Artificial intelligence is the set of technologies that enables computer systems to learn from data, recognise patterns, process language and perform tasks that require cognitive abilities. For a company, understanding how it works also means distinguishing what current systems can do from prospects that remain open, such as artificial superintelligence.
This guide starts from the definition and history of AI, explains how the models work and moves on to business applications. It also addresses a question for decision-makers: how to choose tools, assess their results and build an AI governance system capable of managing costs, accountability and risk. All the way to Corallo.ai, the Italian platform that integrates AI into companies’ operational processes.
What is Artificial Intelligence: definition and meaning
Artificial intelligence (Artificial Intelligence, or AI) is the branch of computer science that studies and develops systems capable of performing tasks that, if carried out by humans, would require intelligence: reasoning, learning from experience, understanding language, recognising images and solving complex problems.
The most widely cited definition in academic circles is that of John McCarthy, who in 1956 described AI as
“the science and engineering of making intelligent machines, especially intelligent computer programs”.
Today the definition has evolved: artificial intelligence is a set of technologies that enable machines to simulate human cognitive abilities, with an increasing degree of autonomy and adaptability. Just how far that degree will increase, however, remains to be seen.
The main sub-disciplines
When people talk about artificial intelligence, many concepts are often lumped together. The main subdisciplines of AI are:
- Machine Learning (ML). A subfield of AI that enables systems to learn from data without being explicitly programmed. The model is trained on a dataset, identifies patterns and improves its predictions with experience. It underpins most current AI applications.
- Deep Learning (DL). An evolution of machine learning based on artificial neural networks with many layers. Particularly effective for image recognition and natural language processing. It has produced the most significant breakthroughs of the past decade.
- Natural Language Processing (NLP). The field concerned with how machines understand and generate human language. It underpins voice assistants, chatbots, machine translation and large language models.
- Generative Artificial Intelligence. The category of systems capable of generating original text, images, video and code. GPT, Claude and Gemini are examples of Large Language Models (LLMs) belonging to this category. Released to the public from 2022 onwards, they have redefined the AI debate worldwide.
- Agentic Artificial Intelligence. This covers systems capable of planning a sequence of actions, using external tools and operating with varying degrees of autonomy to achieve a goal. To understand how these capabilities enter the processes of an SME, you can explore the topic further in our article ‘what is an AI agent’.
History of Artificial Intelligence: from its origins to 2026
1950–1956: the foundations
1950 is conventionally regarded as the year in which thinking about AI was born. It was then that Alan Turing published Computing Machinery and Intelligence, the paper in which he posed the question “Can machines think?” and introduced the Turing Test: if a machine can hold a conversation indistinguishable from that of a human being, it can be considered intelligent.
In 1956, at the Dartmouth Conference, John McCarthy, Marvin Minsky, Claude Shannon and other researchers officially coined the term Artificial Intelligence. It was the moment AI was born as a formal academic discipline.
1957–1974: the optimism of the early years
The early years of AI were marked by extreme optimism. Researchers developed the first chess-playing programs, the first problem-solving systems and the first attempts at machine translation. DARPA funded the research generously. Minsky and other scientists were already predicting machines capable of general intelligence within twenty years. The computing power of the time, however, could not sustain those ambitions.
1974–1980: the first AI Winter
The term AI Winter describes the periods in which research funding contracted because results fell short of expectations. The first AI Winter, between 1974 and 1980, was triggered by the 1973 Lighthill report, which questioned the discipline’s real progress and led to cuts in government funding in Britain and the United States.
1980–1987: the era of expert systems
The 1980s saw a revival with expert systems: programs able to simulate the decision-making of a human expert in a specific domain. Digital Equipment Corporation’s XCON cut server configuration costs by around 40 million dollars a year. The AI market reached one billion dollars for the first time. This cycle, too, ended in a second AI Winter, when the structural limitations of expert systems became evident: they were fragile, costly to maintain and unable to generalise.
