Artificial intelligence: definition, types and maturity levels

SECTION 1

Defining artificial intelligence: simulating human cognitive faculties

Artificial intelligence refers to the branch of computer science that seeks to reproduce human cognitive faculties within machines: learning from data, reasoning about new situations, perceiving an environment, understanding language and making decisions. According to the definition popularised by John McCarthy, one of the founding fathers of the field, it is the ability of a computer to perform tasks that normally require human intelligence, using language, forming concepts, solving problems once reserved for humans.

The field actually covers a set of subdisciplines long developed separately: machine learning, natural language processing, computer vision, robotics, planning and optimisation. These building blocks now converge within single systems able to handle text, images and sound, which blurs academic boundaries but does not remove the need to understand what each one contributes, or what it costs.

The Dartmouth research project, sketched out as early as 1955, had identified seven major areas of work, from simulating brain functions to abstraction, self-improvement and creativity. Seventy years later, that programme remains strikingly current: today's systems learn, generalise and produce original content, even though their mechanisms have little in common with the symbolic approaches of the early days.

### From symbolic AI to foundation models

The history of the field alternates summers and winters: rule-based expert systems in the 1980s, the rise of statistical learning in the 1990s and 2000s, the deep learning revolution from 2012 onwards, then the Transformer architecture in 2017. The latter opened the way to foundation models: neural networks trained on immense corpora, able to serve as the base for thousands of applications without full retraining. That shift is what explains the spectacular acceleration observed since 2022.

SECTION 2

Weak AI, general AI and super AI: the three main categories

Weak AI, also called narrow AI, covers systems designed to excel within a bounded set of tasks, without claiming to reproduce thought. IBM's Deep Blue chess computer was its historical illustration. Notably, today's large language models, however versatile, still formally belong to this category. They cover an unprecedented range of tasks, but remain statistical systems with no intentions of their own and no general understanding of the world.

General AI (AGI, for artificial general intelligence) refers to a system able to match human cognitive abilities across every domain: learning a new skill from few examples, transferring reasoning from one field to another, pursuing a goal over time. The topic has left pure speculation behind: frontier labs make it a stated objective, and the debate now concerns timelines and measurement criteria, with no established scientific consensus.

For a company, this taxonomy is anything but theoretical: it is a reminder that every system deployable today is narrow AI, with a validity perimeter to define, edge cases to test and guardrails to put in place. Announcements of imminent general intelligence should not dictate a roadmap; it is the measurable capabilities of the moment, confronted with precise use cases, that determine it.

### Super AI, a speculative horizon

Super AI refers to a system that would surpass human intelligence across all domains. It remains hypothetical, but it structures a growing share of AI safety research: goal alignment, model control and interpretability, international governance. This work is no longer distant futurology, it conditions the trust that companies and regulators place in the most capable systems.

SECTION 3

The four functional levels: from reactive AI to self-aware AI

A second reading grid ranks systems by degree of functional sophistication. Reactive AI produces predictable outputs from given inputs, with no memory of the past: Deep Blue and the first AlphaGo belong here. Limited memory AI goes a step further by learning from past data and observations, this is the category of virtually every system deployed today, from autonomous vehicles to language models.

### Limited memory, the category of today's systems

Recent models push the boundaries of this category: context windows of several million tokens, persistent memory across sessions, retrieval-augmented generation (RAG) that plugs the model into the company's knowledge. These mechanisms create an impression of continuity, but do not change the nature of the system: it learns from past data, it does not represent other people's mental states.

Autonomous vehicles illustrate this level well: they combine real-time perception with learning from millions of kilometres of past driving to adjust their decisions. Recommendation engines, fraud detection systems and coding assistants follow the same logic, learning from yesterday to decide better today, without ever understanding why the decision matters.

The two upper levels remain research topics. Theory of mind targets machines able to represent the intentions, beliefs and emotions of their interlocutors, the condition for genuine social interaction. Self-aware AI, finally, would hold a representation of itself and understand its own existence. No current system comes close, despite surface behaviours that can be unsettling.

