“Us sticking to 8 bit bytes” is a consequence of having preferred BCD to octal in the past, so the causation is reversed (“12-bit words would have made a lot more sense if it was, say, octal.”) [Edited: actually 12 bit words would make sense in either case, as it's three BCD digits or four octal digits]

The Intel 4004 used four bits to manipulate a single BCD digit. The 8086 had BCD instructions. There were many reasons for preferring BCD when designing computer architectures, though my favourite which was already becoming less relevant at 8086-time was that it meant a full column on a punchcard wouldn’t be “all holes” and reduced the likelihood of the cards tearing.