The head of innovation at one of South Africa's four largest retail banks will not speak on the record. Not because the project is sensitive — it is well past that stage — but because his competitors are watching, and the gap he is currently enjoying is measured in months. What he will confirm is that since the beginning of 2025, his bank has had autonomous AI agents running three distinct back-office processes end to end: credit application processing, disputed-transaction resolution, and a segment of regulatory compliance reporting. No human reviews the output unless the system escalates. The agents are not making recommendations. They are making decisions.
South Africa's enterprise AI market has moved, in roughly eighteen months, from a landscape dominated by chatbot deployments and co-pilot tools to one in which agentic systems capable of autonomous multi-step task completion are entering production at major institutions. A March 2026 survey by Johannesburg-based research firm Emerging Tech South Africa, which canvassed 240 large enterprises across financial services, retail, and insurance, found that sixty-three percent reported active agentic AI deployments, up from eleven percent in the equivalent 2024 survey. More strikingly, forty-four percent reported removing human review from at least one AI-managed process in the previous twelve months.
At Sanlam's Bellville campus, the innovation team spent eighteen months building what they call an agent layer above the insurer's existing claims infrastructure. The system handles the end-to-end processing of motor vehicle claims under R150,000, from first notification of loss through settlement. Claims that previously required an average of eleven touch-points across four departments now require two: an initial human intake and a final sign-off on edge cases flagged by the agent. Average settlement time has dropped from nineteen days to four. "We didn't automate the easy part," says Sanlam's head of claims technology, Pieter van der Merwe. "We automated the whole thing."
We didn't automate the easy part. We automated the whole thing.
The employment implications are arriving, and the industry is discussing them in the careful language of role evolution and workforce transition. Sanlam's claims operation in Bellville employed forty-seven people in the roles that the agent layer now handles. Van der Merwe says thirty-one of those employees have been redeployed to higher-complexity work within the same division. Sixteen are in transition. Advocacy organisations, including the South African Federation of Trade Unions, have raised formal objections with the Department of Employment and Labour, arguing that sector-wide adoption without collective bargaining violates the spirit of the Labour Relations Act. The deployments are continuing.
The technical infrastructure enabling this shift is worth examining. Several deployments described in the Emerging Tech survey run on domestic compute — Vodacom and MTN have both expanded their hyperscale data centre capacity in Johannesburg and the Western Cape since 2024, partly in anticipation of exactly this demand. Microsoft's Azure expansion in Johannesburg, completed in late 2024, added significant GPU capacity specifically oriented toward enterprise AI workloads. South African enterprises are, for the first time, able to build and operate agentic systems without routing sensitive financial data outside the country's borders — a legal and reputational constraint that had previously slowed adoption considerably.
The unnamed bank executive, signing off a conversation that lasted an hour in a glass-walled meeting room overlooking Sandton's construction cranes, returns to the reason he will not be quoted. The competitive advantage is real, but it is also temporary. Within two years, every major financial institution in the country will be at the same point. Differentiation will then shift to what the agents are built upon: the quality of the data, the trust architecture, the proprietary processes they have learned to automate. "The agent is a commodity," he says. "The institution it reflects is not." It is the most honest thing said in a room full of bankers all week.
