Bertrand Meyer suggested you could split most code into decisions and actions and that makes the code more maintainable. My own experiences seem to support this.

Machine learning only (partly) addresses one of those categories. There’s still plenty of work to do even if AI decision making is less buggy than human code.