They measure different populations and are easy to misread against each other. 81.1% of these decisions are automatic and finish in minutes, so the 90th percentile of everything lands only a little way into the reviewed cases rather than at their middle — which means it can fall below the human median entirely when the reviewed share is small. The two groups also overlap rather than sitting one after the other, because every figure here is measured from when the applicant applied, so the vendor's own latency is inside all of them: an automatic decision on an application the vendor sat on for hours is slower than a reviewed one the vendor answered in a minute. So the p90 answers “how long does a slow case take?” across the whole population, while the human median answers “how long does a reviewed case take?”, and which of the two is larger moves with the mix rather than with how fast anyone is working.