In early 2025, a client in the insurance sector came to me with what seemed like a simple request: automate claims processing. Two previous attempts had failed. When I opened the file, I quickly understood why… and the problem had nothing to do with the AI model that had been chosen.
That project reinforced a conviction that has guided my work ever since: the impact of agentic AI is not measured by the sophistication of the model. It is measured by the ability to automate the right process, the quality of the data, and the rigour of the system built around it. But before we get there, let’s clarify what we’re talking about, because the term “agentic AI” is among the most overused of the moment.
Agentic AI is a system capable of pursuing an objective by coordinating reasoning, tools, data, and actions within a given environment, with a certain degree of autonomy.
For a system to truly qualify as agentic AI, five elements must be present:
• An explicit objective and a defined scope of action
• Orchestration capabilities: planning, sequencing, and adjusting
• Access to tools and data, not just text-based content
• Context memory across steps
• Verification loops and escalation mechanisms
It is neither an enhanced chatbot, nor an automated process with a few prompts added to it, nor a “free” agent operating without a control framework. With agentic AI, we move from a response-based logic to assisted execution, across compound tasks performed in sequence or in parallel.
.png)
The Three Main Families of AI: Traditional, Generative, and Agentic AI
In successful deployments, the agent absorbs the repetitive and structured part of the work. Teams can then focus on decision-making, client relationships, and non-standard cases. The profession does not disappear; it is reconfigured.
The value comes from the architecture: tools, data, controls, observability; not the model. Most enterprise use cases can be deployed using available models, within governed cloud environments.
RAG responds. Agentic AI executes. The difference lies in the ability to chain actions, use transactional tools, and maintain state between steps, rather than in the quality of document retrieval alone.
Autonomy is not an end in itself; it is a level to be adjusted based on the use case. On a critical process, execution with human validation is often preferable to a fully autonomous action, even if the technology would allow it.
Three forces are converging and making this a decisive moment. First, today’s models are better at reasoning, using tools, and integrating into enterprise environments than they were a few years ago. Second, organizations are looking to automate entire workflows, not just accelerate isolated tasks. Finally, the operational question has changed: it is no longer simply “does it work?”, but rather what level of autonomy is acceptable, for which process, and with what safeguards.
According to Gartner, more than 33% of new enterprise applications will include autonomous agents by 2028, compared with less than 1% in 2024. This is not a wave to anticipate; it is a transition already underway across several sectors.
Agentic AI is not simply a passing trend. It is a new layer of work orchestration that reshapes roles, processes, and controls. Organizations that clarify what it is — and what it is not — before deploying it give themselves a head start.
In the next article in the series: the six foundations to validate before discussing deployment. Because agentic AI leaves little room for shortcuts, and most failures begin there, well before the choice of model.