1990–2006: machine learning and the silent transition
While the term AI was losing commercial credibility, research continued under other names. In 1997 IBM’s Deep Blue defeated world chess champion Garry Kasparov. Recommendation algorithms were developed that would go on to power Amazon and Netflix. The internet was producing an unprecedented volume of data. That data was the fuel machine learning models had been waiting for.
2006–2012: the Deep Learning revolution
In 2006, Geoffrey Hinton published the paper that revived artificial neural networks under the name of deep learning. The availability of powerful GPUs and large datasets finally made it practical to train networks with many layers. The watershed moment came in 2012, when AlexNet, the network developed by Hinton’s team, won the ImageNet competition, cutting the error rate in visual recognition by more than 10% compared with previous approaches.
2012–2020: commercial AI spreads
Google, Facebook, Amazon, Apple and Microsoft build AI teams and acquire the most promising start-ups. Speech recognition systems improve radically (Siri, Alexa, Google Assistant). In 2016, DeepMind’s AlphaGo defeats the world Go champion: for the first time, a machine beats a human being at a task that required intuition and long-term strategy, not just calculation.
2020–2026: the era of Large Language Models
2020 marks a new phase with OpenAI’s release of GPT-3. In November 2022, the public launch of ChatGPT brings language models to the attention of a far wider audience. In the years that follow, these systems begin to write code, use external tools and tackle increasingly complex sequences of tasks, while still requiring verification and oversight in the most sensitive applications.
In 2026 the industrial scale of the competition is changing too. Training and deploying frontier models requires capital, infrastructure and revenues capable of sustaining enormous investment. It is in this context that the possible IPOs of Anthropic and OpenAI are being discussed: a listing would expose the two companies’ costs, growth and economic sustainability to the judgement of the markets.
Artificial intelligence and geopolitics: the new digital Cold War
Which brings us to the present day. In 2026 artificial intelligence has become an instrument of national power. The competition between the United States and China for AI dominance mirrors the structure of the space race of the 1950s and 1960s, but with a faster pace of change and implications that span the economy, defence and national security.
The Trump administration has published America’s AI Action Plan, a document whose subtitle, Winning the Race, makes the American stance unambiguously explicit: AI is a competition to be won against China. In parallel, Trump has frozen the executive order that provided for forms of federal oversight of frontier models. The logic is clear: any form of governance is read as a competitive brake. In 2026, speed of development has itself become a form of sovereign power. This was demonstrated by the speech delivered on 22 September 2026 at the UN General Assembly, where Trump proposed replacing, in US documents, the expression artificial intelligence with super intelligence, or SI. The new name, superintelligence, technologically inaccurate, clearly expresses the tycoon’s political position: avoid any alarmism, and never slow the process down.
China has responded with the Global AI Governance Action Plan, which aims at multilateral governance of AI on a global scale. Meanwhile, Europe finds itself in an uncomfortable position. On the regulatory front, Europe has chosen its own path: from 2 August 2026 a significant part of the AI Act’s provisions applies, including the transparency obligations. The remaining rules follow later deadlines. As Executive Vice-President Henna Virkkunen stated: “Those who lead technological innovation will shape the future, and we must ensure that Europe plays a leading role in this process.”
In 2026, AI has become the most contested geopolitical asset on the planet. Every trade agreement, every restriction on chip exports, every infrastructure decision has implications that reach far beyond the technology market. For companies that depend on AI providers, cloud infrastructure or international markets, ignoring this dimension means operating without accounting for a structural risk variable.
For Europe, the question also concerns control over the infrastructure and technologies on which businesses and institutions depend. Professor Marco Tupponi explores this relationship between technological dependence and decision-making autonomy in Bliss Faculty’s analysis of European digital sovereignty in the era of frontier AI.
The leading AI superpowers: who controls the future
The AI market is dominated by a small number of organisations that control the computing infrastructure, frontier models and research pipelines. Knowing who they are helps in navigating a rapidly changing ecosystem.
