Nobody has explained to you what AI is. Let us start there.

“Understanding AI” series, note 1 of 12. Twelve notes in order, from zero to the boardroom. This is the first.

Tuesday, 7:40, the train from Zug to Zurich. The managing director of a 60-person company reads her email on her phone. Three messages in a row. A supplier offers her “an AI solution” for her administration department. Her son sends her the school essay he wrote “with the AI” in ten minutes. And a member of the board asks her, with the courtesy of someone who has already decided this is a problem, what the company is doing about artificial intelligence.

Somewhere between Baar and Thalwil she notices something uncomfortable. She has been hearing about this every day for two years, she has paid for a licence almost nobody uses and, if the board member asked her for a definition right now, she does not have one. She can recognise it. She cannot say what it is.

The scene is imaginary. The situation is not. I meet it every week in people who run serious companies, with serious margins, and who talk about AI the way people talk about the weather: fluently, and without quite knowing how it works. It is not their fault. Almost everything published on the subject is written to sell or to frighten, and neither of those explains anything.

This series is my attempt to do so. Twelve notes, from the most basic question to the decisions only management can take. No jargon I have not explained first, examples from ordinary life, and one rule I impose on myself throughout: every figure comes with its source, and when something is a vendor’s estimate, I say so. By the end of this first note you will have two sentences to explain AI to anyone in your company. And I will tell you the most common mistake people make when explaining it, which is also the oldest: it is seventy-five years old.

AI is not a new technology. It is an old question that suddenly has answers that work.

A question from 1950 that is still open.

In 1950 the mathematician Alan Turing published a paper in the journal Mind that opened with a question, “can machines think?”, and dismissed it on the first page as too vague. In its place he proposed a game. A person puts written questions to two parties they cannot see, a human and a machine, and has to guess which is which. If they cannot, the machine has won. Today we call it the Turing test.

Turing also made a prediction: by around the year 2000, an average interrogator would have no more than a 70% chance of identifying the machine after five minutes of questions. The prediction failed, and for decades the game was an exercise in philosophy rather than engineering. What matters about it is not the date. It is that, from day one, we have measured a machine’s intelligence by how human it seems. Hold on to that idea. I will come back to it at the end, because it is the root of the mistake I promised you.

The summer it was given a name.

On 31 August 1955 four researchers signed a two-page proposal asking a foundation for money. John McCarthy of Dartmouth, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Labs. They asked for funding for “a 2 month, 10 man study of artificial intelligence” during the summer of 1956 at Dartmouth College, in New Hampshire. It is the first time the expression appears in writing.

The central sentence of the proposal deserves a slow reading: the study would proceed on the conjecture that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”. Notice two words. Conjecture: it was a bet, not a fact. And simulate: they never said reproduce. Seventy years later, both words are still the right ones.

What came next was not a straight line. There were decades of enthusiasm, decades of cuts and projects that did not deliver, and several generations of different techniques under the same name. That is the first thing worth knowing: “artificial intelligence” does not name a technology but a goal, and very different things fit under that goal. The spam filter in your inbox and the assistant that drafts contracts share a name and have nothing in common inside.

What the law says it is.

When a word means too much, it helps to look at how the people who have to apply it define it. The European Artificial Intelligence Act, in Article 3, defines an AI system as a machine-based system designed to operate with varying levels of autonomy, that may adapt after deployment and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.

It is a legal definition, which is why it sounds like one. But it contains one word that explains almost everything: infers. A traditional program follows instructions. If the customer is over 65, apply the discount; if the invoice exceeds a thousand francs, ask for a second signature. Someone wrote every rule, and the program does exactly that, always, with any data. An AI system does not work like that. Nobody wrote the rule. It deduces, from what it receives, the most probable result.

