At Roku Pay the partner support queue was drowning. The instinct in that situation is to add people to answer faster. That treats the symptom. The volume was the signal, and the signal said the process was making people ask.
I looked at what the inbound actually was. Most of it clustered into a handful of question types, and almost all of those questions existed because information partners needed wasn't where they'd look for it, or a step that could be self-serve required a human. The queue wasn't a staffing gap. It was a design defect, repeated thousands of times.
So I redesigned the points where partners got stuck. Put the answers where the questions happened. Made the self-serve steps actually self-serve. Removed the human from loops that didn't need one. Inbound dropped by about 90 percent, and the requests that remained were the real ones that needed judgment.
The general lesson: when you're flooded with the same request over and over, don't get faster at answering it. Find the thing upstream that keeps generating it and remove that.