De l’automatisation à la cognition: L’évolution de l’IA agentique

Pendant des décennies, l’automatisation des entreprises a suivi une logique familière: définir le processus, coder les étapes et laisser le logiciel les exécuter de façon cohérente. Les scripts automatisaient les commandes, les runbooks cons cons ? (les runbooks) capturaient les procédures, l’orchestration connectait les tâches entre les systèmes, et l’AIOps a introduit l’apprentissage automatique dans le volume croissant de données opérationnelles. Chaque avancée rendait l’automatisation plus performante, mais les humains prenaient encore la plupart des décisions. Ils définissaient le processus ; le logiciel l’exécutait.

Agentic AI change cette relation. Au lieu de commencer avec un jeu d’instructions fixes, un agent peut partir d’un objectif et déterminer comment le poursuivre dans des limites définies. Cela nécessite plus que la génération d’une réponse. Le système doit interpréter le but, rassembler le contexte pertinent, évaluer les preuves, utiliser la mémoire, choisir parmi les outils ou les automatisations, et s’ajuster au fur et à mesure que les conditions évoluent. Anthropic établit une distinction similaire entre les flux de travail, où les modèles et les outils suivent des chemins de code prédéfinis, et les agents, où le modèle peut diriger dynamiquement son propre processus et l’utilisation des outils.

De l’exécution déterministe aux décisions adaptatives

None of this makes traditional automation obsolete. In fact, deterministic automation remains the right answer when the path from a known condition to an action is well understood. If a service fails a health check, restart it. If a request meets specific criteria, route it to the appropriate queue. If an interface crosses a threshold, open a ticket and run a diagnostic. These are exactly the kinds of repeatable actions automation handles well.

AIOps expanded that model by helping organizations interpret what was happening before an automation fired: correlating events, identifying anomalies, reducing noise and adding context. Agentic AI takes the next step by introducing judgment between the signal and the action. The question is no longer only, “Which workflow does this condition trigger?” It becomes, “Given the objective and the evidence we have now, what is the best next step?”

Consider a degrading service. An agent may need to examine telemetry, topology, recent changes, tickets, configuration data, historical incidents and business impact before deciding what to do. In one situation, the evidence may point to a familiar remediation, and in another, the agent may need to gather more information or run a read-only diagnostic. And when the potential impact is high or the evidence is inconclusive, the right decision may be to bring in a human. That ability to adapt the path based on what the system discovers is the meaningful shift.

Le contexte et la mémoire modifient la qualité de la décision

Agency without context is not very useful. The same symptom can mean something entirely different depending on topology, maintenance activity, recent changes, dependencies or customer impact. Experienced operators understand this instinctively; they rarely make consequential decisions from one isolated signal. They assemble a picture, separate observation from inference, and decide how much confidence the evidence deserves.

Memory adds another dimension. Google Cloud identifie la mémoire as a core component of agentic architectures because it helps agents maintain context and use information from prior interactions. In operations, the opportunity is broader than simply retrieving an old ticket. Useful operational memory can preserve what happened when similar conditions appeared, which hypothesis proved correct, what remediation was attempted, whether it worked or was rolled back, and what operators learned afterward. When that experience has been validated, it can become evidence for the next decision rather than knowledge that disappears into a postmortem or remains in an expert’s head.

There is an interesting parallel in cognitive science. Research on cognitive offloading examines how people use external tools and actions to reduce the mental demands of a task, while research on tool use has shown that repeated interaction with tools can alter how people represent their practical reach and ability to act. The analogy to enterprise AI should not be taken literally, but the underlying idea is useful: tools do more than save time. Over time, they can change how we approach a problem. The well-known research on London taxi drivers and navigation expertise offers another reminder that sustained experience can shape cognition itself. For enterprises, the design question is therefore not simply how much work an agent can take over, but whether the human-agent system becomes better at understanding and acting on complexity.

Les agents doivent orchestrer l’automatisation, pas la remplacer

This is also why I do not see agentic AI as a rejection of automation. Earlier in my career, I spent years building operations automation, and one lesson stayed with me: execution alone is not enough. Automation has to earn trust. Operators need confidence that the right action is being used for the right situation, with the right authority and enough evidence to understand why.

A practical agentic architecture builds on that investment. Existing scripts, workflows, APIs, orchestration platforms and domain-specific tools become capabilities the agent can call when appropriate. If the evidence supports a known remediation, the agent can use the proven automation rather than improvising a new action. If the evidence is incomplete, it can investigate further. The value comes from combining reliable execution with more adaptive decision-making around when and how that execution should occur.

Plus d’autonomie rend la gouvernance plus importante

The moment software can choose among actions; governance can no longer be something added after the fact. An enterprise agent needs to understand not just what it is technically capable of doing, but what it is authorized to do in a particular environment, at a particular level of risk, and under a particular identity or role. NIST’s AI Risk Management Framework reflects this broader view of governance, organizing AI risk management around govern, map, measure and manage, with governance operating across the other functions.

In practice, that means an agent should be able to show the evidence behind a recommendation, make clear what systems and data it used, expose uncertainty, respect approval requirements and leave an audit trail of what happened. Reversibility matters too. A read-only diagnostic and a production configuration change should not carry the same autonomy threshold.

Human oversight therefore becomes more precise rather than simply disappearing. The same agent might gather evidence autonomously, recommend a remediation, require approval before a higher-risk action, and then verify whether the approved action actually worked. The goal is not to declare humans universally “in” or “out” of the loop. It is to put human judgment at the points where it adds the most value.

Le véritable avantage réside dans ce qui se passe après l’action

Perhaps the most consequential difference appears after an action is taken. Traditional automation usually asks whether the workflow executed successfully. A cognitive system should ask a harder question: did the action produce the intended outcome, and what should we learn from what happened?

That changes the value of the feedback loop. A successful remediation can strengthen the evidence for a similar situation in the future. A failed hypothesis should make the system less likely to repeat the same reasoning. An operator correction can improve the next recommendation, while a rollback may reveal that confidence exceeded what the available evidence justified. Over time, the important asset is not simply a larger library of automations. It is the accumulated relationship among situations, evidence, decisions, actions and outcomes.

Seen this way, the evolution of enterprise automation is less about replacing one generation of technology with another than about adding new capabilities around what already works. Scripts encoded actions, runbooks captured procedures, orchestration connected them, and AIOps helped interpret increasingly complex operational signals. Agentic AI adds the ability to pursue objectives, reason across context and evidence, draw on memory, choose among available capabilities, act within governed boundaries and learn from the result.

The endpoint should not be unlimited autonomy. It should be accountable autonomy: systems that can recognize when to act, when to investigate, when to ask permission and when to stop. That is the transition from automation to cognition—not machines replacing human operations, but enterprises developing a better way to perceive, decide, act and learn.