If a 12-bit BCD number is written out as

  ABCD  EFGH  IJKL  [a=msb]

it may be converted to binary by multiplying ABCD*100 and EFGH*10,
and then adding those products together with IJKL.  Since 10=8+2
and 100=64+32+8, it's possible to rewrite the expression as the sum
of four numbers:

ABCDABCD--
-ABCD-IJKL
---EFGH---
-----EFGH-

This sum may be performed by using a variety of means; the most ob-
vious would be two use two carry-save adders followed by a full carry
adder; this may be overkill given that only one bit position (bit 3)
has terms on all four lines, but should yield reasonable results none-
theless.

Note that if a sequential-logic (rather than combinational) approach is
acceptable, that may simplify things somewhat.  The design, however,
would be affected by timing and sequencing requrements; since I don't
know what the requirements are here I wouldn't know where to begin (other
than suggesting that this might be a good application for a PIC...)