There’s a great episode of 30 Rock where Jack (the head of the network) replaces all of the pages with computers. Naturally, everything is going great until Jack has to ship a package to his boss. The package doesn’t make it because he typed in the 6th floor instead of the 66th. The computer system proves he’s wrong but Jack pulls a page in to take the blame and throws out the computers. It’s a funny episode that contains a large amount of truth.
One of the advantages of humans being involved in a process is that they can check for other human errors that are difficult to detect (such as someone putting in the 6th floor instead of the 66th floor). AI is a black box to all but a small portion of the population. It follows the “And then a miracle occurs” model. Put something in and then the outcome magically happens. No idea how, but it does.
Black box processes require a high degree of trust. They may be fine for rote activities but do you really want one designing your 10 million square foot real estate portfolio? How will you know if the variables are right? How will you be certain it didn’t miss something? Or forget to factor in politics. If you have to triple check all of the outcomes manually anyway, what good is the AI?
Overcoming this lack of trust is going to be very difficult for most AI. There is one class of AI that it will work really well though – AI that speeds up the decision process without actually making a recommendation. AI has a great ability to support existing processes and make them more robust. Replacement of process will be a bit further down the road after a lot more trust is built.