Every large South African enterprise has now run an AI pilot. Far fewer have an AI system in production. The gap is rarely the model — it is everything around the model that regulated organisations rightly demand before software touches customers or money.
Why pilots stall
The typical stalled pilot shares three traits. It was built on a data extract rather than a governed pipeline, so nobody can vouch for what the model would see in production. It has no owner in the operating model — IT considers it a business experiment, business considers it an IT system. And it answers a question nobody attached a rand value to, so when compliance raises the first hard question, there is no business case to spend against.
What the shipped systems have in common
First, they start narrow: one document type, one queue, one forecast. Narrow scope makes the risk assessment tractable and the value measurable. Second, they inherit existing governance instead of inventing new committees — the same change control, access management and audit logging that already governs core systems. POPIA work is done at the data layer once, not per-project. Third, they keep a human in the loop wherever a decision affects a customer, which converts an unanswerable question ('is the model always right?') into an operational one ('is the review queue staffed?').
The uncomfortable prerequisite
Most organisations discover mid-pilot that their real problem is data plumbing. If claims history lives in four systems with three customer identifiers, no model will reconcile it. The honest sequence is data foundations first, models second — which is less exciting to announce, but it is the difference between a demo and a system.
The practical takeaway: pick one measurable use-case, run it through your existing governance, and budget as much for the pipeline and the operating model as for the model itself. That is the profile of every AI system we have seen survive its first year.