Skip to content
prod
Industry NewsTechnology

Agentic AI is about to change how banks operate

By Elenjical

Agentic AI is about to change how banks operate

When FinTech Summit Africa 2026 convened in Sandton this summer, one topic dominated the agenda: agentic AI. Unlike conventional AI tools that analyse information or recommend an action, agents can carry out a process from beginning to end, with far less human intervention. For African financial services, that matters. Joint research by the Financial Sector Conduct Authority (FSCA) and the Prudential Authority shows banks are already the sector’s leading AI adopters, with 45% of AI-adopting banks planning to invest more than R30 million in the technology.

The attraction is easy to see. Institutions surveyed already expect AI to improve productivity, strengthen fraud detection and help them make better use of their data. Agentic systems extend that idea beyond individual tasks. An agent dealing with an operational exception, for example, could identify the problem, find the information needed to resolve it, update the relevant systems and move the process on. In a bank handling enormous volumes of routine work every day, that starts to change the economics of the operation.

“We are already seeing the first versions of this inside financial institutions,” says Zion van Wyk, analyst at capital markets technology consultancy Elenjical Solutions. “Fraud detection and automatic invoicing are obvious starting points because they are highly repeatable and already close to full automation. The interesting question now is how quickly the scope expands.”

Developments elsewhere in enterprise technology suggest that expansion could happen quickly. AWS Kiro, for example, is an agentic development environment that can take natural-language requirements, turn them into detailed specifications and then produce working code, documentation and tests. What is interesting is less the coding application itself than what it says about the technology: agents are beginning to move beyond answering questions and towards carrying out substantial pieces of work.

For banks, that raises the prospect of agents taking responsibility for larger parts of an operational process rather than simply helping an employee complete it. There is, however, a sizeable gap between what the technology may soon be able to do and what many institutions are ready to allow it to do.

The regulators’ own data makes that clear. Cybersecurity is cited as the most pressing AI risk by 37% of institutions, while only 30% report board-level oversight of their AI frameworks. Roughly one in five applies no explainability methods at all. Those shortcomings become much more important when AI is no longer simply offering advice.

If an agent is able to act, a bank needs to know exactly where that authority stops. Transaction limits, permissions and escalation points need to be designed into the process. Every action must also be traceable, with clear accountability if the system behaves unexpectedly. These questions will ultimately determine how much work banks are prepared to hand over.

Nor is this something that can be left solely to an AI or technology team. Agentic systems introduce new vulnerabilities, including prompt injection and manipulated data, while banks continue to operate within tightly controlled regulatory environments.

“You need people who understand the technology deeply enough to know where the vulnerabilities are, working alongside people who understand the regulatory obligations,” van Wyk adds. “That combination is what allows an institution to move from an interesting pilot to something it can actually trust in production.”

Elenjical Solutions has seen something similar in its benchmark transition work. South Africa’s move from JIBAR to ZARONIA has forced institutions to standardise processes and address underlying systems well before the new benchmark is fully embedded.

“The groundwork can be unglamorous,” van Wyk says. “But it is what makes implementation easier to maintain and monitor. Do that work early and you create the conditions to scale much faster when the technology is ready.”

Regulators are having to keep pace too. South Africa’s outcomes-based Joint Standard on cybersecurity and cyber resilience was an early marker. In April, Cabinet approved a Draft National AI Policy for public comment before it was subsequently withdrawn for reworking. In June, Cabinet confirmed that the revised approach would seek to establish national standards for the ethical use of AI.

For banks, greater clarity could be important. Once institutions know more precisely what regulators expect around governance and accountability, one of the barriers to putting agentic systems into production begins to fall away.

And financial services is particularly well suited to the technology. Banks already have a strong commercial case for AI and vast quantities of structured, process-driven work. As agentic systems mature, institutions that have cleaned up their data, modernised their architecture and understood where automation can safely take control will have a considerable head start. The question may soon be less whether banks will use agentic AI, and more how much of the business they are prepared to let it run.