Sometimes, for a bounded task; there is no evidence here for universal replacement. Treat model size as a design variable and compare complete systems under the same acceptance rules.
QLoRA established an important memory-efficient adaptation method, but its 2023 experiments do not establish superiority over today’s APIs. LoRA Learns Less and Forgets Less reports a learning-versus-retention trade-off in its coding and mathematics settings. Together these studies justify testing adaptation, not assuming its success.
A recent, narrower example is He, Popa and Danilevsky’s October 2026 query-clarification preprint: its authors report a fine-tuned lightweight adapter performing strongly on an ambiguity task. That is a hypothesis generator for support workflows, not evidence that it can safely resolve the buyer’s tickets. This pass reviewed the abstract, not a reproduction of the experiment.
Ask what would falsify the business case: poorer acceptance on a valuable customer segment, increased escalation, missed latency targets or maintenance that consumes the apparent saving. Write those reversal conditions before training.