The 9 SQL queries against a single 8-row table suggest minimal database overhead; compute concentration is almost certainly in LLM synthesis rather than data retrieval. This workload is well-suited for scheduled runs given the modest query volume and small dataset size, though on-demand execution remains viable if latency tolerance is under a few seconds. To reduce future cost, prioritize prompt optimization and result caching—reusing LLM outputs for identical or near-identical input rows will yield the largest efficiency gain, since the bottleneck is synthesis rather than data access.