Machine learning is a branch of artificial intelligence in which a program is not given its rules but derives them from examples. Instead of a set of conditions somebody writes down by hand, a model emerges that finds patterns in historical data and applies them to new cases. The difference from classic programming is not the computing power, it is where the logic comes from.
That sounds abstract, but it has long been everywhere: the spam filter in your inbox, your bank’s fraud detection, the product suggestions in a shop, the demand forecast in a stock system.
The three learning paradigms
How a model learns depends on what the data contains.
Supervised learning
Every example comes with the correct answer: thousands of emails labelled “spam” or “not spam”, quotes labelled “won” or “lost”. The model learns the relationship between input and answer. By far the largest share of business applications sits here, because the benefit is measurable: you know the truth and can count how often the model gets it wrong.
Two task types dominate. In classification the answer is a category (invoice, delivery note, reminder). In regression it is a number, such as the expected delivery time in days.
Unsupervised learning
Here there are no given answers, only data. The model looks for structure on its own: groups of customers who behave alike, transactions outside the usual range. Typical tasks are segmentation and anomaly detection. The catch: nobody but you decides whether the result is useful. There is no correct answer to check against.
Reinforcement learning
An agent tries out actions and gets feedback in the form of reward or penalty. Over many rounds a strategy emerges. The approach became famous through games and robotics; in everyday business you meet it rarely, since it needs an environment where huge numbers of attempts can run without harm.
| Paradigm | What the data contains | Typical question |
|---|---|---|
| Supervised | Examples plus the correct answer | “Which category does this case belong to?” |
| Unsupervised | Examples only, no answers | “Which groups are hiding in my customer base?” |
| Reinforcement | Reward for good decisions | “Which action pays off in the long run?” |
How an ML project runs
The sequence looks much the same across industries. It almost never starts with an algorithm.
The question first, the data second
At the beginning sits a decision somebody currently makes by hand, and the question of what a better one would be worth. Without that anchor there is no judging later whether the model is any good. Only then do you look at which data describes that decision.
Preparation
This step eats the largest part of the project time. Data sits in several systems, in different formats, with gaps and duplicates, and has to be merged, cleaned and translated into features a model can work with.
Training and testing
The data is split. The model learns on one part, a second part tunes the settings, and a third stays untouched until the end to answer the only genuinely interesting question: how good is the model on cases it has never seen? Soften that separation and you are measuring yourself, which produces results that do not survive production.
Going live and watching
A model is software and deserves the same care: interfaces, tests, logging, a way back. It also has a quirk classic applications lack. Models age when reality changes. A new product, a new market, a modified form, and the predictions drift. Without monitoring nobody notices.
Where mid-sized companies use machine learning
The worthwhile cases are rarely spectacular. Usually they are high volumes of small individual decisions.
- Understanding documents: sorting and reading incoming invoices, delivery notes and forms automatically.
- Forecasts: sales, material demand, capacity. Any number somebody currently estimates from experience is a candidate.
- Maintenance: deriving from sensor data when a machine needs attention instead of servicing on a fixed interval.
- Quality control: checking production images for defects.
- Prioritisation: routing support tickets, ranking leads, steering dunning runs.
Why data quality decides everything
A model can only reproduce what sits in the data. If a systematic error sits there, the model learns it and applies it consistently to every new case. That is where the risk lies: human sloppiness is randomly distributed, machine sloppiness is not.
Four traps come up again and again:
- Wrong labels: if the training examples were sorted by different people using different standards, the model learns the disagreement instead of the rule.
- Giveaway features: if the training data contains a column that does not exist yet at prediction time, the results look brilliant and collapse in production.
- Bias: if the data covers only one slice of reality, the model is valid only for that slice.
- Age: data from a different market situation describes a world that no longer exists.
More data helps with none of these points. Better data helps with all of them.
How it relates to deep learning and AI
Artificial intelligence is the umbrella term, machine learning the practically most important part of it, and deep learning in turn a branch of machine learning built on multi-layer neural networks. A large language model is a trained model too, just one with an enormous amount of text as its material. For tabular business data, simpler methods are often the better choice: faster to train, easier to explain, cheaper to run.
Frequently asked questions about machine learning
How much data do I need? There is no universal number, it depends on the task. More useful than the total volume is how many examples you have for each case the model must tell apart. A category that appears twelve times in the training data will not be recognised reliably, however large the overall dataset is.
Is a language model like ChatGPT machine learning as well? Yes. A language model is a very large model trained on text. What differs from a classic ML project is the approach: you train nothing yourself, you use a finished model and steer it through context, for instance with prompt engineering or retrieval augmented generation.
What do ML projects usually fail on? On vague questions and on data that turns out to be unusable. The algorithm is rarely the problem. A small, honest experiment on real data early saves expensive disappointment.
Do I need specialists in house for this? Not necessarily to start, more likely to operate. A model nobody watches quietly loses quality. At least one person should understand what it does and how to spot it drifting.
Conclusion
Machine learning is not a tool for everything. It is one for recurring decisions with plenty of examples from the past. Success depends less on the method than on the question you want answered and the data you have. Whether the effort is worth it for a specific case at all is something we sort out beforehand in AI consulting for companies. How we build software, and where we use AI while doing it, is described on our page about custom software development. What works in practice and what does not is written up in the article AI in software development. For anything beyond that there is a free consultation.