thoughts on when to use expensive models vs. cheap ones

In the slides we looked at today, it seemed like the consensus was expensive models can be run on a small subset of records for QA-like tests, rather than for processing an entire document collection.

Do we have other thoughts to add to or contrast with this idea?