Tom Davenport, one of the gurus of analytics in business, points out that automation is only worthwhile if the task is entirely rote and unchanging for long enough that the machines can be reliably selected and programmed for their entire working life. So far that is not as large a number of tasks as either the "robots are coming for our jobs!" alarmists or the "robot slaves for a better world!" optimists have predicted.

Even in manufacturing, tasks are not as rote as you'd thing. Even Foxconn, which assembles iPhones, has so far found that product cycles are too short to make robots cost-effective for many tasks. 5 years ago their CEO announced he wanted to replace a million assembly-line workers with robots... but so far he still employs a million assembly-line workers and 50K robots.

This is true for now while machine learning is still nascent.

But once the robots are able to learn more complex tasks, this could change quickly.

So the transition will happen slowly until it reaches a turning point of machines' ability to learn.