How to think about AI in your business
If you run a business, you are being sold AI from every direction. By software vendors, by consultants, by the bank, by LinkedIn, by the competitor who mentioned it at a breakfast. Some of it is real and will genuinely change how you operate. Most of what is being sold to you will not make you a cent.
Telling the two apart is the whole skill, and it has very little to do with understanding the technology.
There is one question that does most of the work: am I making more money?
Not "is it impressive." Not "is this the future." Are you making more money than you were before you started spending on it? Hold that question at the front, and most of the noise sorts itself out.
Start with your constraint, not with the technology
At any moment, one thing is limiting your business. Usually only one.
It might be demand — not enough people know you exist, or not enough of them say yes. It might be cash — you can see the growth but you can't fund it. It might be capacity — you are turning work away, or delivery is slipping and the complaints are starting. It might be one person, usually you, that every decision has to pass through.
Here is the pattern that plays out again and again. A business is paying a team to do repetitive back-office work. It costs, say, R150,000 a month, and it works — the job gets done, nobody is complaining about it. Then the business spends two or three million rand building a system to automate that team away.
Read that again. Well over a year of costs, paid up front, to replace something that was already working. And the thing actually holding that business back was never the back office. It was that not enough new customers were coming through the door.
The money wasn't wasted because it went to AI. It was wasted because it went somewhere that wasn't the constraint. You end up with a faster back office and the same empty order book.
So before any AI conversation, do this: write down what is limiting your business, in one sentence. Then ask whether the thing you are about to buy or build actually touches it.
If it doesn't, it doesn't matter how good the demonstration was. You will get an excellent version of something that was never your problem — and you will have spent the money and, more expensively, your attention.
That single question kills most AI projects before they start. That is not a bug.
Foundation first, or you get wrong answers faster
Think about building a tower out of blocks. If someone gives you five seconds, you stack them straight up and hope. Five minutes, you build differently. Five days, differently again. Five years, and you would spend most of that time on what sits underneath.
The height you are aiming for decides the foundation you need. And the businesses that shoot up and then stall are almost always the ones that laid a one-storey foundation and then tried to add nine floors to it. The fix, when it eventually comes, is to go back down and rebuild — slow, expensive, and thoroughly demoralising.
AI is the clearest example of this in years, because it does not fix foundations. It amplifies whatever is already underneath.
If your customer list lives in three places and none of them agree, no tool will reconcile your business for you — it will give you three confident answers. If your stock counts are guesswork, nothing will tell you your true margin. And if your books are three months behind with a pile of unexplained transactions sitting in a suspense account, you can point the most capable system in the world at them and get a fluent, well-argued, beautifully formatted analysis of numbers that are simply wrong.
That last case is the genuinely dangerous one, because the output looks better than anything you have had before. Confidence and correctness are not the same thing, and these tools are far better at the first than the second.
The order of operations doesn't bend:
- Get the data right. One source of truth for the things that matter — sales, stock, debtors, cash. Boring, and everything else rests on it.
- Get the rhythm right. The numbers land on a fixed date every month, whether the month was calm or chaos. A process that only happens when things are quiet isn't a process.
- Get the question right. Know what you are actually trying to decide — whether to raise prices, whether you can afford the hire, whether the big customer is worth keeping.
- Then apply the technology, at whichever step is genuinely slow.
Jump to step four and you have built a fast machine on sand.
What AI is genuinely good at
None of this is scepticism about the technology. In the right place it is remarkable. It is very good at:
- Reading and extracting. Pulling information out of statements, invoices, contracts, application forms, PDFs — work that used to be hours of typing.
- Sorting and matching. Classifying things, checking one list against another, flagging the items that don't fit the pattern.
- First drafts. A proposal, a job advert, a policy, a difficult email. A draft you edit is far cheaper than a blank page.
- Finding the odd one out. Going through everything to find the thing a person would only catch by luck. Machines don't get bored, and boredom is where mistakes hide.
- Checking work. Looking over something already done and asking what's missing. It is better at this than most people expect.
Look at the shape of that list. Every item is a job with a right answer, where the difficulty is volume and attention rather than judgement, and where you can check the result afterwards. That is the sweet spot — and it is a wide one. If something in your business fits that shape, take it seriously.
What AI isn't for
Judgement calls. Do you take the order at a thin margin to keep the team busy, or hold your price? There is no right answer sitting in the data. It depends on your capacity, your cash, and what you are trying to build. Ask a machine and you will get a confident answer either way — and it has no stake in whether it's right.
Owning the decision. Software doesn't pay tax, doesn't sit in front of SARS, and doesn't sign anything. Somebody still has to make the call and carry it. Better analysis doesn't shift that an inch.
Applying pressure over time. A forecast doesn't help you because it was produced. It helps you because someone sits down with you every month, asks why it didn't happen, and doesn't let the same excuse run twice. No tool does that. That is a person, and it is most of the value.
Being your marketing. Your customers don't care that you use AI. They care about the result. Advertising yourself as an "AI-powered" anything is describing your process to someone who only asked about the outcome — and plenty of people are quietly wary of it. Sell the outcome. Use whatever you like to deliver it.
The trap that matters most: don't outsource the thinking
This is the one to underline.
A useful habit: put the same question to two or three different AI tools and watch them come back with different answers. It cures you of the illusion within about ten minutes, and it is why the people who use these tools most seriously still back their own judgement over the output.
The deeper problem is that these systems are agreeable. Frame a question so that it hints at the answer you are hoping for, and you will get that answer — argued well enough to feel like independent confirmation. That isn't analysis. It's an expensive mirror.
And the thinking you would most love to hand over — the hard calls, the ones that keep you awake — is exactly the thinking you most need to stay sharp at. Your judgement about your own business is the most valuable thing you own. Hand it over and it weakens quietly, and you won't notice until the day you need it.
Use these tools to do the work. Don't use them to do the thinking. Easy to say. Much harder to hold.
A sane way to actually try something
If you want to move, move small and honestly.
- Pick one task that is tied to your constraint and fits the shape above — repetitive, has a right answer, you can check it.
- Time it first. Know what it costs you today, in hours or in rands. If you can't measure the before, you will never know whether anything improved.
- Put a review date in the diary now — sixty or ninety days out.
- On that date, ask the only question that counts. Are you making more money, or has something you genuinely needed actually freed up? If neither, stop. Stopping is a result, not a failure.
Most businesses skip the "time it first" step, then can't tell whether anything got better, so the tool stays on forever on the strength of a feeling.
The question to ask yourself
Put the technology aside and answer this one:
What decision did I make in the last six months where I didn't have the numbers I needed?
A price increase made on instinct. A hire committed to without knowing whether cash would carry it. A customer kept on who may well be costing you money. A tax bill that arrived and hurt.
If two or three come to mind quickly, that is your constraint — and it is a foundation problem, not an AI problem. No amount of clever software layered on top will answer a question your business currently can't see.
Build the foundation. Then let the technology make it cheaper.
Further reading: why the answer usually isn't a CFO, and how to read your management accounts.