South African financial institutions may be aiming too low with AI

That is where the distinction between bolt-on AI and built-in AI matters. Bolt-on AI adds new capability to old structures, helping an existing process move faster, cheaper or with less human effort. Built-in AI starts with a harder question: if AI were part of the design from the outset, would this process, service or operating model still look like this? The first leaves the model largely intact; the second begins to rework it.

That strategic choice is now coming into view for South African financial institutions. In our survey work across 12 local banks and insurers, there was no shortage of activity, including experimentation, investment and, in some cases, real momentum. But much of it still sits comfortably inside the current model. Firms were applying AI to recognised parts of the business rather than challenging how those parts should work

This is where the debate needs to mature. Leadership teams that continue to treat AI mainly as a portfolio of use cases are likely to get local wins: a better chatbot, a sharper fraud model, some productivity upside in support functions, faster drafting, more accurate triage and less rework. All of that is welcome, but it can also become a trap. The institution stays busy, the dashboard fills up and little changes in the underlying architecture of the business. That is the weakness of bolt-on AI. It can create enough visible progress to delay harder choices, leaving institutions to modernise fragments of a model that may already be ageing badly.

Senior leaders should be wary of AI programmes that look too tidy. If every initiative appears to succeed, the institution may be learning too little. Real reinvention usually involves experiments that disappoint, assumptions that prove wrong and designs that need reworking. The greater risk may be a programme that is safe enough to show progress, but too cautious to change the business.

Built-in AI is more demanding because it takes aim at the design of the work itself, forcing a rethink of how a journey, function or domain should operate end to end. It also raises harder questions about where judgement should sit in the process. This is not a technology programme dressed up as strategy; it is fundamentally business redesign.

South African incumbents have not been irrational in their cautious approach. Legacy estates are heavy. Data quality is uneven. Investment choices are constrained. Regulation is not getting lighter. A measured first phase made sense. The risk is that discipline hardens into under-ambition, especially when market signals begin to point in a more demanding direction.

Those signals are worth reading carefully, without overreacting. Offshore examples are useful as indicators of where economics and customer expectations can move, rather than as templates to copy.

Klarna, the Swedish buy-now-pay-later firm turned digital bank, for example, showed how quickly AI can shift service economics when it is pushed into the operating core rather than left at the edges. It also provides a useful caution. After leaning heavily into labour-replacement, the company later resumed hiring and shifted focus toward growth and service quality.

Lemonade, an American insurance-tech upstart, offers a different signal. It was built with automation embedded into claims, data capture and decisioning from the outset, illustrating what built-in AI can look like when it shapes the operating model rather than being added later.

None of these map neatly onto South Africa. That is not the point; service models, labour models and customer expectations are already moving.

The broader direction matters more than any single brand. EY-Parthenon’s recent US insurance research points to a shift from isolated experimentation toward scaled value, including cost impact and front-office change. EY’s own public Client Zero AI transformation programme reflects a similar move from scattered experimentation to a more deliberate, strategy-led model.

The implication for boards and executive teams is straightforward. They should now be asking harder questions than they were a year ago: which service models are too expensive, which processes still rely on human effort that adds little value, which product lines carry friction that customers will soon stop accepting, and which decisions could be made faster and closer to the customer with a different mix of human and machine judgement. They should also consider which parts of the control environment could be designed into the workflow rather than layered on afterwards, and which parts of the organisation are still being defended as strategic when they are really just inherited.

Boards and executive teams should have no illusions about the choice in front of them. Many institutions are about to spend meaningful money making legacy models marginally better. They will report progress, point to pilots and claim momentum. Some of that will be true; much of it may still miss the broader opportunity.

In summary

The more serious response is not to launch more experiments indiscriminately. It is to choose a small number of domains where AI could materially change economics, service or control, and treat them as redesign challenges rather than technology deployments. A smaller group of firms will do exactly that. They will use AI to rethink how work gets done, how decisions are made and where the business can be fundamentally simplified, removing friction others are still managing and building operating models that are cheaper to run, faster to adapt and harder to attack.

That gap may widen faster than many incumbents expect.