The financial sector has always been technology intensive. Electronic trading, algorithmic credit scoring, cloud-based core banking, and robo-advice each affected how financial services are delivered without fundamentally changing who delivers them. The human decision-maker remained at the centre. Technology accelerated the execution of decisions; it did not substitute for them.
Agentic AI is different. Not in degree, but in kind.
Agentic AI triggers a re-organisation of the relationships between institutions, individuals, and systems that goes well beyond operational efficiency and, in the process, has a significant impact on governance.
As the Financial Stability Board (FSB), International Organization of Securities Commissions (IOSCO), and the Organisation for Economic Co-operation and Development (OECD) have pointed out in their respective reports earlier this year, emerging AI systems raise governance questions that existing arrangements were not designed to address fully.
What do these questions mean for the South African financial sector and for a regulator such as the Financial Sector Conduct Authority? This is what this article, the first in a series of three, aims to unpack.
AI agent vs agentic AI: A distinction worth making
The term “agentic AI” is often used interchangeably with “AI agents”, and the conflation obscures what is changing. The OECD’s working paper, The Agentic AI Landscape and Its Conceptual Foundations, draws a clear and consequential distinction. An AI agent is a system that perceives and acts on its environment with a degree of autonomy, using tools to achieve specific goals: a fraud detection system flagging anomalous transactions, or a document tool extracting data from loan applications. Powerful, but task-specific, with a scope of action bounded in advance by the humans who deploy it.
Agentic AI describes systems composed of multiple co-ordinated AI agents that can decompose complex goals into constituent tasks, assign those tasks across agents, adapt as new information emerges, and pursue objectives over extended time horizons, all with limited human supervision at the point of action.
IOSCO has similarly observed the rapid emergence of AI systems that can take actions autonomously, with potential real-world consequences, on behalf of a user, with little or no human intervention. In its report, Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, IOSCO reiterates that agentic AI is not simply a sophisticated agent. It is a system that can plan, delegate, and adapt in ways its deployers may not have specifically instructed and cannot always reconstruct after the fact. The significance of this distinction is not technical. It is a governance distinction.
A socio-technical paradigm
The OECD’s description of agentic AI as a “socio-technical paradigm” is an important conceptual contribution in its working paper, and it deserves to be unpacked rather than cited in passing. “Socio-technical” is not a synonym for “widely adopted”. It is a specific claim about the relationship between technology and the social arrangements in which it is embedded: technological systems and social systems co-evolve, and when a new technology is introduced, it does not simply slot into existing relationships and practices. It reconfigures them.
Every previous wave of financial technology was socio-technical in this sense. Electronic trading restructured price formation; credit scoring introduced intermediating logic between lender and borrower; robo-advice changed who could access financial planning and on what terms. Each shift was significant, and each generated regulatory adaptation. But in each case the fundamental governance architecture held: a licensed institution, staffed by human beings, made decisions. Technology shaped and accelerated those decisions, but the decision, and the accountability attached to it, remained with the institution authorised to conduct the relevant financial service.
The OECD’s insight is that agentic AI challenges this architecture at its foundation. Agentic AI systems are not isolated technical artefacts. Their value lies not only in autonomous action but in the interaction between multiple AI agents, humans, and institutional processes: a relational web the OECD describes as essential to understanding what agentic AI is. Their co-ordination operates across contexts that no single human supervisor has fully specified in advance. A tool belongs to its user. An agentic system participates in relationships that its deployers initiated but do not fully control at the moment of operation. And uptake is accelerating even as the OECD flags unresolved gaps in security, privacy, and trustworthiness. Governance frameworks across jurisdictions are still being calibrated to a shift that is already well under way.
Platform with decisional capacity
Enter the working concept of “platform with decisional capacity”. It helps in describing what agentic AI does in financial services, rather than what it technically is. Describing a platform as having decisional capacity does not imply legal personhood or regulatory responsibility; it merely describes where decision formation increasingly occurs. A platform with decisional capacity makes consequential choices, on behalf of others, through processes that the deploying institution did not specifically instruct at the level of the individual decision and may not be able to reconstruct after the fact.
The choices are not arbitrary: they are goal-directed, informed by data, and calibrated to objectives. But they are emergent. The specific decision pathway that produced a given outcome was not pre-specified in the way a rule-based system produces a deterministic output.
Why “platform with decisional capacity”? Because the concern it captures is already visible in the international evidence. IOSCO’s survey work found that providers could seek to disclaim liability for harm arising from AI systems or shift responsibility to others in the AI supply chain, that identifying and holding accountable responsible persons may prove challenging, and that most technology providers sit outside the perimeter of securities regulation altogether.
The FSB has identified the complexity and limited explainability of some AI methods among key vulnerabilities. These observations point to the same underlying phenomenon: decisional capacity is migrating to systems, and the governance arrangements built around human decision-makers are being asked questions they were not designed to answer.
Previous generations of AI changed how decisions were made. Agentic AI changes what makes them. From a distance the distinction appears subtle. Up close, it is the difference between a tool and an actor.
Who decides? And who answers for it?
Internationally, supervisory tooling is following IOSCO’s May 2026 Supervisory Toolkit, which organises supervisory considerations around governance and accountability, model development, monitoring, third-party providers, market conduct risks, and operational resilience, and applies across AI system types including agentic AI.
South Africa is not insulated from the “decisional” shift. The FSCA and the Prudential Authority’s joint work on AI has documented substantial and growing adoption across banks, insurers, and market intermediaries.
Understanding the socio-technical paradigm is the precondition for a coherent response. If we misread the nature of the shift, treating agentic AI as merely a more powerful version of existing tools, we will design governance responses calibrated to the wrong problem.
The next article in this series turns from what agentic AI is to what it does to accountability. When a platform with decisional capacity makes a consequential financial decision, about a credit limit, a risk classification, a product recommendation, or a claims outcome, the questions that follow are not merely technical. They are questions that every conduct regulator must gear up to answer: who decides what? And who answers for it?
Nolwazi Hlophe is senior specialist: fintech at the FSCA.
Disclaimer: The views expressed in this article are those of the writer and are not necessarily shared by Moonstone Information Refinery or its sister companies.
