Do you have a lot of unmapped reads?
Sometimes, there may really be something else contributing to true RNA-Seq sequencing in your sample (where there may actually be benefit to using a joint reference). For example, if you have a virus infection and there are a lot of virus reads (where you may want to use a human+virus reference).
However, I think there is always some amount of cross-contamination and other factors that are going to keep all 100% of your reads from being important to understand. For example, if you have filtered things like adapters and other non-biological sequences, and you still have <2% unexplained reads, I think limits to the genome assembly, purity of de-muliplexed samples, etc. could be a factor.
Not sure if this is relevant to you, but I think there are currently top PhiX BLAST hits that aren't accurate (because someone with a 16S study didn't understand that you can get PhiX sequence in your de-multiplexed samples, even though it wasn't supposed to be there, probably with a single-barcode):
C: Calling Single-Barcode Samples from Mixed Runs as Dual-Barcode Samples | Possibl
In other words, I think it makes a difference whether you have 30% unaligned reads or 5% unaligned reads.