@cite:les-composants-de-l-ia

SECTION 4

Foundation models and generative AI: the turning point of the 2020s

Since late 2022 and the arrival of mainstream conversational assistants, generative AI has pushed narrow AI into a new era. Foundation models, large language models, image, audio and video models, are now natively multimodal: a single system reads a document, analyses an image, listens to a conversation and produces text, code or visuals. They serve as a common base for thousands of business applications, from software development assistance to document analysis.

Concretely, a single enterprise platform can now summarise a call for tenders, extract the sensitive clauses of a contract, generate a tested and documented software component or prepare a presentation from raw data. This versatility explains the speed of adoption: most large European organisations now run at least one generative use case in production, well beyond the experimentation stage.

### Reasoning models and inference-time compute

The striking evolution of recent years is the rise of reasoning models: rather than answering immediately, they allocate compute at inference time to explore several paths, check their steps and correct their mistakes. These extended reasoning chains markedly improve performance in mathematics, programming and scientific analysis, at the price of higher latency and cost, a trade-off that every use case now has to make explicit.

The landscape has also diversified: frontier proprietary models, very capable open-weight models, and small specialised models able to run on a workstation or at the edge. Inference costs have dropped by several orders of magnitude in three years, which shifts the strategic question: it is no longer whether AI can be used, but where it actually creates value.

@cite:ia-conversationnelle-definition-exemples

SECTION 5

Agentic AI: models that plan and act

Agentic AI extends foundation models into systems able to pursue a goal: breaking a task down, calling tools, search engines, databases, business APIs, observing the result and adjusting their plan. Interoperability standards such as the Model Context Protocol (MCP) normalise how agents connect to information systems, which accelerates their adoption in enterprise workflows: customer support, software engineering, IT operations.

The first at-scale deployments concern well-bounded processes: sorting and qualifying support tickets, assisted code review, preparing compliance files, reconciling data across systems. The feedback is consistent: value comes less from maximum autonomy than from the reliability of the agent-process pair, measured against explicit business criteria.

### From copilot to supervised agent

Progress happens through autonomy levels: an assistant that suggests, a copilot that executes under validation, an agent that handles a process end to end with human checkpoints. Mature organisations wrap each level in explicit guardrails, limited permissions, logging, continuous evaluation of outcomes, because a failing agent can act, not merely get an answer wrong.

@cite:l-ia-agentique-au-travail-comment-les-agents-autonomes

SECTION 6

AI's real maturity in 2026: capabilities, limits and regulation

In production, most deployed systems still belong to narrow, limited memory AI, however impressive it may be. The known limits persist: hallucinations that require human verification for critical uses, sensitivity to prompt injection attacks, inference costs to keep under control, dependence on data quality. Calibrating expectations remains the first skill of any AI project.

Evaluation has become a discipline in its own right: academic benchmarks saturate and quickly become outdated, so serious organisations build their own evaluation sets, representative of their use cases and their data. It is the only reliable way to compare models, detect regressions as updates roll out and ground the decision to go to production.

The regulatory framework has taken shape: in Europe, the AI Act entered into force in August 2024, its obligations for general-purpose models have applied since August 2025, and those targeting high-risk systems become enforceable from August 2026. Use-case classification, documentation, human oversight and risk management are now an integral part of any serious deployment.

SECTION 7

Turning this reading of the landscape into an AI strategy with Adservio

At Adservio, we start from this clear-eyed map, what is operational, what is emerging, what remains speculative, to anchor every project in the capabilities that are actually available. Telling apart a reactive AI, a reasoning model and a supervised agent is not an academic exercise: it is what makes it possible to choose the right technology for the right use and avoid costly dead ends.

Our conviction: technology is only worth as much as its link to clear business objectives and a realistic trajectory. We help your teams frame their use cases, select the relevant models, put in place the evaluation and governance required by the new regulatory framework, and master operations to turn the potential of AI into measurable results.

FAQ

Frequently asked questions