Paradoxes of scaling AI in South African financial services

South African financial institutions are past the “what is AI?” phase. Money is committed, teams are mobilised, and deployments are starting to show up in the wild. Every surveyed bank had already rolled out or soft‑launched GenAI. Similarly, most surveyed insurers had moved beyond experimentation into pilot, beta or launch‑ready releases. That is meaningful momentum, but it is not yet scale.

Across banking and insurance, the same issues keep resurfacing: weak data foundations, limited access to usable data, difficult integration into legacy environments, and the challenge of embedding AI into day-to-day workflows. In banking, data quality and legacy integration emerged as the top two blockers. Insurance pointed in much the same direction. These are not minor technical frustrations. They are the structural reasons why promising use cases fail to become a repeatable capability.

That leaves many firms in an awkward middle ground. AI is visible enough to sound reassuring. Boards hear updates. Executives can point to pilots and early wins. But the underlying foundations are still too fragile to support consistent enterprise-wide scale.

The US comparison matters here only as a reminder of where this goes next. In a more scaled market, the firms that pull ahead are not the ones with the most pilots. They are the ones that fix data, integration and architecture early enough for value to compound. If South African institutions do not treat those foundations as a strategic priority, AI will remain active but uneven.

There is no shortage of activity in the South African data. Institutions are experimenting, piloting, and expanding their use-case portfolios. On the surface, that looks encouraging.

What is harder to see is whether organisations are learning with the same discipline they use when launching. That is where the picture starts to look less mature. There is far less evidence of hard-edged value measures, explicit stop rules, or a strong institutional habit of shutting down weak ideas and reusing what has been learned.

This is where the US comparison becomes useful. US institutions appear further along not only because they are larger and better funded, but because many seem more comfortable absorbing controlled failure as part of capability building. In South Africa, by contrast, the number of reported visible failures is strikingly low relative to the amount of activity underway. In banking, respondents reported virtually no failures. Insurance showed the same broad pattern: plenty of activity, very little failure being surfaced. Some of this may be due to timing, and some to definitional issues. However, there are likely to be elements of avoidance, where financial institutions are more comfortable launching pilots than retiring them, or where their ambitions are set too low.

That is a problem. A healthy AI portfolio should produce some dead ends. What matters is whether the institution learns quickly enough to stop weak ideas, redirect investment and build on what works. Without that discipline, innovation generates noise rather than momentum. In practice, this often shows up as pilots that remain isolated within functions, with little reuse across the organisation.

Across both banking and insurance, the current AI agenda still leans heavily toward productivity, efficiency and operational use cases. That is sensible. Claims, underwriting, servicing, fraud, risk operations, and internal productivity are all rational starting points. They are measurable, easier to justify, and usually lower risk.

The survey data makes that bias clear. In banking, use cases clustered heavily in the back office, often around 40% to 50% of the portfolio, while front-office use cases tended to sit closer to 10% to 20%. In insurance, the centre of gravity was claims, underwriting, and servicing, with distribution notably less prominent. In other words, firms are mostly using AI to improve how the machine runs, not yet to rethink how the business competes. But that is still only a starting point.

The bigger strategic opportunity lies elsewhere: in changing how decisions are made, how customers are served, how products are designed, and how value is created. That is where AI stops being a tool for optimisation and starts becoming a force that can reshape competitive advantage.

This is one of the clearest places where the US benchmark sharpens the picture. In a more scaled market, the conversation has moved further toward broader commercial integration and more advanced deployment. In South Africa, many firms still appear to be in the safer phase: using AI to improve the current model rather than challenge it.

That may be the right sequence in the short term. It becomes a risk if it lasts too long. Efficiency creates breathing room. It does not, on its own, prepare an institution for a market in which the basis of competition may be shifting.

The survey data suggests that governance is getting more attention, and that is the right instinct. In a regulated sector, AI without guardrails is not innovation; it is unmanaged risk. But governance only helps if it accelerates clear decisions. When it becomes a collection of committees and reviewers, it can slow delivery, dilute accountability and turn AI into a compliance exercise rather than a performance lever.

This is especially visible in banking. Half of the respondents said the primary AI decision-maker was a dedicated governance committee, while only a quarter pointed to executive leadership. That structure may feel safe, but it often means that no single executive is accountable for value, trade-offs, and delivery pace. The practical outcome is predictable: priorities drift, teams build in parallel, approvals take longer, and scale stalls quietly.

The “so what” is simple: AI at scale needs both control and cadence. Keep the guardrails strong but make decision rights explicit, empower accountable executives, and measure governance by outcomes (speed, reuse, risk incidents avoided), not by the number of forums.