Artificial intelligence (AI) is the branch of computer science that studies how to build systems capable of performing tasks that, done by people, would require intelligence: perceiving, reasoning, learning from experience, understanding language or making decisions. The term was coined in 1956 at the Dartmouth conference, but it is worth deflating the myth from the outset: what we call AI today does not think or understand the way a human being does. These are statistical systems that recognise patterns in large volumes of data and produce predictions. That distinction is not pedantic; it is the basis for understanding what these tools can and cannot do.
Narrow AI versus general AI
The most useful conceptual distinction is the one proposed by the philosopher John Searle between weak AI and strong AI. Weak (or narrow) AI is designed to solve a specific task: classifying emails as spam, recommending products, transcribing speech or driving a vehicle. It is the only AI that exists today, without exception. However impressive a conversational assistant may seem, it remains a system specialised in predicting plausible sequences of text, not a general mind.
Strong AI, or artificial general intelligence (AGI), would be a system with cognitive ability comparable to a human's in any domain, capable of transferring knowledge from one task to another and of reasoning autonomously. It does not exist, and the scientific community has no consensus on when (or whether) it will arrive. Confusing the two concepts leads to unrealistic expectations and misguided investment decisions: no current system replaces human judgement; it assists it.
Machine learning: how a machine learns
The engine of almost all practical AI is machine learning, an approach in which the system infers rules from data rather than receiving them hand-coded. Three paradigms are distinguished:
- Supervised learning: the model learns from labelled examples (inputs with their correct answer). This is the case of regression, which predicts continuous values such as a price, and classification, which assigns categories such as "fraud" or "not fraud". It needs quality labelled data, which tends to be expensive to produce.
- Unsupervised learning: the model looks for structure in unlabelled data. Clustering, for example, groups customers with similar behaviour for market segmentation.
- Reinforcement learning: an agent learns by trial and error while maximising a reward, a paradigm used in robotics, games and process optimisation.
Deep learning, neural networks and transformers
Deep learning is a branch of machine learning based on artificial neural networks with many layers, capable of learning hierarchical representations: the first layers detect edges in an image, the next ones shapes, and the last ones complete objects. Convolutional networks (CNNs) dominate computer vision; recurrent networks (RNNs) were designed for sequences.
The decisive leap came in 2017 with the Transformer architecture, described in the paper "Attention Is All You Need". Its attention mechanism makes it possible to process long sequences while taking the full context into account, and it is the basis of the large language models (LLMs) that underpin today's conversational assistants. It is important to understand that these models predict the most likely next token: that is why they can produce false statements in a confident tone, a phenomenon known as hallucination. It is not an accidental glitch but a direct consequence of their statistical nature, and it makes it essential always to verify the outputs in critical contexts.
Types of AI: summary table
| Type | What it does | Practical example |
|---|---|---|
| Narrow (weak) AI | Solves a specific task | Spam filter, recommender |
| Supervised machine learning | Predicts from labelled examples | Fraud detection |
| Unsupervised machine learning | Finds patterns without labels | Customer segmentation |
| Deep learning | Learns complex representations | Image recognition |
| Generative AI (LLM) | Generates text, images or code | Assistants and document summaries |
| General AI (AGI) | General intelligence (hypothetical) | Does not exist today |
The European AI Act: the legal framework that changes the rules
Any organisation deploying AI in the European Union must be familiar with Regulation (EU) 2024/1689, known as the AI Act, the world's first comprehensive regulatory framework for artificial intelligence. Its logic is a risk-based approach with four levels. Unacceptable risk (for example, social scoring or subliminal manipulation systems) is prohibited. High risk (AI in recruitment, credit, medical devices or critical infrastructure) is subject to strict requirements for risk management, data quality, technical documentation and human oversight. Limited risk imposes transparency obligations, such as informing the user that they are interacting with a machine or labelling artificially generated content. Minimal risk is essentially unregulated.
The prohibitions began to apply in February 2025 and the rest of the regulation rolls out in stages. On top of this comes the GDPR, which remains fully in force: if an AI system processes personal data, it requires a legal basis, transparency and respect for the data subject's rights, including the right not to be subject to solely automated decisions with significant effects. The Spanish Data Protection Agency has published specific guidance on AI and data protection.
Real-world applications by domain
Beyond the theory, AI already generates measurable value in specific, well-bounded tasks. In computer vision, models inspect parts on a production line and detect defects at a speed and consistency unattainable for the human eye, or read number plates and documents to automate paperwork. In natural language processing, they classify incoming emails, extract data from contracts, summarise long documents and power assistants that resolve frequent queries. In predictive analytics, they anticipate demand to optimise inventory, estimate the risk of default or identify which customers are most likely to abandon a service.
What unites all these successful cases is their bounded nature: they work because the problem is well defined, the data is available and there is a clear criterion for success. The projects that fail tend to be those that expect AI to solve ambiguous or strategic problems that not even people have well defined. The practical rule is simple: AI shines in repetitive, pattern-based tasks with clear feedback, and falters in decisions that require common sense, broad context or ultimate accountability.
Responsible AI: bias, explainability and oversight
An AI system inherits the biases of the data it is trained on, and it can amplify them at scale. A recruitment model trained on biased historical decisions will reproduce that discrimination with the appearance of technical objectivity. That is why responsible AI is not an optional ethical add-on but a quality requirement and, in high-risk cases, a legal one. Its pillars are fairness (auditing and correcting bias in data and outputs), explainability (being able to justify why the system produced an output, as opposed to "black box" models), traceability (documenting data, versions and design decisions) and meaningful human oversight, which the AI Act expressly requires for high-risk systems. The ISO/IEC 42001 standard, published as the first AI management system standard, offers a framework for governing these requirements systematically.
Common mistakes when adopting AI in a company
- Starting with the technology and not the problem: buying an AI solution without a measurable use case that justifies the investment.
- Ignoring data quality: a model trained on biased or dirty data perpetuates and amplifies those biases (garbage in, garbage out).
- Trusting the outputs blindly: using generated text without human verification in legal, medical or financial contexts.
- Forgetting compliance: deploying a high-risk case without the documentation or oversight the AI Act requires.
Frequently asked questions
Is artificial intelligence going to replace my job?
AI automates tasks, not whole professions. It tends to transform roles by taking on the repetitive part and leaving people the judgement, the customer relationship and the oversight. The real risk is not being replaced by AI but by someone who knows how to use it better.
What is the difference between AI, machine learning and deep learning?
They are concentric circles: AI is the broadest field; machine learning is a subset that learns from data; and deep learning is in turn a subset of machine learning based on deep neural networks.
Why do language models make up facts?
Because they predict the most likely text, not the truth. They do not consult a database of facts; they generate plausible sequences. That is why, in any serious use, their outputs must be verified against reliable sources.
Do I need a lot of data to apply AI in an SME?
Not always. For common tasks there are pre-trained models that adapt with few examples, and many valuable cases (customer service, document classification) are solved without building a model from scratch.
Conclusion
Understanding artificial intelligence means accepting an uncomfortable idea: its power comes not from thinking but from recognising patterns at a scale no person can reach. That statistical nature explains both its spectacular successes and its hallucinations, and it marks exactly where human judgement remains indispensable. For a company, the value lies not in jumping on the bandwagon but in identifying the specific task AI can take on best, feeding it quality data, maintaining oversight and respecting a legal framework (AI Act and GDPR) that in Europe is no longer optional. General AI has not arrived; useful AI has, and the advantage will go to those who integrate it with rigour rather than with enthusiasm. At Summum we support that integration by measuring results, not by promising magic.