The better argument, I think, is "Do we need that much optimization?"
Does the trial and error performed by a casual switch user cost us so much that spending the money on custom layout light panels makes sense? I think only in a very small set of applications.
To counterpoint your second paragraph: The existing wall switches are much more than a poor first time experience and require more cognition. When using the "efficient" vertical wall switch panel, you already have to import your full object recognition library without errors. You then have to call up your list of mappings between rooms and switch position on the panel so you can flip the correct one.
With the map, you only need to load your scheme of the office into active memory to select the appropriate switch. Further, the map provides error checking for your cognitive model, confirming / reinforcing the true layout (in a complex map, a user is bound to forget a room or take longer to recall it's location without the external map).
Anecdotally, I lived in my family's third house for a year and a half, and never completely memorized the functions of the 6 switch panel in our front hall.
Imagine an office, where different panels control different light banks depending on where they are in that large office space. The common user (average person in the office seldom turns on or off various things) will need to hunt for the right switch.