On this page
Examples in, a rule out
In machine learning nobody writes the rule: you show the software enough examples and it finds the pattern itself. You give the system data plus the answers that belong with it; the link between them is what it learns. That pattern is called a model. From then on, the model predicts the answer for data it has never seen.
Classic software works the other way round: someone writes that an email containing the word invoice goes to accounts. That holds until a customer calls it a bill. A model that learned from a thousand sorted emails weighs hundreds of signals at once and still routes that email correctly.
You probably use machine learning already. Smart Bidding in Google Ads trains machine learning algorithms on large amounts of data to predict, auction by auction, which bid leads to a conversion. The language models behind ChatGPT and Claude are machine learning too, at a scale no small business trains itself. How they read text in small pieces, and what that costs per language, is explained under AI tokens.
Arthur Samuel and a checkers program, 1959
The term is attributed to Arthur Samuel, a researcher at IBM. In 1959 he published 'Some Studies in Machine Learning Using the Game of Checkers' in the IBM Journal. His aim was a computer that would learn to play better checkers than the man who programmed it. That idea, software that improves through experience rather than through extra rules, has not changed in sixty years. What changed is the amount of data and computing power.
Five dates, from checkers to language models
- 1959Samuel at IBMArthur Samuel describes a checkers program that learns from experience, and uses the term machine learning.
- 2012Recognising imagesKrizhevsky, Sutskever and Hinton train a deep neural network on 1.3 million photos in 1,000 classes. Deep learning becomes the standard for images.
- 2017The transformerVaswani and seven colleagues publish Attention Is All You Need, the architecture under today's language models.
- 2020GPT-3A language model with 175 billion parameters learns a new task from a few examples in the request, without extra training.
- 2024Nobel PrizeJohn Hopfield and Geoffrey Hinton receive the Nobel Prize in Physics for the work that made machine learning with neural networks possible.
Train, test, use
Every project runs through the same three steps. First training: the system gets examples and adjusts its internal numbers until its predictions sit as close as possible to the real answers. Then testing: you show the model examples it never saw and measure how many it gets right. Only then use, which practitioners call inference: the model receives new data and returns a prediction, a label or a number.
How it learns depends on what you give it. There are three main kinds.
| Kind | What you give the model | What it learns | Example in a small business |
|---|---|---|---|
| Supervised learning | Examples with the right answer (labels) | To give a new case the right label | Sorting incoming email into quote, complaint or invoice |
| Unsupervised learning | Data only, no answers | Groups and outliers | Grouping customers by what and how often they order |
| Reinforcement learning | An environment and a score per attempt | Which action earns the most points | A schedule that learns which order of deliveries costs the fewest kilometres |
Most business uses are supervised: you already hold years of answers in your inbox, your accounts or your support tickets, and those become the examples. That is also why automation with AI usually starts with the data you have, not with a model you buy.
The formula turned around
Classic programming
Rules you writeDataAnswers
Machine learning
DataPast answers
Three ways to learn
Learning from examples with the right answer
- 2,000 old emails, each labelled quote, complaint or invoice
- The model labels every new email
- Needs: examples whose answer you already know
Finding groups in data on its own
- 1,400 customers of a wholesaler, only their orders
- The model finds groups that behave alike
- Needs: lots of data, no answers
Learning by trying and scoring points
- Trying thousands of delivery schedules
- Points for every kilometre saved
- Needs: a place to try and a score
Made-up examples. A model only learns what is in its data, including the mistakes and the bias.
600 emails a week, routed automatically
Say a plumbing wholesaler near Maastricht receives 600 emails a week on one general address, in Dutch, German and English. Someone forwards each to sales, the order desk or accounts: at thirty seconds an email, five hours a week. The wholesaler has 2,000 old emails already filed in the right folder. Those are labelled examples, the raw material of supervised learning.
There are three ways to automate it, at very different prices.
| Approach | What you need | What it costs | Fits when |
|---|---|---|---|
| A fixed rule | A list of words and senders | A few hours to set up | The emails are predictable |
| A language model labels them | A prompt with your three categories and a link to the mailbox | Model cost under 2 euros a year for 31,200 emails on Claude Haiku 5.5 or GPT-6 Luna, plus the build | You want to start fast without training data |
| Train your own model | The 2,000 labelled emails and a developer; scikit-learn is free and open source | Mostly hours, no licence | The data is sensitive or the volume is very large |
The sum behind those 2 euros: 31,200 emails of 300 words at 1.67 tokens per Dutch word is about 15.6 million tokens; at 0.10 dollars per million that is 1.56 dollars. French mail costs a little less, German a little more. Run your own volume through the AI cost calculator. The model is rarely the expensive part; the integration and the testing are.
Planning to tune a language model on your own examples? Check what the providers offer first. OpenAI is winding down its fine-tuning platform and no longer admits new users. For most companies, examples in the prompt or RAG, where the model looks the answer up in your own documents at the moment of the question, do the job.
AI, machine learning and language models: what sits inside what
Language models are deep learning on text: pre-trained on vast amounts of text, then steered with examples and rewards from people. The difference with the models above is scale: you do not train one yourself, you steer it with your request. Which model suits which office task is laid out in ChatGPT vs Claude; for pricing of the consumer plans, see ChatGPT pricing.
When a model beats a rule
Machine learning pays off when three things meet: a task that comes back often, enough examples with the right answer, and errors you can catch. Sorting email, reading invoices, flagging a customer who orders less: yes. Decisions about hiring or credit: those fall under the EU AI Act, whose high-risk rules for such systems apply from 2 December 2027. Check where your use lands with the EU AI Act checker. If a fixed rule is enough, it is cheaper, faster and easier to explain. For repetitive work without judgement, such as copying data between two systems, RPA is often enough, and the ROI of either is easy to calculate up front.
What buyers ask before they start
What is the difference between AI and machine learning?
AI is the whole field of software doing tasks we consider intelligent. Machine learning is the part that learns that skill from data. All machine learning is AI, but not all AI is machine learning: a system of fixed if-then rules is AI that learns nothing.
What is the difference between machine learning and deep learning?
Deep learning is a kind of machine learning built on large neural networks with many layers. It needs far more data and computing power, but it also learns from images, sound and free text. Classic machine learning works well on tables of a few thousand rows, such as orders or invoices.
Is ChatGPT machine learning?
Yes. The model behind ChatGPT was trained with machine learning: first on huge amounts of text, then refined with examples and ratings from people. As a user you no longer train it; you steer it with your request and the context you add.
Does a model keep learning while I use it?
Usually not. A model learns during training and is fixed after that. To improve it, you collect the cases it got wrong, label them and retrain, or add them as examples in the prompt.