An ontology defines what exists and how the pieces connect. In computer science it is a formal representation of concepts and relationships: which information is available, how that information links together and what conclusions follow from the links. I have spent years building knowledge systems on that basis. With AI assistance the precision of the queries and the practical use of the results have both grown, and that growth has opened possibilities the earlier systems did not have.
My own practice sits inside that shift. I use structure as a working system, a way to move from raw data to a decision, not as knowledge organised for its own sake. Another wiki that describes what exists helps nobody. A small change in emphasis, from description to tool, changes what the structure is for. Modern computing power makes that tool faster and more capable than it used to be.
Five principles
I recently connected an agentic AI system to an ontological structure I had built, and the exercise answered a question I had not set out to ask: what principles sit behind each approach. The principles behind a classical ontology-driven study and the principles behind agentic AI turn out, in my experience, to be the same five.
Decomposition means breaking a complex problem into its parts. I advised US companies on entering European markets for many years, and the decision was never one decision. It was twenty, all connected: regulation, distribution channels, pricing structures, logistics chains, target audiences, languages, competition and product variants. In an ontology I decompose the problem by hand and decide what counts as a part. An agent decomposes it on its own, and sometimes cuts the problem along different lines than I would have chosen.
Relationships means understanding how the parts connect. Regulation feeds into pricing structure. Pricing structure feeds into which distribution channels are realistic. That chain has a direct effect on which product goes to which target audience, and the tactical marketing that follows looks completely different depending on the audience and the product. What works in France does not work automatically in Germany; two separate markets need two separate marketing approaches. An ontology maps these links explicitly, so I can trace each one. An agent infers the links implicitly, and it sometimes finds connections I had missed.
Dependencies means knowing what has to come before what. I built a training system for dental filling materials for the 3M ESPE group, and the project taught me early that context works best as a fully separate layer, mapped apart from the rest of the structure. In an ontology I sequence the dependencies as designed steps. An agent works out its own order, and that order does not always match the sequence I would have planned.
Context means the result depends on the conditions. QALO, the American lifestyle jewellery maker, sold plastic rings with highly scaled customisation through engraving. Margins in physical retail were already thin, and the customisation had trained the target audience toward buying online. In the ontological value-chain model I built at the time, that shift became visible quickly. Years later I repeated the exercise with AI agents, and it was not immediately clear. The complex decision chain produced too much trial and error and took too long to reach a conclusion. On a simple mapping the agent reaches an answer fast. Beyond a certain complexity, it does not, at least not yet.
Traceability means knowing where a decision comes from. My own ontology is traceable end to end: I built it, every connection is known, and I can explain and sell what it produces. An agent’s result is often better than what I would have built myself, but the path to that result is not always traceable, and an untraceable result is hard to defend in front of a client.
Where the principles diverge
The five principles are identical across both, yet the results differ every time I run the agents. My ontological model is traceable, explainable and follows the plan I set for it. Agents mostly do what they want, unless I force them into a rigid scaffold, and once I do that it stops being agentic AI at all.
After countless attempts, my experience is this: the result with agents gets better, sometimes markedly better than anything I built myself, but it follows no plan I set in advance. The agent finds its own way there, and I do not always understand the route it took. Its strengths are real. It sees connections that escape me, tries paths I would not have thought of, works faster than I do and can run twenty variants before settling on one. The problem sits on the other side of that same strength: I cannot always explain the result, and what I cannot explain I cannot sell, because the uncertainty is too large for a client to accept. Companies need more certainty in their internal planning exactly when markets around them are less certain, and an unexplainable result cuts against that need.
What can be controlled
The catch sits in the question I started with, which system delivers the better result. Years of running both systems point somewhere else, to what I can control and what I cannot.
My own ontology I control fully. I can make any adjustment, I know every connection, I know where the structure is strong and where the gaps sit, and when something goes wrong I know exactly where to look. Agents I do not control. I can give them context and guardrails, but what they make of that context is different every time I run them. Watching a multi-agent system work live, the impression that sticks with me is of a pack of unchecked individualists, each pushing hard for what it thinks is the best result. I know that description is overstated the moment I write it down, but it is the impression that keeps confirming itself when I watch the system work.
One fact sits underneath the impression: the system does not deliver perfect results. Full automation goes wrong at least as often as a hand-built model does.
The question that actually remains is when agentic AI will deliver a perfect ontological model on its own, when the agent builds the structure I still build by hand today, faster and perhaps better than I do. I work with AI agents every day, and I do not yet have a convincing answer, because the agents I run still behave too wildly for that. Taming the algorithm is a task for the people who build it.