Think of the difference between a recipe and a cook. The recipe says two hundred grams of flour and twenty minutes in the oven, and anyone who follows it gets the same result. The cook has made that dish five hundred times, looks at the dough and knows that today, with this humidity, it needs five more minutes. The cook does not consult a rule. The cook infers. The recipe is software as we have always known it. The cook is AI. And like the cook, it is sometimes wrong, and it cannot always explain why it did what it did.

Three ways of building a machine that infers.

Many have been tried over seventy years, and three coexist today. I explain them with the same everyday example, a new employee in your company, because it is the image that works best when a decision has to be made.

The first is to write the rules. A group of experts puts everything they know in writing, as conditions, and the machine applies them faster than anyone. This is the employee who arrives with a thousand-page manual and follows it to the letter. In May 1997 Deep Blue, an IBM computer, beat Garry Kasparov 3.5 to 2.5 in a six-game match in New York. It evaluated two hundred million positions per second. It had not learnt to play: it calculated, with rules and brute force, better than any human can calculate. This approach is still alive in many companies and has one enormous virtue: it can be audited. If it goes wrong, someone can find the rule to blame.

The second is to learn from examples. Instead of writing the rules, you give the machine many solved cases and ask it to find the pattern. This is machine learning, and the idea has a date: in 1959 Arthur Samuel of IBM described a program that learnt to play draughts better than the person who wrote it. Today three variants are distinguished. Learning from labelled examples, like training a new employee with a thousand invoices already classified. Finding groups in unlabelled data, like discovering that your customers fall into four profiles nobody had defined. And learning by trial and error, with a reward for every success, which is how AlphaGo was trained: in March 2016 it beat the champion Lee Sedol 4 to 1 in Seoul, after studying thousands of human games and playing thousands more against versions of itself.

This second employee is very good at what it has seen many times and very bad at what it has never seen. And it has one trait management must know about: it learns what is in the examples, including the mistakes and the prejudices of whoever produced them. If for ten years your company granted credit in a biased way, the machine will learn to be biased with the same elegance.

The third is the one that changed the conversation: generative models. In June 2017 eight researchers published a paper called “Attention Is All You Need”, describing an architecture called the Transformer. The large language models were built on it. The underlying idea is almost offensively simple: a network is trained on an enormous amount of text to do one single thing, predict the next word. It is then tuned, with examples and human corrections, to behave like an assistant. And it turns out that, done at sufficient scale, this produces something that drafts, translates, summarises and writes code.

On 30 November 2022 ChatGPT was opened to the public. According to a note from the bank UBS reported by Reuters, by January 2023 it already had one hundred million monthly users, two months after launch; TikTok had taken around nine months and Instagram two and a half years. That is an analysts’ estimate, not a company figure, and I quote it as such. But it explains the scene on the train: for the first time, AI was something anyone could use without knowing anything about it. And that, which is good news, is also why so many people who use it do not know what it is.

This third employee is the most disconcerting. It has read more than anyone in your company, writes with a fluency that commands respect and, at the same time, has never been in your office, does not know your customers and does not know what day it is unless you tell it. I describe it in detail in the next note, because understanding how it works inside is what avoids the expensive mistakes.

What it is not.

Definitions by what something is not have a bad reputation, but on this subject they are the ones that save the most money. Five things AI is not, which almost every sales presentation takes for granted.

• It is not one thing. Under the same name sit a spam filter, a system that reads invoices, an assistant that writes and a program that plays go. When someone offers you “AI”, the first question is which of the three families it belongs to, because the risk, the cost and the way to test it differ in each.
• It is not new. It has seventy years of history, several generations of techniques and a long list of broken promises. What is new is that the third family works well enough for anyone to use, and that its price has fallen very fast.
• It does not know what it is saying. A language model produces the most probable continuation of a text. It has no intention, it does not check, it does not know it is wrong when it is wrong. This is not a defect the next version will fix, it is the nature of the method, which is why it has to be surrounded by checks.
• It is not an employee. The new-employee analogy helps me explain, but it should not be taken as far as the contract. An AI has no liability. Your company answers for what it does in your name, and courts have already said so. I will discuss those rulings in the note on security.
• It is not a vendor. A handful of laboratories build models, hundreds of companies package them and thousands resell them under another name. That a product “has AI” says almost nothing about what is inside.

