Generative AI has dominated headlines for three years. But for most South African SMEs, the path from experimentation to production value has been unclear. Here's what we're seeing on the ground.
Every boardroom in Johannesburg has had the AI conversation. The slide decks are polished, the pilot budgets approved, and the ChatGPT licences issued. Yet when we audit production systems across mid-market South African businesses, fewer than one in five have moved beyond isolated experiments.
The gap isn't technical talent — it's clarity. Leaders are being sold AI as a product category when it's actually a capability layer that must attach to a specific business outcome.
Where AI is actually delivering value
The projects we see succeeding share three traits: a narrow use case, owned data, and a human-in-the-loop workflow. Document processing for finance teams, customer support triage for call centres, and sales forecasting for distributors are delivering measurable ROI within 90 days.
What fails? Generic 'AI strategy' engagements with no delivery path, chatbots trained on outdated knowledge bases, and models deployed without monitoring for drift or hallucination.
A practical starting point for SA businesses
Start with one workflow where staff already spend hours on repetitive cognitive work. Map the inputs, outputs, and approval steps. Only then evaluate whether a rules engine, classical ML, or an LLM with RAG is the right tool.
POPIA compliance must be designed in from day one — not bolted on after procurement. That means data minimisation, purpose limitation, and clear retention policies for any training or inference pipeline.
2026 is the year AI stops being a slide and starts being a system. The organisations that win will be those that treat it as engineering, not marketing.
Written by Phillip Malele
Khemo IT Solutions