Agentic AI: Navigating opportunity and risk
Jan de Villiers, Head of Cloud Academy at PBT Group
Agentic AI¹ refers to a class of artificial intelligence (AI) systems that can plan, reason, act, and execute multi-step tasks to achieve defined goals with limited, or in some cases no, direct human supervision. Unlike traditional AI systems that generate recommendations or respond to prompts, agentic systems are designed to move work forward autonomously.
In practical terms, this could mean an AI that continuously analyses security logs, detects any anomalous behaviour, isolates a compromised endpoint, initiates containment playbooks, and alerts the security team. It could also be a system that negotiates meeting times across participants, books resources, reschedules if time conflicts arise, and learns preferred patterns over time. The defining feature is not intelligence alone, but the ability to decide and act within a defined scope.
That ability is why agentic AI has become a prominent topic in boardrooms. If a system can interpret context, make decisions, and execute actions, the loop from idea to outcome appears to shrink. The attraction is clear: fewer hand-offs, faster responses, and reduced human intervention in routine processes.
However, this is also where clarity is essential.
“Agentic” is not a binary label. Instead, think of it as a spectrum. Capabilities such as autonomy, persistence, planning, memory, tool use, and goal-directed behaviour appear at different levels across systems. For instance, simple reactive assistants or context-aware helpers to goal-oriented agents. In practice, you can think about the degree of agency a system exhibits rather than a yes/no category.
A useful mental model for evaluating agentic AI is the OODA² (observe, orient, decide, act) loop, a decision-making cycle designed to support fast, effective decisions in changing situations. In simple terms, it describes how organisations take in information, interpret it, choose a course of action, execute, and then repeat the cycle. For a clear explanation, see: https://fs.blog/ooda-loop/.
Traditional analytics systems perform well in the first two stages: observing and orienting. They help organisations understand what is happening and what it might mean. Agentic systems aim to complete the loop by making decisions and executing actions automatically.
That shift is powerful, but it introduces new requirements. Completing the loop responsibly depends on measurable outcomes, validated decision logic, and mechanisms to detect unintended consequences. If an agent acts incorrectly, you must be able to trace why.
This is where the data foundations become decisive.
Despite some vendor narratives, you cannot simply “add data” and expect autonomous systems to produce reliable outcomes. Value still depends heavily on data quality, well-defined models, clear metadata, and governance structures. An agent that pulls the wrong KPI definition or acts on incomplete context will compound the error rather than eliminate it.
In other words, agentic AI does not remove the hard lessons learned over the past decade in data management. It raises the stakes.
Taking responsibility
Accountability is the crux. You cannot outsource it to an agent or a vendor. If an agent decides and acts inside your business, your organisation remains accountable for the result. That principle should guide where agents are allowed to operate, which controls are mandatory, and which decisions require human oversight. Aim to reduce human effort where it’s safe, not remove human judgement where it’s necessary.
There’s also a practical risk that’s easy to miss during experimentation: cost exposure and dependency. Today, much of AI remains a net cost for providers; pricing can change. If your workflows become tightly coupled to a single API or model family, a pricing shift or policy change can break your business case overnight. Explore, but be careful how integral you make any external service until you understand your long-term control points and switching costs.
Acting for change
So, what should data teams do today?
Start with controlled exploration. Build sandboxes where teams can learn how agentic loops behave with your data, under your constraints. Pair every “decide/act” capability with explicit guardrails. These provide the agent with the scope in which they may operate, the actions that should never be performed, and audit trails. Expect to keep humans in the loop in the near term, especially when decisions carry financial, regulatory, or safety impacts.
Secondly, reinforce the basics that make any AI workable: stewarded data, fit-for-purpose models, and clear metadata. If an agent pulls the wrong definition of a KPI, you’ll automate the error at speed. If you cannot trace a result back to its source, you cannot debug or explain it. Governance is an enabler that makes faster loops safe.
Third, design for measurement. Before an agent acts, define what success means and how you’ll detect variation, harm, or side effects. Treat telemetry, rollbacks, and kill switches as essential features. That mindset keeps you honest about whether an agent actually shortens your OODA loop or gives that impression.
Finally, maintain critical distance from the hype. Ask who benefits from the loudest claims. Validate with your own hands. The pace of change is remarkable, and there is real progress behind the headlines. But not every breakthrough will fit your environment.
References:
¹ https://www.databricks.com/glossary/agentic-ai#