The size of what is happening.

It helps to have two or three figures to hand, with their caveats. Stanford University’s AI Index 2025 report records that 78% of the organisations surveyed said they used AI in 2024, up from 55% the year before. It is a survey, and “using AI” covers everything from a trial in one department to a system in production, so read it as a trend, not a measurement.

The figure that interests me most is a different one. According to the same report, the cost of running a system at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. What was expensive to try two years ago now costs pennies, and this changes the question management should ask: it is no longer whether you can afford it, but what is worth doing with it and what is not. Also according to Stanford, the gap in results between the best model and the tenth best fell from 11.9% to 5.4% in a year. The models increasingly resemble one another. The advantage, when there is one, lies in how they are used, not in which one you buy.

And one figure that explains why I write to you from Switzerland and Spain: of the notable models published in 2024, Stanford counts 40 from the United States, 15 from China and 3 from Europe. The models are made elsewhere. What is made here, and where the value for a company like yours is decided, is the application.

How to explain it in two sentences.

Go back to the train. The board member has asked what the company is doing about AI, and this time the managing director has an answer. I propose this definition for you to say out loud at your next meeting, and for anyone on your team to repeat without needing to understand anything else:

A program that, instead of following exact instructions, has learnt from millions of examples to produce the most probable answer. It is right far more often than we expected, and it does not know when it is wrong.

Those two sentences contain everything management needs to begin. The first explains why it works for things that could not be programmed before: nobody wrote the rule. The second explains why it has to be surrounded by checks, contracts and people: it will never warn you that it has made a mistake.

And now the mistake I promised you. It is seventy-five years old and it was made, with all the intelligence in the world, by Turing himself when he proposed his game. It consists in judging AI by how human it seems. It is the mistake of those who trust an assistant because it writes with impeccable fluency, and of those who dismiss it because “it only predicts words”. Fluency measures nothing. A system can draft a perfect email to a customer and contain a false fact in the second line, stated with the same confidence as the true ones. Judge AI the way you judge a supplier: by what it delivers, measured, in your specific case. Never by how it sounds.

1. Today

   Write the definition down

   The two sentences, on a whiteboard or in the first message you send your team on the subject. Make sure everyone uses the same one.

2. This week

   Take stock of what you already have

   The spam filter, the phone’s autocorrect, the translator, the map that calculates the route. You already use AI in ten places; seeing them removes the fear and adds perspective.

3. This month

   Ask every department one single thing

   Which repeated task costs them the most time every week. Do not ask where they would put AI; that is not the question yet. The list of tasks is what counts.

4. Before the next meeting

   Read the second note

   How a language model works inside and why it sometimes makes things up. It is the one that prevents the mistake of trusting fluency.

How I approach it.

I do not sell AI. I design websites for clinics, estate agencies, boarding schools, hotels and law firms, and behind them I build automations that do the work nobody sees: answering in the client’s language at eleven at night, preparing the file the team finds on Monday morning, not losing the enquiry that arrived on a Sunday. To do it well I have to explain to management what this is and what it is not, and I have found that this explanation is worth more than any demonstration.

Before you go on, a simple test. Explain to someone on your team what artificial intelligence is in two sentences, without the words “magic”, “brain” or “think”. If you manage it, you have something many presentations lack. If not, do not worry: that is what the next eleven notes are for.

If you would rather see all this applied to your business, with your tasks, your systems and your figures, that is what I do in the Strategic Diagnostic. But you do not need to speak to me to keep reading.

Next note: how an AI “thinks” and why it sometimes makes things up. Tokens, working memory and the reason, demonstrated in 2025, why a model would rather guess than say “I don’t know”.