The three kinds of AI, which jobs each one is good at, and why a simple rule often beats a model.
Since ChatGPT arrived, many businesses reach for generative AI first and look for a problem to use it on. That gets the order wrong.
We asked Alex Shabala, Group Head of Data Science at Capitec Bank, how the bank decides which kind of AI fits which job…
The move: start with the problem, not the tool
Capitec uses a simple shared language across the bank. Analytical AI is the judge. Generative AI is the author. Agentic AI combines the two. None of them is the default; the job decides.
“We don’t start with the tool and look for a problem. The client shouldn’t care what’s under the hood.”
How to decide which kind of AI to use
1. Try a simple rule first
Rule one of Google’s Rules of Machine Learning is that the best machine learning model is often not a machine learning model at all. A rule-based system is frequently the best place to start. Alex sees it as a scale rather than a step: begin with rules, and move up only when rules stop being good enough.
2. Use analytical AI as the judge
When the job is a decision or a forecast, use analytical AI, the classic machine learning that predicts and classifies. For many system tasks it’s still the best tool available. Alex’s example: if you want to forecast your business’s account balances, ChatGPT won’t do it as reliably or at the same scale as a traditional forecasting model.
When the job is writing, generative AI earns its place: personalised messages, offers, content and help with code. Capitec uses it to change how it speaks to clients during a fraud alert, explaining exactly why a payment looks risky instead of sending a generic warning.
4. Combine them when the job needs both
Agentic systems join the judge and the author. Capitec Pulse is the example: when a client calls in, real-time systems pull up their details, predictive models rank the most likely reasons for the call, and a language model handles the messy parts of what the client says. The client-care agent sees the top five reasons before the conversation starts.
5. Measure it with a plain number
New tools still need old-fashioned measurement. For Pulse, Capitec tracks how often the real reason for the call was in the top five it showed, and works to push that number up. Alex’s point on reliability is the same: ChatGPT is useful to a person, but a business needs systems that work almost every time, so he wouldn’t want his bank built on it.
The big payoff
Matching the tool to the job keeps costs down and results reliable. You use rules where rules are enough, machine learning for decisions and forecasts, and generative AI where it genuinely writes better. The client never needs to know which one is running.
You might also be interested in
Want the full story?
Alex’s full panel from The Open Letter Cape Town 2026 is available to members inside the Founder Collab, where he goes further than we could cover here:
Why fraud tactics have a half-life of four to six weeks, and how Capitec keeps its models up to date
How Capitec writes scam warnings customers actually listen to
Why Capitec puts an expert in the loop, not just a human
When AI hallucination is fine, and when it isn’t
Why he hires for curiosity over credentials
You’ll also get access to 40+ masterclasses from SA founders and operators on sales, fundraising, UX, paid media and more inside The Founder Collab.
Get more SA tech and business news, tips and business-building workflows and subscribe to The Open Letter.
