No child learns to ride a bike by being taught the physics behind it. They get on, fall off, and try again. After a few hours, and a few scrapes, they get the hang of it. The child knows intuitively how to keep their balance, without ever having seen a single formula.
AI learns in a similar way. Not through rules that someone programmes into it, but through examples. Lots of examples. And through mistakes.
What happens during training
Before a model can even begin to learn, someone has to prepare and evaluate the training data, which means manually classifying thousands of customer enquiries: cancellations, complaints, meter readings, general enquiries. This takes time and money, and is often underestimated in practice.
Only then does the actual training begin. The model analyses the classified data, makes a prediction, compares it with the correct answer and corrects itself. This process is repeated millions of times. Gradually, the model learns which features are relevant and which are not. Ultimately, it is able to assess new, unknown queries.
But even then, the model doesn’t simply run along quietly on its own. It needs a person to monitor it, evaluate the results and retrain it if necessary. Not because the model is bad, but because no model is aware of all the exceptions that arise in everyday life. What’s more, reality is constantly changing: new products, new processes, new formulations. Even after all that training, AI will still make mistakes.
Children sometimes fall off their bikes, even long after they’ve learnt to ride. That’s just part of it. If you can cope with it, you’ll reach your destination faster than on foot.
More data isn't necessarily better
A common assumption is that the more data there is, the better the result. This is true, but only if the data is of good quality. The training data must reflect reality, i.e. it must capture the variety of situations the model will later be confronted with. A model that has only been trained on customer enquiries from the summer may perform differently in winter.
I learnt to ride a bike on an old one without gears and with a coaster brake. I got the hang of it, but even today I still can’t really get to grips with modern gears.
In the same way, AI learns incorrect patterns from poor or biased data. Data quality is therefore not a technical detail; it determines how accurate the results are in the real world.
What exactly is a model?
A model is a mathematical construct that recognises patterns in data and uses them to make predictions. It is not programmed with fixed rules; instead, it develops its own rules from examples. This is precisely what sets it apart from traditional software.
The danger of overlearning
There is another phenomenon that often comes as a surprise in practice: overfitting. A model that is trained on the same data for too long ends up memorising the data rather than understanding the underlying patterns.
The result: the model performs excellently on the training data, but fails when faced with new, real-world data. In practice, you often only realise this when it’s too late.
When is a model suitable for practical use?
A model is not simply good or bad; it is more or less suitable for a particular purpose.
The error tolerance must be appropriate for the application. A detection rate of 85% may be excellent for an initial prioritisation of customer enquiries, but it might be too low for an automated decision without human oversight. And if a process allows for no fault tolerance at all, an AI model is simply the wrong tool, as no model is ever 100% accurate. In such cases, it is worth considering other automation approaches.
The same principle, everywhere
Whether a model recognises images, identifies chatters, generates text or detects anomalies in usage data, the basic principle is the same. Show examples. Correct errors. Repeat.
What differs is the architecture of the model, the type of data, and the amount of care that has gone into preparing the training data. The latter is the factor that, in practice, has the greatest effect on success or failure.
A child on a bike needs a good bike, a bit of time, and someone to encourage them every time they fall off. AI needs good data, sufficient computing power, and someone to ask the right questions.