Most conversations about AI in South African boardrooms skip the most important question: how does this technology actually work? Not in technical detail, but in strategic principle.
Understanding the three distinct stages of AI development − machine learning, large language models and the emerging NeuroSymbolic paradigm − is not optional for leaders making significant AI investment decisions. It is the foundation on which every governance, deployment and value measurement choice rests.
I have sat in enough enterprise AI steering committees to know that most senior leaders approve AI budgets without a working model of what the technology can and cannot do. That is not a criticism − it reflects how quickly this field has moved.
But the gap matters, because the limitations of AI are not random. They are structural. And understanding those structures is what separates organisations that deploy AI successfully from those that deploy it expensively and then quietly retire it.
Three stages define the evolution of modern AI. Each builds on the last. Each has specific strengths, specific failure modes and specific implications for South African enterprise deployment.
Understanding all three, not just the one your vendor is currently selling, is the most important strategic literacy a technology leader can develop in 2026.
Stage one: Machine learning − the power and limit of pattern recognition
Machine learning (ML) is the foundation on which everything else is built. Instead of being explicitly programmed with rules, an ML model is trained on large datasets to identify statistical patterns.
Show it thousands of images labelled ‘fraud’ and ‘not fraud’ and it will learn to classify new transactions. Show it millions of customer records and it will cluster them by behavioural similarity without being told how.
ML does not understand what it is looking at. It finds correlations − the features that most reliably distinguish one outcome from another and applies them to new inputs.
Machine learning finds patterns. It does not understand them.
This is genuinely powerful. ML has transformed credit scoring, predictive maintenance, demand forecasting and fraud detection across South African financial services and telecommunications. But it has a structural ceiling: it can only generalise from patterns it has seen.
When it encounters a situation outside its training distribution − a novel fraud pattern, an unusual customer profile, a context its training data did not cover − it fails, often silently and confidently. ML does not know what it does not know.
Machine learning finds patterns. It does not understand them. That distinction is not a flaw to be patched. It is the architecture, and understanding it is the starting point for every serious AI investment decision.
Stage two: Large language models − a genuine leap, with a stubborn weakness
Large language models (LLMs) − the technology behind ChatGPT, Claude and Gemini − represent a significant advance on traditional ML.
Trained on vast quantities of human-generated text across every domain imaginable, LLMs develop a remarkably rich representation of language, concept relationships and contextual meaning.
Where a traditional ML model is trained narrowly on a specific task, an LLM generalises across domains. It can draft a legal summary, explain a financial product, write code and translate between languages − all from the same model.
The leap is real. But the underlying mechanism has not changed: LLMs still generate output by predicting the most statistically likely next word given their training. They do not reason from first principles. They do not verify against ground truth.
OpenAI’s own September 2025 research confirmed that standard LLM training rewards confident guessing over calibrated uncertainty, meaning the model is incentivised to produce a fluent, authoritative-sounding response even when that response is factually wrong.
The industry calls this hallucination. It is not a software bug. It is the inevitable consequence of building a system on pattern completion and asking it to reason.
In low-stakes consumer applications, hallucination is an annoyance. In credit decisioning, regulatory reporting, clinical decision support and AML reasoning − the domains where South African enterprises stand to gain the most from AI − it is a deployment blocker.
The SARB, FSB and POPIA together create a compliance environment in which AI-generated outputs must be explainable, auditable and defensible. A system that confidently produces wrong answers cannot satisfy those requirements, regardless of how fluently it writes.
Stage three: NeuroSymbolic AI − the way forward
NeuroSymbolic AI (NeSy) is the field’s most substantiated response to the hallucination problem − and the logical next step in the evolution of enterprise AI capability.
It combines the pattern recognition and language fluency of neural networks with the logical rigour of symbolic reasoning: formally encoded rules, knowledge graphs and logical inference engines that operate on the output of the neural layer.
The mechanism is elegant. An LLM generates a candidate response. A symbolic reasoning engine checks that response against a formally encoded knowledge base or ontology.
Where the LLM’s output violates a known fact or logical rule, the symbolic engine flags the contradiction and generates feedback. The LLM iterates to a compliant response. The result is both fluent − from the neural layer − and logically verifiable from the symbolic layer.
DeepMind’s AlphaGeometry, which combines a neural language model with a symbolic deduction engine to solve International Mathematical Olympiad problems at gold-medallist level, is the landmark proof of what this combination can achieve.
Amazon deployed NeSy techniques in 2025 across its Vulcan warehouse robots and Rufus shopping assistant. CoreThink, a symbolic reasoning layer, achieved 62.3% on software engineering benchmarks without fine-tuning the underlying LLM. Research published in October 2025 demonstrated 96% accuracy on composite reasoning tasks using adaptive symbolic methods − 25% above the best purely neural baseline.
For South African enterprises, the practical entry point is the knowledge graph: structuring regulatory rulebooks, product catalogues and compliance obligations as machine-queryable structured data, then coupling that structure to existing LLM infrastructure as a verification layer.
This is deployable today. It requires no retraining of underlying models. And it produces the step-by-step logical reasoning traces that audit requirements, POPIA accountability obligations and regulatory explainability frameworks are beginning to mandate for high-risk AI decisions.
The strategic question for every South African board
These three stages − pattern recognition, language modelling and symbolic reasoning − are not merely technical history. They are the diagnostic framework every technology leader needs to evaluate the AI systems currently on their agenda.
The most important question is not “what can this AI do?” It is “what kind of system is this, and what does that mean for what it cannot do?”
My question for every South African leader reading this: when your board next approves an AI deployment, will someone in the room be able to answer that question with precision and govern the gaps accordingly?
If not, the technology is ahead of the governance. And in regulated industries, that gap eventually closes at significant cost.
* Eugene Perumal is a strategy and architecture principal with over 20 years' experience in enterprise technology across telecoms and financial services, including senior roles at Vodacom Group and Absa Group. He holds Master’s degrees and certifications in enterprise architecture, AI governance, cloud and analytics. He writes on enterprise AI strategy, ROI measurement and the shift to agentic AI deployment.
Share
