The amount of training data varies dramatically from application to application. There might be successful Machine Learning solutions that only need 100 examples. However, for many use cases, we need tens of thousands of examples to find the expected results.
In general, there is a tradeoff between the amount training data your company has, the amount of error in your application that you are willing to tolerate, and the difficulty of the problem you are trying to solve. If you have a fairly easy problem and you are willing to tolerate a certain amount of error (e.g., you only need a correct prediction three out of four times), then you will not need very much data. Nevertheless, harder problems and problems that require high accuracy (e.g., medical diagnoses) require more data.