Machine learning is software that gets better at a task from examples rather than from rules a person wrote. It is the technique behind most of what is marketed as AI, and the distinction is worth keeping because it tells you what a system needs to work: data.
In screens and video the honest applications are narrow and useful. Estimating audience counts from a sensor, predicting which creative performs in which context, ranking what a viewer sees next. All learn from accumulated outcomes.
The caveat is that a model is only as good as the data it learned from, and retail data carries the biases of the situation it was collected in. A model trained on one store's traffic will not transfer to a different format, and one trained mostly on a majority group performs worse for everyone else, which in audience estimation means systematically misreading part of your customers.
The practical question for any proposal: what did it learn from, does it keep learning in your environment, and can somebody explain a specific decision after the fact. If none of those have answers, it is a fixed model sold as an adaptive one.