Lets look at a few cases where AI can do better assuming the machine has all the relevant data.
This is a real scenario in every industry. Lets take Credit Cards. Detecting a fraud, in a real time, is very critical to minimise financial damage. It requires going through millions of credit card transactions to probably uncover 10 frauds. The scale of the problem is so huge that a human cannot solve the problem in a time bound manner, so we can take corrective action. Moreover, fraud detection requires looking at each transaction from multiple angles (demographics, location, time, amount, frequency and many more variables). As the number of variables to analyse grows in number, the capacity of humans to analyze millions of transactions across thousands of variables diminishes.
Lots of cancerous tumors and other diseases are identified by expert doctors through Biomedical Imaging. But the ability to spot a tumor depends on the expertise of the doctor and his/her experience. In such situations, and as the cases grow in number, it is a better option to train a machine to understand the nuances of disease detection(by learning) and replace doctors. As more data is fed to the machine it also starts detecting outlier cases that only few doctors can detect. But the advantage here is that the machine is faster, cheaper and removes the human component.