Dr. Christoph Nieuwoudt | The age of Agentic AI

AI will not replace you, but a person using AI may.” - this popular quote has captured the imagination of a world that witnessed the rise of generative AI. Indeed, tools like ChatGPT, Claude, Bard, and Copilot have shown extraordinary capabilities in reading, writing, and summarising - essentially mastering the traditional white-collar tasks of comprehension and communication.

However, despite these powerful capabilities, generative AI hasn’t fundamentally changed how most businesses operate yet. It has augmented the way we find information, write emails, or generate code snippets, but the way organisations operate remains largely unchanged.

Agentic AI refers to autonomous, goal-directed systems that can plan, act, reflect, and adapt - often in collaboration with other agents. Where generative AI is reactive, agentic AI is proactive. It doesn't just respond to prompts; it can take initiative and perform tasks. It can retrieve information, reason through multi-step tasks, monitor its own progress, and adjust its course of action. It represents a new class of digital worker, and it promises to be revolutionary.

Key Approaches in Agentic AI Implementation

As we scale this technology, several core design principles and key learnings have emerged that underpin effective agentic systems.

Unified Data Infrastructure

A successful agentic AI system depends on access to clean, structured, and unified data. Agents need timely, reliable inputs to make good decisions and to act autonomously. Integrating diverse data sources into a single, accessible layer enables agents to retrieve context-rich information and operate effectively across use cases.

At FNB we have made significant strides in building a data platform that collects data from close to 400 source systems, containing over 99% of what we deem to be key data, covering over ten million customers, billions of payments and receipts, trillions of Rands of advances, deposits, investments and sum-assured, billions of streaming interactions etc. While there is ongoing work on such ‘structured’ data, it is largely in place.  Ironically, ‘unstructured’ data covering our products and solutions, internal policies, processes etc. was not as centralised and is requiring a lot of effort to pull together.

Reasoning Architecture

Unlike earlier automation tools, agentic AI relies heavily on reasoning. This means designing control architectures that support logic-based decision-making, contextual memory, and inference across tasks.  Agents developed using chain-of-thought strategies can decompose challenges and learn from missteps, greatly increasing accuracy and achieving ongoing improvement.

At FNB, we’ve found that a simple thing like using a reflection agent to evaluate the output of a primary agent and provide feedback for refinement, greatly improves accuracy.  This approach is particularly effective in complex interactions - such as those handled by our eBucks bot - where multiple factors influence the final response. More generally, optimisation and re-enforcement learning will re-shape how agents form a view of their context and learn on an ongoing basis.

Federated AI Agents

One of our earliest learnings on agents was to focus initially on narrow use cases that deliver value before repeating or scaling the approach wider.  For example, judgemental credit is a huge process in the bank but is incredibly nuanced depending on the client sector and many other factors.  Building an agent to help with compiling credit applications in agriculture and even a single commodity (maize) was easier and allowed learnings to be transferred to further areas. 

More generally though, rather than deploying monolithic systems, federated AI involves specialised agents working in concert. Each agent focuses on a narrow domain, but together they form a distributed intelligence capable of addressing complex tasks. Multi-Agent Collaboration Protocol (MCP) is a major new development that helps orchestrate this federation, assigning tasks, sharing context, consolidating outputs, and enabling cross-agent learning.

API Integration

One of the most significant constraints on scaling agentic AI is not model performance, but Application Programming Interface (API) availability.  APIs are the bridges that allow software including agents to interact with business systems in real-time — a key difference from earlier approaches like RPA, which relied heavily on screen scraping and brittle user interface automation. Modern agents need robust, well-documented APIs to perform tasks reliably. In FNB we have an internal API standard and registry and tools to assist with documentation and discovery. For multi-agent environments, shared API contexts and metadata standards are essential to ensure consistent invocation and interpretation.  Building a scalable API fabric is often the prerequisite for effective agent deployment.

Responsible AI

For Agentic AI to thrive in financial services - and indeed in any organisation, industry or sector - it must be deployed responsibly.  The concerns around data privacy, explainability and transparency, bias and fairness and regulatory compliance are all valid. But they are also manageable provided financial institutions deploy Agentic AI responsibly and in a controlled way. 

We have implemented rigorous governance via amongst other things, a dedicated AI Technical committee (for model and use case approval), an AI architecture and AI infrastructure standards (that includes guardrail components, vendor, tool and cloud including region assessment and approval) and a senior AI Steerco (driving business oversight).  But the most important point is perhaps the next one- human collaboration and oversight.

Human-AI Collaboration

Last, but definitely not least, Agentic AI does not replace human expertise - it augments it.  At FNB, we design systems with a human in the loop philosophy, where people remain in control or at very least oversee critical decision-making.  Agents will take on the repetitive and procedural load, allowing staff to focus on creative problem-solving and high-value customer interactions.

Incorporating feedback - from users, outcomes, or other agents - is critical to driving continuous improvement. This includes mechanisms for reflection, self-evaluation, and supervised fine-tuning.

These foundational approaches are not only technical enablers; they shape how we think about AI as a collaborative partner in real-world systems.

Agentic AI Applications

At FNB, we are deeply invested in exploring the promise of agentic AI and we have over 100 people across data science, engineering, and business domains now working on agentic AI initiatives. To date, we have over 80 AI agents deployed or in development and multiples of this under consideration. 

Our agentic AI portfolio spans three primary categories:

Customer Experience: Agents that proactively assist customers, personalise interactions, resolve issues, and operate across channels.  Many of these agents are in beta testing before deployment to customers later in 2025.
Employee Empowerment: Tools that help employees find almost any information in the group with references to trusted internal documents, basic human capital assistance including things like leave, help with coding tasks as well as agents embedded in email, meetings and a range of general staff activities.
Operational Effectiveness: A range of agents developed for specific purposes with API integrations etc. for them to perform their tasks.  An example is agents for credit - from pre-screening to financial analysis, credit application compilation etc.  Other examples are our agents for financial crime – covering everything from customer and transaction risk assessment, uncovering various types of fraud and writing reports to submit to regulators, some of which we have been doing using AI for many years already.

A Transformative Shift for the Knowledge Economy

We are still early in this journey. While the immediate impact of agentic AI is incremental, the long-term potential is exponential.  Just as the internet transformed communication and cloud computing transformed infrastructure, agentic AI is poised to transform how knowledge work gets done. It’s not just a new way to build software.  It’s a new way to operate and at FNB, we believe we are just getting started.

*Dr. Christoph Nieuwoudt, Chief Data and Analytics Officer at FNB.