Rule of thumb: Demos = as low as 10% reliable (read: we are showing you the successful cases). Product = 99% reliability required and even that is still not great (costly to operate and maintain, sub-par reviews from customers). So the answer to your question depends very much on the use-case and form-factor of the robot. Watching that Neo robot closing a dishwasher was so painful, and even that was still teleop-ed.
Another data point: in academia people thought robot localization was solved back in the 90's. In practice, mislocalizations has been an issue for robotics companies in the real-world until very recently. The number of 9's really matter and is where all the sweat is.