I look at p95 before I look at the average. A workflow that is fast most of the time and slow once in twenty is the one that shows up in support.
On the HR APIs, SQL tuning, Redis caching and moving long work onto background jobs cut p95 by about 40% on the key workflows. The cache key includes the tenant and the query shape, so a speed-up does not become a data leak.
Earlier, on the consultancy systems, indexing, query refactors and read/write separation cut slow-query volume by over 70% on the paths we measured. Redis covered the reads that did not need a fresh join.
The order of work stays the same. Measure the slow statement, fix the index or the query, then cache what is actually hot.