American dominance
In the United States, a handful of players control resources that are decisive for AI development. Nvidia supplies the hardware on which many frontier models are trained. OpenAI, Google DeepMind and Anthropic develop systems used by individuals, businesses and institutions; Meta contributes to the spread of open-weight models. Each of these organisations makes decisions that affect access to the technology, its costs and the ways it can be used.
The relationship between business and government has become part of the same competition. In 2026, the clash between Anthropic and the Pentagon over the terms of use of Claude showed how far control of a model can have political consequences. The story is reconstructed in Anthropic vs the Pentagon: the full story.
The Chinese response
China is developing its own ecosystem of models, infrastructure and platforms. Alongside established groups such as Baidu and Alibaba, the arrival of DeepSeek in 2025 made Chinese labs’ ability to compete on advanced models more visible. For companies, this competition widens the range of choice, but makes it even more important to assess the model’s origin, data handling and dependence on suppliers.
The European exception
In Europe, the only player of scale is Mistral AI, a French start-up founded in 2023 that has raised over 600 million euros and developed open-source models that compete with American ones. It is the only European lab competing in the arena of frontier models. The rest of the European ecosystem is fragmented, underfunded and structurally dependent on American or Chinese cloud infrastructure.
The concentration of the AI market in a few players is one of its most significant structural features. For companies integrating AI into their processes, this means depending on infrastructure they do not control, on rules set by others and on pricing decisions that can change unilaterally.
The Vatican’s involvement
On 15 May 2026, one hundred and thirty-five years after Leo XIII’s Rerum Novarum, Pope Leo XIV signed Magnifica Humanitas, devoted to artificial intelligence. Just as the labour question accompanied the Industrial Revolution, AI today reopens the debate on work, economic power and human dignity. The question also concerns those who lead businesses: who decides how these systems are used, and who is accountable for their consequences? We examined the Pope’s position in Even the Pope is talking about AI.
Artificial intelligence and the market: investment, Stargate and the capex of the AI era
2026 is the year in which investment in AI infrastructure reached a scale unprecedented in the history of the tech industry. For a sense of magnitude: in the first half of 2025, AI capex accounted for more than a third of all US GDP growth. No technology has ever moved so much capital in so little time. Ever.
The Stargate Project
On 21 January 2025, OpenAI, SoftBank and Oracle announced the Stargate Project, with the stated aim of investing up to 500 billion dollars over four years in US AI infrastructure. The project shows how far data centres, energy and computing capacity have become part of technological competition. To read the outlook for AI, therefore, one must also consider the resources needed to develop and deploy it.
Big Tech capex
The five largest US technology companies, Amazon, Alphabet, Meta, Microsoft and Oracle, will spend between $660 and $690 billion on AI infrastructure in 2026. That is a 69% increase on the previous year. Amazon has announced investments of up to 200 billion in the same year. Alphabet stands at between 175 and 185 billion, six times the 31 billion it invested in 2022. Goldman Sachs estimates that annual AI capex will reach 765 billion during 2026 and $1.6 trillion by 2031.
Valuations and paradoxes
Investment in and valuations of AI companies raise a question about the sector’s sustainability: how much economic value will the models generate relative to the capital required to develop them?
The comparison with history is a common one. The American railways of the 19th century were a real and transformative technology: they changed the American economy irreversibly. Yet between 1870 and 1900 they destroyed enormous amounts of investor capital, because valuations ran years ahead of actual adoption. AI is real. The open question is whether current valuations are pricing in twenty years of development that has not yet happened.
Meanwhile, a study by the MIT Media Lab found that 95% of organisations that invested in generative AI reported no measurable financial return. The gap between capital invested and documented returns is the most important number to keep in mind when reading any market analysis of AI.
For an Italian company assessing whether and how to invest in AI, this context has direct implications. The infrastructure is controlled by a handful of American and Chinese players. Costs can change unilaterally. The rules are written elsewhere. The answer is not to abstain, but to build adoption governance that does not depend on a single supplier and that measures return on investment before scaling it.
How Artificial Intelligence works
Understanding how AI works at a conceptual level helps in assessing its applications. Modern AI systems rest on three fundamental elements.
