Optimising Calls to Large Language Models with Uncertainty-Based Two-Tier Selection