As artificial intelligence evolves from simple assistants to autonomous agents capable of making complex real-world decisions, South African financial institutions are facing both enormous operational opportunities and sophisticated new fraud threats.
For most people, using AI boils down to help with calculations, document editing, or image generation. However, the future looks different as more individuals and organizations begin to adopt and rely on agentic AI.
The essence of agentic AI lies in trusting AI to make complex and significant decisions on your behalf. It goes far beyond drafting emails or answering simple queries; instead, agentic AI takes on much more complex tasks, such as independently managing investment strategies, executing multi-stage business operations, or booking complex travel itineraries.
A few weeks ago, the author attended a roundtable discussion focused on bridging the trust gap in AI agents and exploring how they are actively changing banking, payments, and business workflows. The event was organized by Sumsub—an AI-powered trust infrastructure provider that combines identity verification, fraud prevention, transaction monitoring, and risk management to optimize compliance efficiency.
Under the guidance of Mick Amelishko, an AI advocate at Sumsub, the discussion revealed the potential of agentic AI for South Africa as local adoption steadily grows. Industry insights were also shared by Hannes Bezuidenhout, Vice President of Sales for Africa at Sumsub, and Jarrid Jensen, Regional Director for South Africa.
Amelishko emphasized: 'The biggest shift towards an 'agentic' way of working is agency. Last year, AI was just an assistant—it helped check writing or suggest basic ideas. Now, AI can access the outside world, read the live internet, interact with other digital products, make independent purchases, register on websites, or directly interact with bank accounts with permission.'
Amelishko continued, explaining that for centuries, society functioned on the assumption that every transaction or action was performed by a human. Today, this human element is gradually being replaced.
For many, this rapid transition causes legitimate concern. Nevertheless, with proper education and a clear understanding of how to regulate agentic AI, these concerns can be largely mitigated.
Despite these safeguards, the threat of AI-driven fraud remains a rapidly escalating problem. The situation is reminiscent of Kyle Reese's famous warning about the Terminator from the classic 1984 film: 'You can't negotiate with him. You can't bargain with him. He feels no pity, remorse, or fear. And he will never stop... until you're dead!'
While Reese described an unrelenting cybernetic killer, this quote vividly reflects modern concerns about autonomous digital threats—without physical violence, but with equal persistence in exploiting digital vulnerabilities.
Addressing the local situation, Jensen noted how corporate banking trends are developing in South Africa. He observed: 'The large-scale adoption of true AI agents in South Africa is still quite low compared to global markets. South Africa has a complex market with advanced digital banking and instant payment systems, and consumers are already very comfortable with digital transactions. However, full agentic adoption is still in its early stages.'
Adding this perspective, Bezuidenhout observed that while forward-thinking institutions like Discovery Bank are actively moving into the agent space, South Africa as a whole lags behind international leaders such as China.
One of the main issues associated with agentic AI concerns accountability: what happens if something goes wrong, and who bears the ultimate responsibility—the person who issued the initial request, or the machine that executed the action?
The prevailing consensus at the event was that the AI itself cannot bear legal liability for illegal activities; responsibility must always be traced back to the human operator or system controller. However, new methods of financial fraud are evolving rapidly, and the regulatory framework often lags behind these technological shifts.
Illustrating this growing threat, Sumsub data shows that while the global rate of identity fraud slightly decreased to 2.2% in 2025 (compared to 2.6% in 2024), cases of high-tech fraud increased by 180% year-on-year as attackers shifted from low-effort scams to advanced, AI-driven attacks.
Ultimately, as agentic AI transitions from a new experiment to core financial infrastructure in South Africa, balancing automated convenience with vigilant human oversight will be key to maintaining security and trust.