1. Data: machine learning models learn from examples. The more abundant, diverse and high-quality the data, the better the performance. Data quality is the variable that most determines the effectiveness of an AI system in a real business context.
2. Model architecture: the mathematical structure that processes the data. Transformer networks, on which GPT, Claude and Gemini are based, are effective for language. The choice of architecture depends on the type of problem to be solved.
3. Training: the process through which the model learns. Over millions of iterations, the model adjusts its parameters to minimise prediction error. Training requires significant computing power; inference, that is, using the already-trained model, requires far less.
The most powerful AI models in 2026: mapping the frontier market
As of September 2026, the map of frontier models is changing rapidly. Performance must be read in relation to the task: programming, research, document processing, use of external tools and cybersecurity require different capabilities. Cost, access and security controls also affect the choice of a model for the business.
The main proprietary models
OpenAI has introduced GPT-6 Astra, presenting it as its most capable model deployed at scale. Its capabilities, including in computing and cybersecurity, make the question of the safeguards needed when a system can use tools and tackle complex tasks even more central.
Google continues to develop the Gemini family, offering models with differing capabilities and costs. The launches of the Gemini 3.8 series show how far competition now also extends to speed, modes of interaction and the ability to integrate AI into products used every day.
Anthropic has introduced Claude Opus 5.5, focusing on complex professional tasks, programming and the balance between performance and cost of use. The same company developed Claude Mythos and Project Glasswing, opening the debate on access to the most capable models. The project has since evolved: Anthropic offers Mythos’s capabilities to the public through Claude Fable 5.1, with additional safeguards, while reserving Mythos 5.1 for selected organisations for specialist uses.
xAI, with its Grok family, is also competing in proprietary models. For a business, however, the model’s name alone says little: it must be assessed against the specific use case, the data it will handle, its costs and its terms of use.
From models to specialised applications
The frontier does not advance only through a more powerful model. Anthropic, for example, built Claude Science as a working environment for researchers, connecting existing Claude models to scientific tools and databases. It is a novelty in how the model is applied, not the name of a new model.
A different path emerges with Midjourney Medical: the case concerns the company’s entry into medical imaging through a hardware project. The two examples show why, to understand where AI is heading, one must also look at the applications and expertise that connect the technology to a specific sector.
The rise of open-source models
One of the most significant developments of 2026 is the closing of the gap between proprietary and open-source models. DeepSeek V3.2 (China, open source) delivers 90% of GPT-5.5’s performance at one fiftieth of the cost. GLM-5.1 from Zhipu AI briefly held first place on SWE-bench Pro: the first open-source model to achieve that result. For companies running internal workloads with privacy requirements, hybrid architectures have become the norm: open-source models for internal processes, proprietary APIs for high-criticality production tasks.
Choosing a model has become a strategic decision, no longer merely a technical one. It depends on the specific use case, compliance requirements, cost structure and tolerance for the risk of hallucinations. There is no single best model: there is the model best suited to the organisation’s specific problem.
The European AI Act: the world’s first regulatory framework for artificial intelligence
The European Regulation on artificial intelligence, known as the AI Act, entered into force on 1 August 2024 and applies in phases. From 2 August 2026 many of its provisions are in effect, including the transparency obligations under Article 50, examined in depth by Professor Marco Tupponi. Other rules, in particular those concerning several categories of high-risk systems, follow later deadlines.
The AI Act classifies systems according to the risks they may pose to safety and fundamental rights. Obligations vary depending on the type of system and the role of the organisation that develops or uses it.
| Unacceptable risk: banned | AI systems that manipulate human behaviour subliminally, social scoring systems operated by governments, real-time biometric recognition in public spaces for law enforcement purposes (with limited exceptions). Banned across the EU since 2 February 2025. |
| High risk: stringent obligations | AI systems used in recruitment, credit scoring, critical infrastructure (energy, water, transport), education, law enforcement, migration management and the administration of justice. Mandatory: prior conformity assessment, registration in an EU database, human oversight and detailed technical documentation. |
| Limited risk: transparency | Systems that interact with humans (chatbots, deepfakes) are subject to a transparency obligation: users must be informed that they are interacting with an AI system. |
| Minimal risk: free use | The vast majority of AI applications: spam filters, recommendation engines, logistics optimisation systems. Subject only to the general rules of EU law. |
It is also worth noting that, alongside the European framework, Italian law regulates certain unlawful uses of AI in Law 132/2025.
