cloud system does not, and your customers are on their own phones. The bot is answering hardest at the hour you are least able to.
Rands, not dollars. Most well-known tools price in USD, so your monthly cost moves with the exchange rate whether or not the vendor changes its price.
POPIA. Stricter on direct marketing than most people assume, and it applies to whatever you automate. There is a full write-up in AI and POPIA, and why South Africa covers the supplier argument.
5 · What you should never automate
Complaints. Exceptions. Anything with a legal, medical or financial consequence. Anything where a confident wrong answer costs more than no answer.
The important part is that a well-built system refuses these rather than improvising, and hands the person to somebody who can help. That is a configuration decision made before you go live, not a hope that the model behaves. If a system improvises once in an hour of you deliberately trying to break it, it will improvise on a customer.
And if somebody asks outright whether they are talking to a person, the answer is the truth. A denial found out later costs more than the question ever would have.
6 · How to train it properly
The build takes 30 minutes. The part that decides quality is two to three hours of your own writing: your prices exactly as you charge them, the twenty questions you actually get, what you do not do, your escalation rules, and three replies you were happy with.
"What you do not do" is the single most-skipped input and it prevents the worst answers. Then spend an hour trying to make it fail before anybody else sees it. There is a step-by-step version in how to train AI on your business.
7 · Where to go next
Two pages carry on from here. How to choose an AI chatbot is the eight checks that separate a good supplier from a bad one, each of which you can run yourself in an afternoon. Every question, answered straight is the short version of everything above — cost, POPIA, install, limits and supplier choice — in one page you can read in ten minutes.
And if you are still deciding whether it is a chatbot you need at all, compare the options puts a chatbot, live chat, a contact form and a person in one table.