Machine Learning.

According to Google, machine learning (ML) refers to the process of training software—known as a model—to make useful predictions based on data or to generate content such as text, images, or videos. Instead of manually programming every rule, the model learns the relationships from examples. Google explains this using a rain forecast as an example: Instead of calculating the physics of the atmosphere, a model learns from large amounts of weather data which patterns lead to how much rain. In business, ML is used in spam filters, sales forecasts, and product recommendations, as well as in every LLM.

How Machine Learning Works

Google distinguishes between four ways in which ML systems learn. In supervised learning, the model is fed data with known correct answers. This results in regression models that predict numbers—such as travel time—and classification models that assign items to a category, such as spam or not spam. In unsupervised learning, the model searches for clusters in unlabeled data on its own.

In reinforcement learning, according to Google, a model receives rewards or penalties for actions taken in an environment and uses this to develop a strategy. The fourth type is generative AI, which generates new content from inputs. Many of these models are based on deep learning.

What You Should Keep in Mind at Work

Start without ML as long as you don't have any data. Google's guide, "Rules of Machine Learning," begins with exactly this advice. According to Google, ML requires data, and simple rules often deliver a good portion of the benefits right from the start.

First, determine how you will measure success. Google's second rule, in essence, is to design and implement metrics first. Without data from before the model was implemented, it will be impossible to show later whether it has improved anything.

Switch to ML when your rules become too complicated. According to Google, a simple rule is a good place to start, whereas a complex set of rules is difficult to maintain. If data is available, Google says a trained model is easier to update.

Check whether a ready-made model is sufficient. For text-based tasks such as summarizing, categorizing, or answering questions, you usually don’t need to train your own model, because language models can already do this with a good prompt. In projects, we’ve found that developing your own ML model is particularly worthwhile when dealing with numerical data from your systems—for example, when making forecasts for demand or capacity utilization.

Frequently Asked Questions About Machine Learning

Is machine learning the same as AI? No, ML is a subset of AI. It encompasses the processes in which software learns from data rather than simply following fixed rules written by humans.

How much data do you need? There's no set number—it depends on the task. What's more important is that the data reflects the real-world cases for which the model will later make predictions, and that the correct answers are accurately recorded.

Does an ML model continue to learn on its own while in operation? Usually not. A model is trained, tested, and then deployed. In its guidelines, Google advises planning from the outset to release new model versions on a regular basis.

In "The 5 Processes Where AI Pays Off First," we outline where machine learning really pays off in your workflows. For implementation, check out " AI for Businesses."