We cannot afford three years of unmeasured AI spend. That is our advantage.

There is a story we tell ourselves on this continent about leapfrogging, and I want to be careful with it, because it is half true and the half that is false does real damage.

The true half is that arriving late to a technology means arriving after someone else has paid for the expensive mistakes. Mobile money did not have to unwind a century of branch infrastructure. Payments did not have to migrate a cheque-clearing system.

The false half is the implication that this happens automatically. It does not. Leapfrogging is a choice made under constraint, and the constraint does not make the choice for you. Plenty of markets have arrived late and simply reproduced the incumbent's mistakes on cheaper hardware.

So let me put the AI version of the argument precisely, because I think it holds, and then let me put the case against it, because I think that holds too.

The argument

Gartner expects roughly 40 per cent of agentic AI projects to be cancelled by the end of 2027, on grounds of cost escalation, unclear business value and inadequate risk controls.

Read that as a statement about capital tolerance rather than technology. A large European bank or a US insurer can absorb three years of unmeasured AI spend. It is uncomfortable, someone loses a job over it, and the institution continues. The write-off is survivable because the balance sheet is large relative to the experiment.

Most South African and African institutions cannot do that. Not because we are less sophisticated, but because the ratio is different. An unmeasured three-year programme here is not an embarrassing line item. It is the AI budget, and there is no second one.

That constraint has a specific consequence. If you cannot afford to discover in year three that you cannot prove anything, you have to build the measurement in at the first agent instead of the fiftieth. Baseline capture before the pilot. Cost per solved task from day one. Human-touch rate reported alongside it.

Those are exactly the practices that the cancellation statistics say are missing. We are being pushed toward them by scarcity rather than by wisdom, and the outcome is the same either way.

There is a similar argument on skills. We cannot build data centres faster than anyone else, and pretending otherwise has consumed a lot of policy attention. We can build fluency faster than anyone can build data centres, because it requires teachers and practice rather than capital and power. That is a genuine comparative advantage and it is chronically under-invested in.

The case against, which deserves an honest hearing

I have made this argument to rooms of executives and I have had it pushed back on well. The strongest objections are these.

Constraint often produces caution, not discipline. A scarce budget can just as easily produce a small, timid, unmeasured pilot as a small, rigorous, well-instrumented one. There is nothing in scarcity that selects for measurement. It selects for cheapness, and cheapness frequently means skipping the evaluation harness, which is the expensive part that produces no demo.

We have infrastructure constraints the argument ignores. Load shedding, connectivity cost, latency to the nearest region, data residency requirements and foreign exchange exposure on consumption-priced services are all real, and none of them are solved by better measurement discipline.

Skills leave. Building fluency faster is only an advantage if the fluent people stay, and a lot of them do not. Any argument about human capital as comparative advantage has to survive contact with emigration data.

Late arrival can mean structural dependence. Leapfrogging past your own infrastructure sometimes means leapfrogging into somebody else's. If the models, the platforms and the compute are all foreign, arriving late may mean arriving as a customer rather than a participant.

I think the argument survives all four, but only conditionally, and the condition is worth stating plainly.

The condition

Constraint becomes advantage only when it is converted into a deliberate policy of measuring first. Otherwise it is simply a smaller budget.

That conversion is a decision somebody has to make, and it is not the technology team's decision. It is an executive one. It sounds like this: we will not fund an agent until the workflow it touches has a captured baseline, and we will report cost per solved task and human-touch rate alongside every AI initiative from the first quarter.

That policy costs nothing. It slows down the first project by about a week. And it is the difference between a constraint that produces discipline and a constraint that produces a small failure instead of a large one.

What this means for the region

South Africa has chosen a sector-specific, multi-regulator approach to AI oversight rather than a single AI Act, with sector instruments expected from 2027/28. That timing coincides almost exactly with the window in which most enterprise roadmaps promise autonomous, cross-organisational agents.

I read that as an opportunity rather than a threat, for a specific reason. Institutions that already measure will find supervisory expectations easy to meet, because the artefacts a regulator wants, a named owner, a decision record, validation evidence, are the same artefacts a well-run programme produces anyway for its own reasons.

Institutions that do not measure will experience the same requirements as a compliance burden arriving on top of a programme they already cannot justify. Same regulation, two entirely different experiences, determined by a decision taken years earlier about whether to capture a baseline.

The organisations in this region that build measurement in now will be shaping those sector instruments through consultation rather than absorbing them through enforcement. That is a narrow window and it is open at the moment.

The honest version of the leapfrog claim

Every technology this continent has adopted late, we have had the opportunity to adopt better, because we arrived after the expensive mistakes had already been made by someone else with more money.

We have sometimes taken that opportunity and sometimes not. The difference has never been the technology. It has been whether someone senior decided, early and explicitly, to do the unglamorous thing first.

With agentic AI the unglamorous thing is measurement, and the window for deciding it is now rather than in 2028.

Let us do it on purpose this time.