Well, adjusting nudges in aggregate but diced in various ways. Measured very much not in aggregate. We’d see positive and negative outcomes roll in over multiple years and want it per identifier (an individual). I’ve heard of companies generating a model per person but we didn’t.

A silly amount of work but honestly lots of value. Experimentation optimising for short term goals (eg upgrade) is such a bad version of this, it’s just all that is possible with most datasets.