When planning for the scale of a new system, do you plan for the worst case or best case? Which method would create a more resilient system?

Put another way, if a service can only respond to 93% if requests, I wouldn't call it a very well architected service. EVs have a long way to go for many people.