In short
Machine learning (ML) is the branch of artificial intelligence in which software learns patterns and relationships from data instead of being explicitly programmed for every situation. An ML model is not fed with rules ("if X, then Y") but with examples, and derives from them how to handle new, unknown cases.
This makes machine learning strong wherever rules are too complex, too numerous or simply unknown: from sales forecasting and fraud detection to speech recognition.
Classification: AI, ML and deep learning
The three terms are often mixed up but form a clear hierarchy. Artificial intelligence is the umbrella term for systems that solve tasks that would normally require human intelligence. Machine learning is today's most important route to it: learning from data. Deep learning, in turn, is a form of machine learning that works with multi-layered neural networks and forms the basis of the current wave of AI; large language models, and thus all of generative AI, are essentially deep learning systems trained on very large amounts of text.
The three types of learning
How a model learns depends on what kind of data and feedback is available. Three basic forms cover most use cases:
| Type of learning | Data basis | Typical task | Example |
|---|---|---|---|
| Supervised learning | Examples with a known correct answer (labels) | Classification, forecasting | Predicting payment defaults, categorizing product images |
| Unsupervised learning | Raw data without labels | Finding structures and groups | Customer segmentation via cluster analysis, anomaly detection |
| Reinforcement learning | Reward signal for actions | Optimizing strategies | Warehouse routes, price control, robotics |
In business practice, supervised learning dominates, because many business questions take the form "tell me for the new case what came out for the old case". The bottleneck is almost always the data basis: enough clean, correctly labeled examples.
How an ML project runs
Regardless of the use case, ML projects follow a recurring pattern:
- Sharpen the problem: which decision is to get better, and how is success measured? Without a measurable goal, ML remains an experiment.
- Review and prepare the data: bring sources together, deal with gaps and outliers, define the target variable. Experience shows this is the largest share of the project time.
- Choose features and model: determine relevant input variables (features) and test model types, from a simple decision tree to a neural network. Simple models first: they are easier to explain and often surprisingly competitive.
- Train and evaluate: the model learns on part of the data and is tested on held-back data. What matters is performance on unseen cases, not on the training data; otherwise there is a risk of overfitting.
- Roll out and monitor: the model is integrated into processes or systems and observed continuously. Data changes over time (data drift), so monitoring and regular retraining are part of operation; the keyword is MLOps.
Where machine learning makes an impact at mid-sized companies
Forecasts: sales, demand and capacity forecasts make purchasing and staff planning more precise than simple comparisons with the previous year. Recommendations: online shops suggest suitable products based on purchase and click behavior, one of the most measurable ML use cases in e-commerce. Quality and maintenance: predictive maintenance recognizes failure patterns in machine data before downtime occurs; image recognition inspects components in production. Risk and fraud: suspicious transactions, payment defaults and return patterns can be flagged early. Text and language: classifying incoming emails, extracting invoice data, searching internal documents, increasingly in combination with LLMs.
An example from practice
A typical scenario from retail: a shop operator wants to reduce returns. From historical order data (product attributes, size information, customer history, payment method), a supervised model learns which orders are highly likely to come back. Orders with a high risk of return then receive targeted measures: more precise sizing information on the product page, adapted recommendations or an advice option in the checkout. The point is: the model does not make individual decisions against customers but systematically improves the points in the process at which returns arise. Demand forecasts in purchasing or failure predictions in production work according to the same pattern.
Limits and typical pitfalls
Data quality beats model choice. No algorithm can rescue incomplete or biased data. Anyone who has historically served only certain customer groups gets models that perpetuate precisely this bias. Correlation is not causation. A model recognizes relationships, not causes; decisions with far-reaching consequences require expert review. Explainability: the more complex the model, the harder it is to understand why it decides as it does, which is relevant for acceptance and, for example in credit or personnel decisions, also legally. Operating costs: a model is not the end of a project but an ongoing operation with monitoring, retraining and responsibilities. Incidentally, a separate route has become established for adapting the style and behavior of pre-trained models: fine-tuning builds on a finished base model instead of training from scratch.
