What is machine learning, how does it differ from classical software, and why does it matter if you are not a techie?
In classical software you program rules: “if amount > 10,000, ask manager approval”. In machine learning you teach a system to recognise patterns from examples, instead of writing the rules out explicitly. That sounds subtle, but the consequences are enormous.
The difference in one sentence
Classical software follows rules you wrote. Machine learning learns rules you don’t have to write.
The three flavours
Supervised learning
You feed the model examples together with the correct answer. Tens of thousands of invoices labelled “fraud” or “not fraud”. Hundreds of thousands of emails labelled “spam” or “not spam”. The model learns the pattern and can then classify new cases. This is by far the most common type in business.
Unsupervised learning
No labels. The model finds structure in the data itself - for example groups of similar customers (segmentation). Useful for exploration and as a stepping stone to more targeted analyses.
Reinforcement learning
The model learns by trying things and getting feedback. This sits behind game AI, robotics and the fine-tuning of modern language models (RLHF). You will rarely build this yourself in office work.
What is a model, really?
A model is not a crystal ball. It is a collection of numbers - often millions or billions - that together form a function: input in, prediction out. During training those numbers are adjusted until the predictions on known examples are good enough. Then the model can handle new input.
More and better data means a better model. But more data hits diminishing returns. At some point you gain nothing without also improving the architecture or features.
Where is it actually used?
- Fraud detection in banking - scoring thousands of transactions per second.
- Recommendation systems in e-commerce, streaming and news.
- Demand and inventory forecasting in retail and logistics.
- Image analysis in healthcare (radiology, pathology) - usually supportive, not decisive.
- Text classification: routing customer emails, triaging legal documents.
What are the pitfalls?
A model learns what is in the data. If your data is skewed, your model is skewed (see our article on bias). If your data is old, your model predicts yesterday’s world. And if your objective is wrongly defined, the model will brilliantly solve the wrong problem.
A famous example: a hospital trained a model to triage pneumonia patients. The model recommended sending asthmatic patients home - because asthmatic patients in the data died less often. Not because their pneumonia was milder, but because they were immediately given intensive care. The model learned the wrong pattern.
Four questions you can ask today
- 1What problem are we actually solving - and is that exactly what we are letting the model optimise?
- 2Where does the data come from, and who is in it (and who is not)?
- 3How do we measure “good”? Per subgroup, or only on average?
- 4What do we do when the model is wrong - who checks, who corrects?
Whoever asks these four questions routinely prevents more damage than a team of consultants. Our AI course practises this with cases from various sectors - no maths required.
