The typologies of artificial intelligence overlap depending on whether one adopts a technical, functional, or regulatory angle. Classifying AI systems by capability (narrow, general, superintelligence) remains common, but this framework is no longer sufficient to describe the current landscape. Since the gradual implementation of the European regulation on AI and the emergence of autonomous systems capable of orchestrating complete workflows, the very way of categorizing these technologies has changed.
Agentic AI: the functional type that classifications ignore
Most guides distinguish between predictive AI, generative AI, and supervised or unsupervised machine learning. Since 2025, Gartner has identified an additional category: agentic AI. These systems do not merely produce a response or prediction. They perceive an environment, make decisions, and execute actions autonomously or semi-autonomously to achieve a defined goal.
Related reading : How to Explore the Different Sections and Content Available on Zazie Web
In practical terms, an AI agent can manage an entire logistics process, orchestrate multiple specialized language models, or handle a data pipeline without human intervention at every step. The difference with a chatbot or a traditional generative model lies in this ability to act autonomously in a real environment, whether digital or physical.
What makes this category structuring is that it transcends the usual typologies. An AI agent can leverage supervised learning to analyze data, generative AI to produce a report, and predictive AI to anticipate an incident, all within the same sequence. The boundary between “types” then becomes porous, and understanding the different types of artificial intelligence now requires integrating this agentic dimension.
See also : Black and White Feather: Decoding the Spiritual Meaning of this Powerful Symbol
Regulatory classification of the AI Act: four levels of risk
The European regulation on AI (AI Act, regulation (EU) 2024/1689), which came into force on August 1, 2024, and is applicable in phases, introduced a reading framework that does not rely on the underlying technology or the degree of autonomy, but on the level of risk to individuals. This approach changes the way companies must think about their AI systems.
The four levels defined by the text are:
- Unacceptable risk: purely prohibited practices, such as social scoring or manipulating vulnerable individuals through subliminal techniques.
- High risk: systems used in health, safety, recruitment, or critical functions, subject to strict obligations for risk management, technical documentation, and human oversight.
- Transparency obligations: systems that interact with individuals (chatbots, deepfakes) and must disclose their artificial nature.
- Minimal risk: the vast majority of common applications, with no specific obligations beyond common law.
For European companies, this risk-based classification takes precedence over the technical distinction between generative AI and predictive AI. A generative model used to produce advertising images falls under minimal risk. The same model, applied to generating medical reports, shifts into the high-risk category with heavy compliance requirements.
Implementation timeline and concrete constraints
Prohibitions related to unacceptable risk apply first. Obligations for high-risk systems follow a progressive timeline. Companies deploying general-purpose AI models must provide technical documentation and comply with copyright regulations.
Field feedback varies on the ability of SMEs to absorb these new compliance requirements, especially when using third-party models hosted outside the European Union. The issue of cascading liability (model provider, integrator, end user) remains a legal friction point.
Generative AI and language models: state of play in 2026
Generative AI has captured most of the media attention for the past two years. Language models (LLMs) like those powering ChatGPT, Claude, or Gemini produce text, code, images, and video from prompts. Their application domain has expanded well beyond automated writing.
In business, uses focus on document synthesis, code generation, customer assistance, and marketing content creation. The available data does not allow for concluding that generative AI is replacing entire jobs at this stage, but it is clearly redistributing tasks within teams.
Persistent technical limitations
Hallucinations (factually incorrect responses presented confidently) remain a structural problem for LLMs. No generative model guarantees the factual accuracy of its outputs, which limits their deployment in regulated fields without a layer of human verification.
The energy consumption of training and inference is another open issue. Comparisons between models show significant performance discrepancies depending on the tasks, complicating the choice for companies looking to industrialize a specific use case.
Machine learning and predictive AI: mature applications
Before the explosion of generative AI, machine learning already constituted the technical foundation of the majority of AI systems deployed in production. Predictive AI, which analyzes historical data to anticipate trends or events, remains the most widespread type of AI in industrial sectors.
Predictive maintenance, fraud detection, credit scoring: these applications rely on supervised models trained on labeled datasets. Their reliability directly depends on the quality and representativeness of the training data, a point often underestimated during the transition to production.
Unsupervised learning finds its place in customer segmentation, network anomaly detection, or exploratory analysis of large databases. Unlike supervised learning, it identifies structures in data without prior labels, making it useful when categories are not known in advance.
Where predictive AI and agentic AI converge
The most advanced systems now combine a predictive module (which anticipates an event), a decision-making module (which chooses an action), and an execution module (which acts on the environment). This convergence blurs the boundaries between types of AI and makes rigid classifications increasingly ineffective.
The European regulatory framework pushes in the same direction: what matters is no longer so much the technique used as the final use and the associated risk. For organizations deploying AI systems, the relevant reading framework now combines the technical nature of the model, its degree of autonomy, and its regulatory risk level. None of these three dimensions alone is sufficient to describe what an AI system actually does.