Explainability and regulation
As soon as ML models make or prepare decisions that affect people, two requirements are added that pure forecasting quality does not cover. The first is explainability: in the long term, business departments only accept models if it is traceable which factors drive a forecast, and if it can be corrected when there are objections. Methods of explainable AI (such as feature importance analyses) are therefore standard for models close to decisions, not an optional extra. The second is regulation: the EU AI Act classifies AI systems by risk; applications in areas such as recruitment, lending or critical infrastructure may count as high-risk AI systems and then trigger documentation, oversight and quality obligations. The typical sales forecast or product recommendation does not fall under this, but anyone planning ML applications should check the risk classification early, because it influences architecture and process decisions that can only be retrofitted at great expense later.
Build or buy: own models vs ready-made AI services
Not every ML application requires a model of its own. In practice there are three levels of expansion, and the right choice saves more money than any model optimization.
Level 1: use ML as a product feature. Many tools that companies already use contain pre-trained models: the product recommendation in the shop system, anomaly detection in monitoring, text recognition in document management. Here there is nothing to train, only to configure. For standard cases, this is almost always the most economical route.
Level 2: connect ready-made models via APIs. Translation, speech and image recognition, document extraction and language models can be consumed as a service and fed with your own data. The effort lies in integration, not in building the model; a typical project for weeks, not months.
Level 3: your own model on your own data. Only when the question is company-specific (the returns forecast on your own range, the failure prediction on your own machine data) is a specially trained model worthwhile. Here the project steps described above apply in full, including operation and monitoring.
The decision resembles the classic make-or-buy question for software: if the capability differentiates your own business, that speaks for building it yourself; if it is standard, almost everything speaks for buying it in.
What makes ML projects laborious
Reliable effort estimates depend less on the algorithm than on three factors. First, the data situation: is the necessary data available completely, accessibly and in usable quality, or does it first have to be brought together from ERP, shop and Excel islands? Data preparation regularly dominates the project effort. Second, the degree of integration: a model whose forecast ends up in a dashboard is productive quickly; one that intervenes automatically in ordering processes needs interfaces, error handling and approval logic. Third, the operating requirements: how often does it need retraining, who monitors forecast quality, what happens in the event of drift? A model without an operating concept is a proof of concept, not a product.
Why ML projects fail:
- There is no measurable target variable; "doing something with data" is not a project goal.
- Data quality is only checked after the project has started and does not support the question.
- The model achieves good metrics, but nobody builds it into everyday work; what is missing is the process, not the technology.
- After go-live, nobody feels responsible; the model quietly becomes outdated and after a year delivers worse decisions than gut feeling.
Related terms explained briefly
Feature: an input variable of the model, such as basket value, season or customer age; the selection and preparation of features often decides more about quality than the algorithm. Label: the known correct answer in the training data, e.g. "was returned: yes/no". Training vs inference: training is learning on historical data (compute-intensive, infrequent); inference is application to new cases (ongoing, cheap). Overfitting: the model memorizes the training data instead of the pattern, recognisable by strong performance in training and weak performance on new data. Data drift: the gradual change in real-world data compared with the training state, the main reason why models without monitoring get worse over time. Neural network: a type of model made up of many connected computing units in layers; "deep" networks with many layers are the foundation of deep learning and thus of current generative AI.
Frequently asked questions about machine learning
Does machine learning always need huge amounts of data?
No. Many operational forecasting tasks work with a few thousand clean data records. Huge amounts of data are mainly needed by deep learning models trained from scratch; mid-sized companies usually use pre-trained models or classic methods on their own data instead.
What is the difference between ML and generative AI?
Generative AI is an application of machine learning: models that generate new content. Classic ML in companies, by contrast, usually answers decision and forecasting questions with numbers, categories or scores. Both are based on the same learning principles.
What is the best way for a company to start?
With a decision that is made frequently, has a measurable impact and for which data is already available, such as demand forecasting or lead prioritization. A small, measurable pilot almost always beats the big platform project: it provides proof within weeks of whether the data situation is sufficient and builds the internal trust without which no model ever reaches regular operation.
Further definition: machine learning on Wikipedia.