The limitation of PRQL is that it only does SELECTs, by design. If you want to insert/update/delete data, you're back to SQL.
That means that your team's data scientist might give you a query written in PRQL, but if you want to actually incorporate it into the data pipeline, you'll need to translate it into SQL.
I wish that PRQL would support at least a limited ability to insert -- for example, maybe just the case of inserting into a new temp table. No update or ON CONFLICT logic to worry about. It could look like this:
from tracks
filter artist == "Bob Marley"
save bob_marley_tmp