South Africa uses AI more than any other African country. It does not yet have enough people who can build, ground and run it. The scarcest skills are also the least glamorous.
Every African AI strategy assumes the models will be imported and the value will be local. That holds only if someone local can wire a model into company data, deploy it and explain what it did. The skills behind that, retrieval, agents and deployment, are the difference between an AI strategy and an AI pilot that quietly dies.
I keep meeting South African executives who can tell me, in detail, which AI tools their staff use. Far fewer can tell me who in the organisation would build, host and fix an AI system if the vendor disappeared tomorrow. That gap is the story of 2026.
By Microsoft's measure, 23.1% of South Africa's working-age population used generative AI in the first quarter of the year, up from 19.3% a year earlier and well ahead of Nigeria and Ghana at 10.1% and Kenya at 8.7%.
Using AI, however, is not the same as engineering it. For the first time in almost a decade, Xpatweb's Critical Skills Survey has placed AI specialists among South Africa's 10 hardest-to-fill occupations, and its preliminary 2026 findings show 22% of employers struggling to recruit AI and emerging technology skills. Adoption has run ahead of the people needed to make it work.
We do not need more engines
I find one distinction useful when I explain AI careers to executives. Machine learning researchers build the engine: they train foundation models from scratch, publish papers and usually hold advanced degrees. AI engineers build the car: they take a model that already exists and wire it into a system that does useful work, by connecting it to data, giving it tools and memory, and wrapping it in guardrails.
Africa has spent years asking when it will build its own frontier engine, and that is the wrong first question. An earlier State of AI report found that Sub-Saharan Africa holds about 15% of the world's population but under 1% of AI patents and conference publications, and no skills programme closes that gap quickly.
AI adoption has run ahead of the people needed to make it work.
The car is a different matter. The engines are already available, many of them open source, and the value for an African bank, miner, retailer or municipality lies in the system built around them. It is also a skill set that can be learned by building, which matters in a country where, as Salesforce's Ursula Fear has pointed out, qualification frameworks run on five-year cycles while AI skills change in months.
Three tiers, read from the bottom
I think of the craft in three tiers, and I insist that the order matters, because people who jump straight to building agents end up relearning the basics. The foundation is unglamorous: Python fluent enough to read what your AI assistant writes, Git for sharing work, the command line and Linux, where most AI tooling runs, and APIs, the standard way software talks to software and to models.
Python, tellingly, leads demand in South African job listings, and local surveys suggest experience often counts for more than the qualification on the CV.
The second tier is where AI engineering becomes its own discipline: embeddings, which turn text into numbers so a computer can search by meaning; retrieval-augmented generation, or RAG, which fetches the right company documents and hands them to a model so it answers from facts instead of guesswork; and agents, systems that choose their own next step, call tools and check the results in a loop.
The third tier is shipping: containers and Kubernetes to deploy reliably, observability to see why a system decided what it did, and monitoring to keep cost and security in check.
The hiring data follows the same shape. Pnet figures reported by Daily Investor show AI-specific roles up 252% over six years and roles that simply require AI skills up 488%, led by AI software developers, data scientists, data engineers and machine learning engineers, with Gauteng accounting for 58% of vacancies.
LinkedIn's 2025 Emerging Jobs findings, as summarised by ALX, put AI engineer, machine learning engineer and data scientist among the five fastest-growing job titles in Nigeria, Kenya and South Africa.
RAG is the African business case
Of the three tiers, the middle one is where African organisations will win or lose first, because RAG answers the question every board asks: can the model work with our data without that data leaving our control? A bank's credit policy, a hospital's clinical guidelines, a municipality's by-laws and a mine's safety procedures are all documents a generic model has never seen.
RAG lets an engineer ground the model in them, and increasingly that can happen on African soil. Altron launched an AI factory in a Teraco data centre in October 2025, a global AI player of African heritage followed in March 2026 with what it describes as the continent's first NVIDIA-powered AI factory, and in August Stratos Lab announced an exascale AI cloud at Digital Parks Africa.
South Africa already hosts more data centres than any other African country, although commercial access to local GPU capacity is still maturing. The regulatory picture is thinner: AI here is governed mainly through the Protection of Personal Information Act, which was never designed for it, so the engineer, not the statute, often has to build in the safeguards.
