Three stages of AI that every leader needs to understand

Most conversations are[–>AI In South African boardrooms, we skip the most important question: how does this technology actually work? It’s about strategic principles, not technical details.

Understanding the three different stages of AI development (machine learning, large-scale language models, and emerging neurosymbolic paradigms) is not essential for leaders making important decisions about AI investments. it is the foundation upon which everything is supported[–>governancethe choice of deployment and value measurement remains.

I have sat on many corporate AI steering committee meetings and found that most senior leaders approve AI budgets without a working model of what the technology can and cannot do. This is not a criticism. It reflects how quickly this field moves.

But the gaps are important because AI limits are not random. They are structural. And understanding these structures is what separates organizations that successfully implement AI from those that deploy it at high cost and then quietly retire it.

The evolution of modern AI can be defined in three stages. Each builds on the last. Each has particular strengths, particular failure modes and particular implications for South African enterprise deployment.

Understanding all three, not just what vendors are currently selling, is the most important strategic literacy technology leaders can develop in 2026.

Machine learning (ML) is the foundation on which everything else is built. ML models are trained on large datasets to identify statistical patterns, rather than being explicitly programmed using rules.

Show me thousands of images labeled “”[–>fraudLearn how to tell when a new transaction is ‘not a scam’ and classify it. If you show it millions of customer records, it will cluster them by behavioral similarities without being told how.

ML doesn’t understand what it’s looking at. Find correlations, the features that most reliably distinguish one result from another and apply to new inputs.

Machine learning finds patterns. It doesn’t understand them.

This is really powerful. ML has transformed credit scoring, predictive maintenance, demand forecasting and fraud detection across financial services and telecommunications in South Africa. But it has a structural ceiling, and we can only generalize from the patterns we’ve seen so far.

They often fail quietly and confidently when they encounter situations outside the scope of their training distribution, such as new fraud patterns, unusual customer profiles, or situations not covered by training data. ML doesn’t know what it doesn’t know.

Machine learning finds patterns. It doesn’t understand them. This difference is not a flaw that can be fixed. This is an architecture, and understanding it is the starting point for any serious AI investment decision.

Large-scale language models (LLMs), the technology behind ChatGPT, Claude, and Gemini, represent a significant advancement over traditional ML.

Trained on vast amounts of human-generated text across every conceivable domain, LLM develops an extremely rich representation of language, conceptual relationships, and contextual meaning.

While traditional ML models are trained specifically for a specific task, LLM generalizes across domains. Draft legal summaries, describe financial products, write code, and translate between languages, all from the same model.

That leap is real. However, the underlying mechanism remains unchanged. LLM still produces output by predicting the next word that is statistically most likely based on its training. They do not reason based on first principles. They don’t verify against ground truth.

OpenAI’s own September 2025 study confirmed that standard LLM training rewards confident guesses over adjusted uncertainty. This means that the model is encouraged to produce responses that sound fluent and authoritative, even if those responses are factually incorrect.

In the industry we call this an illusion. It’s not a software bug. This is the inevitable result of building a system based on pattern completion and asking for inferences from it.

In low-risk consumer applications, hallucinations are a nuisance. AI hinders adoption in credit adjudication, regulatory reporting, clinical decision support and AML inference – areas where South African businesses can benefit the most from AI.

SARB, FSB, and POPIA are working together to create a compliance environment where outputs generated by AI must be explainable, auditable, and defensible. A system that confidently generates wrong answers cannot meet these requirements, no matter how fluently written.

NeuroSymbolic AI (NeSy) is the field’s most proven response to the hallucination problem and the logical next step in the evolution of enterprise AI capabilities.

It combines the pattern recognition and language fluency of neural networks with the logical rigor of symbolic reasoning: formally encoded rules, knowledge graphs, and logical reasoning engines that operate on the output of neural layers.

The mechanism is elegant. LLM generates response candidates. A symbolic reasoning engine checks its response against a formally encoded knowledge base or ontology.

If the output of the LLM violates known facts or logical rules, the symbolic engine flags the contradiction and generates feedback. LLM iterates the compliant response. The results are fluent from the neural layer and logically verifiable from the symbolic layer.

DeepMind’s AlphaGeometry combines a neural language model and a symbolic deduction engine to solve International Mathematics Olympiad problems at gold medalist level, and is groundbreaking proof of what this combination can accomplish.

Amazon introduced NeSy technology in its warehouse robot Vulcan and shopping assistant Rufus in 2025. CoreThink, a symbolic inference layer, achieved 62.3% on the software engineering benchmark without any tweaks to the underlying LLM. A study published in October 2025 demonstrated 96% accuracy on a complex reasoning task using adaptive symbolic techniques. This is 25% higher than the best pure neural baseline.

For South African businesses, a practical entry point is the knowledge graph. Structure your regulatory rulebooks, product catalogs, and compliance obligations as machine-queryable structured data and combine that structure with your existing LLM infrastructure as a validation layer.

This is available starting today. There is no need to retrain the underlying model. It then generates a step-by-step logical reasoning trace that audit requirements, POPIA accountability, and regulatory explainability frameworks are beginning to mandate for high-risk AI decision-making.

These three stages (pattern recognition, language modeling, and symbolic reasoning) are more than just technological history. These are the diagnostic frameworks technology leaders need to evaluate the AI ​​systems currently on the agenda.

The most important question is not “What can I do with this AI?” The question is, “What kind of system is this, and what does it mean that it cannot be done?”

I have a question for all South African leaders reading this. The next time your board approves the implementation of AI, will someone in the room be able to accurately answer that question and manage the gap accordingly?

If not, technology is ahead of governance. And regulated industries will eventually close that gap at great cost.

* Eugene Perumal is a Head of Strategy and Architecture with over 20 years of experience in enterprise technology across telecommunications 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 about enterprise AI strategy, ROI measurement, and the transition to agentic AI deployments.