Most boards approving artificial intelligence (AI) investments in South Africa today could not tell you, from their own experience, what a capable AI model can do, or where it will fail.
They are signing off budgets, strategies and risk appetites for a technology they have met mostly in slide decks.
I say this after many years in boardrooms, and I say it without contempt, because the problem is not intelligence or diligence. It is exposure. And it is costing organisations real money.
On 11 October, Andrej Karpathy, a founding member
On 11 October, Andrej Karpathy, a founding member of OpenAI who later led AI at Tesla, posted some deliberately rough arithmetic about who actually understands AI. He described a funnel, and said it is widening.
At the top sit about six billion people, three-quarters of humanity, who have barely touched a large language model, the technology behind tools such as ChatGPT.
Below them are one billion to two billion casual users on free tiers, who treat AI as a better search engine or an e-mail polisher.
Everyone is using it. Hardly anyone is banking it.
Come roughly 20 million professionals using frontier models,
Then come roughly 20 million professionals using frontier models, the most capable ones available, for serious coding and mathematics, where projects that once took months now take days. At the tip are around 5 000 people with internal access to systems the public has not yet seen.
Add the top two layers together and you have roughly 95% of the world. My honest observation is that most directors and executives sit squarely inside it. Many have formed their view of AI from an occasional free chatbot. Very few have watched a capable model take a messy, real business problem and either solve it or fail at it in an instructive way.
The research supports the observation. In Deloitte’s 2025 survey of nearly 700 board members and executives across 56 countries, 66% said their boards had limited or no knowledge or experience of AI, and almost a third said AI was not on the board agenda at all.
In August 2025, MIT’s NANDA initiative published
In August 2025, MIT’s NANDA initiative published The GenAI Divide, which found that about 95% of generative AI pilots delivered little or no measurable effect on profit and loss.
A month later, the Boston Consulting Group reported that 60% of companies were getting hardly any material value from AI despite substantial investment, while only 5%, the ones it calls future-built, achieved five times the revenue gains and three times the cost reductions of everyone else.
In August this year, McKinsey found that nearly nine in 10 organisations now use AI regularly, yet only 37% attribute any impact on EBIT (earnings before interest and tax, a core measure of operating profit) to it, essentially unchanged from the year before.
Everyone is using
Everyone is using it. Hardly anyone is banking it. That the failure rate matches the top of Karpathy’s funnel is a coincidence of arithmetic, but I doubt it is a coincidence of cause.
MIT’s researchers called the problem a learning gap, and its boardroom version shows up in predictable ways.
A director who has never seen a capable model at work cannot tell a genuine use case from a polished demonstration. Boards overestimate AI in the strategy session and underestimate it in the risk register.
Budgets follow what looks impressive: MIT found
Budgets follow what looks impressive: MIT found that more than half of generative AI spend goes to sales and marketing tools, while the largest returns came from unglamorous back-office automation.
Organisations insist on building their own tools, when buying from specialist vendors succeeded roughly three times as often.
The remedy is not to turn directors into data scientists. It is to move them far enough down the funnel to judge well.
First, hands-on time
First, hands-on time. Give every director and executive access to a frontier model and a real task from their own portfolio: a board pack to critique, a contract to summarise, a strategic assumption to test. An hour of genuine use teaches more than a day of presentations.
Second, measurable outcomes before money moves. No pilot should start without an agreed baseline, a target, and a date on which the board will see the result and decide whether to scale or stop.
Third, bring the 20 million into the room. Somewhere in your organisation is a developer or analyst who uses these tools seriously every day, and who probably knows more about practical AI than your entire exco. Invite them into board and exco discussions, and listen.
Fourth, make it governance, not enthusiasm
Fourth, make it governance, not enthusiasm. King V, which applies to financial years beginning on or after 1 January 2026, expects the governing body to govern data, information and technology, including emerging technologies such as AI, and, in the words of the Institute of Directors in South Africa, to be technologically- and not just financially-literate.
ISO/IEC 42001, the international standard for AI management systems, places leadership commitment squarely in clause five.
Add AI competence to the board skills matrix, and treat any gap as a governance finding with an owner and a deadline.
The organisations in that top 5% are not
The organisations in that top 5% are not using different tools. They are making better decisions about the same ones, and those decisions start at the top.
If the people approving your AI strategy have only ever seen AI from the wide end of the funnel, your returns will most likely reflect it.
So, I put the question directly to the directors and executives reading this: which tier of the funnel does your board sit in, and how do you know?