Language adds a second local advantage. That same AI factory is meant to support models for African languages, including Swahili, Zulu and Afrikaans, Johannesburg's Lelapa AI builds for isiZulu, and a farmer in South Africa is already asking agricultural questions in Setswana through the AI-Farmer app. A RAG system that answers accurately in the language a customer actually speaks is something no imported chatbot can easily match.
The factory floor is leading the way
One global player in AI, with African roots, shows what leading in AI engineering looks like in practice. It has deployed what it describes as the continent's first NVIDIA-powered AI factory in South Africa, became Africa's first NVIDIA Cloud Partner in November 2025, and has committed up to US$720 million alongside NVIDIA to the build-out, with expansion planned for Nigeria, Kenya, Egypt and Morocco.
What I find noteworthy is not the hardware. It is the engineering layer on top: a multi-model exchange that lets developers access, build and deploy applications on leading AI models, running on infrastructure that keeps sensitive data in-country, and partnerships with the CSIR and the Zindi data science community that put that capacity in the hands of researchers and builders.
That is the car, not the engine, and it is the clearest local example I know of the move from consuming AI to engineering it. It also signals what South African engineers need next: the skill to orchestrate several models, route work between them sensibly, and keep every decision traceable.
The unglamorous skill Africa is missing
Here is the finding that should worry boards. In its 2026 talent mapping across 48 African countries, the GSM Association (GSMA) concludes that the continent over-indexes on linguistic AI while under-investing in evaluation, MLOps (the discipline of deploying and maintaining AI systems) and governance, the very disciplines that make AI deployable at scale. It sets a target of 200% growth in those specialists by 2030.
In my three-tier view, that is the third tier, and it is exactly where projects die. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and none of those is a failure of the model.
For an African organisation the exposure is sharper. Token bills arrive in dollars while revenue is earned in rands, so cost monitoring is not an engineering nicety but a treasury control. And an agent whose decisions cannot be traced is an agent that cannot be defended to a regulator, an auditor or a customer.
The scarcest skill in African AI is not building the model. It is keeping the system honest once it is running.
Brain drain or brain gain?
There is a hard commercial reality underneath all of this. Xpatweb cites a global ratio of AI talent demand to supply of 3.2 to 1, which means every capable South African engineer is also a global asset. The GSMA warns that brain drain to Europe and the Gulf is hollowing out senior expertise on the continent, and industry reports say global employers, including Amazon Web Services, are recruiting African data and AI engineers for remote roles.
Resisting that portability is a losing strategy. The better one is to offer work worth staying for: African language models, agriculture, health and financial inclusion problems that offshore employers rarely prioritise.
An engineer who builds a portfolio on those problems becomes more employable globally and more valuable locally, and the three portfolio projects I would point any aspiring engineer to, a RAG knowledge system, a data-querying agent and a deployed application, map neatly onto them.
Five actions for African business and education leaders
Hire for the car, not the engine. Most organisations do not need a research scientist. They need engineers who can ground a model in company data, give it tools and deploy it. Specify RAG, agent and deployment experience, and test for it with a practical build rather than a degree check.
Respect the order of the tiers. Fund Python, Git, Linux and API fundamentals through apprenticeships, bootcamps and internal academies before sending staff on agent courses. Skipping the base means paying to relearn it later.
Make the first project a RAG system on your own documents, hosted in-country. Choose a policy, product or compliance collection, keep the data under POPIA-aware controls on local infrastructure where it makes sense, and measure whether answers are correct and sourced. One build teaches embeddings, retrieval and governance together.
Fund evaluation, observability and governance before you scale agents. Treat traceability and cost monitoring as launch requirements. Budget for dollar-denominated token costs in rand, and make sure any agent decision can be explained after the fact.
Give local talent problems worth staying for. Partner with universities and innovation hubs to place learners on live projects in local languages, agriculture, health and financial inclusion. The GSMA's targets include half of AI students completing industry internships by 2028 and 100 000 practitioners trained through hubs by 2030, and business can help meet them rather than wait for them.
My question for every African business and education leader reading this: if your most capable AI engineer accepted an offshore offer next month, would anyone left be able to explain how your AI system reached its last decision, and fix it when it was wrong?
The models are already imported and the infrastructure is arriving. What is not yet certain is whether the people who run them will be here to run them.
* 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 is currently a global practice leader in the field of AI software factories and AI engineering. 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.

