It's good to stay away from the hype as much as possible. In what applications do you think using ML is not necessary, or worse still, a bad idea?
Most often, "machine learning" refers specifically to supervised learning. Generally, such technology requires an historic set of examples of sufficient volume and quality (what this means varies greatly from problem to problem) from which to learn. Some problems do not come with a reasonable set of training data, and none is forthcoming: these situations are a poor choice for machine learning. Each of these factors makes example data less desirable: candidate explanatory variables which are few in number or highly redundant, missing values and poor measurement. All of the challenges one faces in traditional statistical data collection and sampling apply.<p>Some problems are not learning problems at all. Retail companies with poor management and consequent poor service, for instance, need machine learning less than they need competent management.