Machine Learning

Short answer: Machine learning is a way of building software that learns from examples instead of following rules that someone writes by hand. A team trains a model on past data, such as orders, transactions or customer behavior, and then uses it to predict an outcome for new cases. The output is usually a score, a category or a forecast. Generative AI is one type of machine learning. This entry covers predictive machine learning, which estimates an outcome instead of creating new content.

Common business uses include forecasting demand, spotting fraud or unusual activity, predicting which customers will leave, scoring sales leads, setting prices and recommending products. Each one needs enough past data that is accurate, consistently defined and similar to future conditions. A model trained on one market, season or customer mix may perform poorly when those conditions change. For many smaller problems, simple rules or a good report give most of the value at lower cost and risk. Machine learning is the better choice when the pattern is too complex, or the volume too large, for rules to handle well.

First, the team decides what outcome to predict and gathers past examples where the outcome is known. It cleans the data and splits it in two. The team trains the model on one part and tests it on the other part, which the model has not seen. It compares several methods on the test data, using measures such as accuracy or forecast error. The team then connects the chosen model to an existing system or report, so it scores new records automatically. After launch, someone watches its accuracy and retrains the model when customer behavior, products or markets shift.

A subscription business has 10,000 customers and loses about 300 of them each month. It trains a model on two years of usage, support and billing history to score how likely each customer is to cancel. Each month, the account team calls the 500 customers with the highest scores. Without the calls, about 45 of those 500 would leave. If the calls keep one in five of them, the business keeps about 9 extra customers a month. It can then compare that revenue with the cost of making 500 calls.

Machine learning matters when a decision repeats across many cases and the pattern behind the right answer is hard to describe with rules. Before approving a model, leaders should name the decision it will change. They should also agree how to compare its accuracy with the current method, and who owns the model after launch. Teams often stop using models that have no clear owner or purpose.

A model is only as good as its training data, including any bias in it. Poor or incomplete data produces confident but misleading predictions. Some uses, such as credit, hiring or insurance decisions, may fall under rules on discrimination or on explaining decisions, depending on the country. Models get worse as conditions change, so they need ongoing monitoring and a maintenance budget. A prediction is an estimate with a margin of error, so treat it as one input to a decision.