None of these points are big obstacles. Airports may be "busy", but the system is already set up such that pilots can perform a whole flight under next-to-zero visibility. It's mainly an issue of airspace control.

En route diversions are easy to handle by a rudimentary flight crew and communication equipment. The computer just needs to know the new destination and off it goes. It could even find one on its own, if all the crew is knocked out and all communications have failed.

That being said, people won't trust an AI enough to make this a reality soon. Even though a system more secure than an average pilot is certainly conceivable.

In all three of those circumstances the human pilots are already relays, and are not really not necessary as executors.

Consider the current ATC data links such as CPDCLC. The pilots read what appears on the screen and press ACCEPT to load the ATC commands into the flight management computer which instructs the autopilot.

Pilots don't land in thick fog, autoland has been doing that for 50 years. If autoland self-aborts they divert to another airport and follow the ILS needles to land there...

TCAS negotiates collision-avoidance in busy airspace and generates emergency action commands for the pilots.

Military drones are also at least partially overseen by a pilot as well.

When we've mastered the automation of cars, then we can start turning our eyes to be able to handle planes as well. Just like cars, automating clear and calm days is relatively simple; but most days are neither clear nor calm.