You ran one experiment but sliced the results twenty ways and found two significant segments. What do you tell the team?

Treat those as hypotheses, not findings. At alpha 0.05, testing twenty slices gives roughly a 64 percent chance of at least one false positive even when nothing is real, so two hits is about what pure noise produces. Apply a correction such as Benjamini Hochberg, or better, pre register the segments you care about and confirm any surprise in a fresh test.

I would tell them we found two leads, not two results. With twenty independent tests at 0.05, the chance of at least one false positive under a true null is about 64 percent, and the expected count is one. So two is squarely inside what noise produces. Then I would apply Benjamini Hochberg rather than Bonferroni, because Bonferroni across twenty tests is brutally conservative and would probably kill a real effect along with the noise; false discovery rate control is the right tool when you are screening. Usually one or both drop out. If something survives, I still treat it as a hypothesis and say so: the honest next step is a confirmatory test targeting that segment specifically, with the segment declared before we look. I have seen this go the other way, where a team shipped to a segment based on a sliced result and the effect vanished at scale, and the cost was not just the wasted feature, it was that nobody trusted the experiment platform for a quarter afterward.