Artificial Intelligence in business: practical applications in 2026
91% of Fortune 1000 companies increased their AI investment in 2026. 63% of marketing and communications professionals use generative AI tools in their workflows. The global AI market is expected to reach 1.8 trillion dollars by 2030.
The point, for a business, is to understand which applications solve a concrete problem and how to measure their results. Use cases vary widely across business functions: automating a repetitive request, querying internal documents and supporting a sales forecast each require different data, controls and indicators.
Not to mention the management side. When employees adopt tools without a shared assessment, the problem of shadow AI emerges: the organisation can lose visibility over the data being shared, the costs and actual usage. All of this has delicate consequences for the organisation.
| Process automation | Periodic reporting, request routing, database updates, notifications between systems: tasks that consume hours without generating strategic value. AI systems that automate these workflows give operational capacity back to the people who used to perform them. Organisations that have implemented AI automation in these areas report reductions in time spent of between 40 and 70%. |
| Business intelligence | Traditional BI systems show what happened. AI systems explain why it happened and project what will happen if no action is taken on a specific variable. A system that flags a drop in conversion on the same day, identifies the cause and suggests the intervention is worth more than a report that records it at month-end, when the budget has already been spent. |
| Knowledge management | Documents, contracts, procedures, manuals: every organisation holds a body of information that is inaccessible in practice because it is buried in folders and formats that cannot be queried. Systems based on RAG (Retrieval-Augmented Generation) make it possible to query this information in natural language, obtaining answers in seconds rather than minutes or hours. |
| Reputation monitoring | Reputational crises build up before they erupt. An AI system that monitors brand mentions, sentiment by channel, competitor moves and emerging trends in real time makes it possible to pick up weak signals before they become problems. The difference between managing a crisis and preventing one is almost always a matter of hours. |
| Personalised communication | AI systems produce tailored versions of communications materials from a single brief, adapting tone, format and register to the channel, the segment and the stage of the funnel. On average, personalised communications outperform standardised ones by 40-60%. |
| Forecasting and planning | From demand forecasting models to strategic scenario simulations: AI systems trained on historical company data produce more accurate forecasts than those generated with traditional approaches, in far less time. In supply chain, budget allocation and workforce management, the impact of more accurate forecasts is directly measurable in margins. |
How to implement AI in a company: a practical guide
The main mistake in adopting AI is one of method. Organisations that have generated no return on their AI investments, or that use it without any governance, almost always share the same profile: they bought tools before defining the problems to be solved.
A corporate AI policy clarifies which tools may be used, for which activities and with what responsibilities. To develop one, you first need to establish these five points clearly.
1. Audit and mapping. Before any investment, a precise mapping of processes, data quality and availability, and technology infrastructure is required. Without this step, any estimate of return is an untested assumption.
2. Identifying quick wins. The AI projects with the fastest ROI are those where data is already available and structured, processes are documented and the team has a basic knowledge of the tools. Starting here reduces risk and builds the organisational confidence needed for more ambitious projects.
3. Architecture design. The solution must integrate with the workflows and tools already in use, without requiring infrastructural upheaval. An AI system that requires 30% of existing processes to change in order to work carries an adoption cost that often exceeds the expected benefits. For the most sensitive processes, it is therefore necessary to define human oversight of AI systems: who reviews the output, when they intervene and with what authority.
4. Training. An AI system without a team that knows how to use and govern it is a wasted tool. Training is not an ancillary cost: it is a precondition for a return on the investment.
5. Continuous measurement and optimisation. AI systems improve with use and feedback. Performance measurement must be part of the architecture from the outset, not an afterthought.
