JOHAN FOURIE | What SA’s next AI policy should say

Last week government withdrew its new Draft National AI Policy 16 days after its launch, having found it cited many fabricated academic references. It acknowledged the “embarrassing” lapse, stating AI tools were used to draw up the policy.

Though the withdrawal was an act of accountability worth acknowledging, it was not the only lapse. The draft also proposed the creation of eight new bodies, including an AI Insurance Superfund modelled on the rickety Road Accident Fund (RAF), without acknowledging the RAF model has saddled the state with R400bn in unfunded liabilities and desperately needs an overhaul.

The episode questions the state’s readiness to regulate the AI systems that are already being heavily used throughout the country. Today the Bureau for Economic Research (BER) releases a research note on what is a potentially better approach.

The note uses insights from a special Outlook and AI usage survey conducted among registered users of the BER website during the first quarter of 2026. Of the 358 respondents, 95% of managers and 91% of private individuals reported using AI in their working week.

The BER study echoes the findings of others. The Visa and Discovery Bank SpendTrend26 survey of 1,000 credit card-using South Africans earning more than R100,000 a year found 40% of respondents reported using AI weekly for purchase decisions and 43% to pay for an AI subscription.

According to the Anthropic Economic Index, South Africans racked up just over 4,000 Claude.ai conversations (0.41% of Claude’s global total) in the week of February 5 to 12 this year, placing South Africans in the same band for usage as several other middle-income countries on a per capita basis.

South Africa, like the rest of the world, is in the early years of the diffusion of a new general-purpose technology that, like electricity or the internal combustion engine, will eventually affect almost every sector. The gains are already vast and measurable

These surveys show AI use has become routine for a wide cross-section of South Africans. But it is the productivity impact that is the arguably the issue that matters most for policymakers.

In the BER survey 35% of managers/professionals reported AI had added 5% or more to their productivity over the past three years. More interesting is that 82% of them expect AI to add 5% or more to their productivity over the next three years.

The international evidence reflects the same J-shaped pattern of modest initial realised productivity gains coupled with large, expected gains as access deepens and firms reorganise around new AI tools.

South Africa, like the rest of the world, is in the early years of the diffusion of a new general-purpose technology that, like electricity or the internal combustion engine, will eventually affect almost every sector. The gains are already vast and measurable.

The policy question is not whether to permit AI adoption. It is whether the framework being designed in Pretoria will help that adoption broaden, deepen and become more productive, or whether it will impede it.

The withdrawal of the draft AI policy opens the door for the next draft to be built on stronger foundations. It should draw on the recent economics literature on AI by acknowledging:

• Productivity gains come from the spread of AI and practical investments that encourage its diffusion.
• Most of the welfare accrues to AI users — up to 12 times more than the providers of AI earn. Any policy designed against the value AI creates for its providers will be designed against the wrong baseline.
• AI safety regulation is vulnerable to capture by the firms it oversees.
• Autonomous AI systems need clear rules on identity, registration and accountability more than new ethics boards.
• Concentration among AI providers is a real issue, but the response is to ensure fair access to existing AI tools, not to create new licensing regimes.
• When regulators attempt to use AI tools to monitor AI-using firms but lack the capacity to keep pace, monitoring expands but doesn’t necessarily reduce the harm AI may have caused.

Creating an adoption-first programme

The evidence suggests any national AI policy should first be preoccupied with the fostering the right conditions for AI tools to spread — cheap bandwidth, last-mile connectivity and exchange-control rules that do not penalise software-IP transactions. Because without these complements AI generates limited gains.

Unfortunately the withdrawn draft completely misses this point by placing issues of supervision before those of diffusion. To avoid repeating this mistake the test for any future AI policy should be whether it empowers South Africans to put the defining technology of this era to work for their prosperity or stands between them and it.

Several instincts in the withdrawn draft should not be repeated in the next one. The withdrawn policy proposed the creation of eight new supervisory institutions. Even though one binding constraint on the South African economy is the state’s weak implementation capacity, the withdrawn policy’s response was to build more state; to create a wide architecture of supervision.

Instead of proposing eight new institutions the policy should halve this number and make use of existing bodies where possible

A more useful response, where a real market failure justifies state action, is to commit to a small number of targeted interventions and scale up only what works. Examples could include:

• an AI-assisted clinical decision-support system in a public hospital;
• an indigenous-language AI resource built jointly with a university; or
• an AI-assisted revenue administration tool at the South African Revenue Service.

None of these requires new legislation.

Instead of proposing eight new institutions the policy should halve this number and make use of existing bodies where possible. For instance, it should build a national measurement capability for AI adoption inside Statistics South Africa and use the tools the Competition Commission has to police market concentration concerns.

It should also specify any proposed new bodies’ rules of operation from the start to ensure they are resistant to regulatory capture given that the complexity of AI safety compliance gives incumbents an advantage in shaping the rules.

Lastly, any new policy should embed AI into every existing university and TVET programme, rather than try to create new AI-only qualifications. The problem is not that markets are failing to deliver education; it is that the state’s quality assurance institutions, designed for a slower world, will prevent the educational system from adapting at the speed required.

The next draft must keep what was right about the first (the desire for fairness and protection, for example) and let go of what was wrong. In so doing it must be guided by the understanding that its role is not to insert administrative friction between users and tools that are already producing measurable gains, but to ensure these gains translate into broad-based prosperity.

• Prof Fourie is chair of economics, history and policy at Stellenbosch University. The full research note can be found at www.ber.ac.za/growth