For organisations that need to turn these criteria into an operating system, Bliss’s AI Governance programme starts by mapping the tools in use and ends with measuring the results.
Corallo.ai: the future of AI in Italy
Corallo.ai is the artificial intelligence platform developed in Italy for organisations that want to make AI operational in their day-to-day work.
The name is no accident. Coral does not create resources: it filters them. It takes in turbulent water and returns nourishment, working in the depths where no one is looking, building structures that last. The metaphor describes what the system does with company data: it brings it to the surface, makes it queryable and turns it into decisions.
The services available include an intelligent Knowledge Base that can be queried in natural language, decision dashboards that interpret anomalies and suggest actions, automation of repetitive processes, real-time reputation monitoring across more than 47 channels, AI Audit & Readiness Assessment and training for corporate teams. The entire team is AIFA-certified.
Tomorrow has arrived
There is a moment, in every major transformation, when change stops being a forecast and simply becomes the world you find yourself in. That artificial intelligence is that transformation is no longer a thesis to be defended: it is the context in which every organisation already operates, regardless of sector, size or will. The race between the United States and China, which treats AI as military infrastructure, says so. Leo XIV says so, having written an encyclical on it in the same year OpenAI reached a valuation of 730 billion. The 660 billion in capex invested in 2026 by five private companies says so.
The question is not whether AI will change the way we work, decide and compete. The question is who will reach that change with a system and who will not. In 1950, Alan Turing asked whether machines could think. Seventy-six years later, the more urgent question is: what will those responsible for leading an organisation choose to do with this capability?
Domande frequenti
What is artificial intelligence, in simple terms?
Artificial intelligence is the ability of a computer system to perform tasks that normally require human intelligence: understanding language, recognising images, making decisions, learning from experience. In practical terms, it is the technology behind chatbots, recommendation engines, autonomous driving, market forecasting and business process automation.
What is the difference between artificial intelligence, machine learning and deep learning?
Artificial intelligence is the broadest field: it includes all systems that simulate cognitive abilities. Machine learning is a subcategory: systems that learn from data without being explicitly programmed for each task. Deep learning is a subcategory of machine learning: it uses neural networks with many layers and produces the most advanced results on complex tasks such as visual recognition and language understanding.
When was artificial intelligence born?
The term was officially coined in 1956 at the Dartmouth Conference. The theoretical foundations had been laid in 1950 by Alan Turing. As an operational discipline with commercial applications, it developed mainly from the 1980s onwards with expert systems, and reached commercial maturity with deep learning in the 2010s. The breakthrough for the general public came in November 2022 with the launch of ChatGPT.
Can artificial intelligence replace workers?
AI automates tasks, not roles. The most exposed tasks are repetitive, high-volume and low-value: data collection, standard reporting, request routing. Roles that call for judgement, creativity, relationships and the handling of unforeseen situations are far less exposed. The best-documented effect in organisations that have implemented AI is a redefinition of what is worth working on, not a reduction in headcount.
How does an Italian SME start implementing AI?
The starting point is the audit: mapping the processes where time spent generates no strategic value, checking the quality and availability of existing data, and identifying quick wins with an estimated ROI. SMEs have an advantage over large organisations: fewer information silos, processes that are simpler to map, fewer obstacles to implementation. The constraint is not budget: it is methodology. Corallo.ai offers an AI Audit & Readiness Assessment service designed specifically as a starting point.
What does generative AI mean?
Generative AI is the category of systems capable of producing original content: text, images, video, code, audio. ChatGPT, Midjourney, Claude and Gemini are examples. Unlike discriminative systems, which classify or predict, generative systems create new outputs from an input.
What is the difference between AI and traditional automation?
Traditional automation executes predefined sequences of instructions: if X happens, do Y. It works in stable, predictable contexts, but does not adapt to unforeseen variations. AI learns from data and generalises: it handles variation, ambiguity and new situations. The practical difference is that traditional automation requires every case to be explicitly programmed, whereas AI learns how to handle them from experience.