Two people understand billing. One is on a plane.

Every team has a subsystem only two people truly understand, and a quiet bet that they are never both unreachable at once. AI is shortening that list without telling you. A Microsoft Research survey of 319 knowledge workers found that the more people accepted AI output as-is, the less they engaged in the effort of comprehending and double-checking the work. The study.

It is self-reported, so read it as a signal, not a measurement. But run the signal forward across a team: if individuals understand less of what ships under their name, then for any given subsystem, fewer people can actually reason about it. The bus factor was already thin. Passive acceptance shaves it.

Comprehension debt: the gap between what gets shipped and what the team can still reason about.

This is the version of the debt that sizes up to your VP cleanly. It is not "morale" or "tech debt." It is a single number: how many people can safely change the thing that takes payments. When that number is one, your roadmap has a dependency on one person's calendar, and you will discover it the week they take leave. Ask, for your two or three load-bearing systems, who the second name is. If you cannot answer, that is the work.

One for your next review:

A scheduled report runs at 9pm every night and lands correctly. Twice a year, for about a week, it runs at 8pm or 10pm, and someone files a bug. What is going on?

Does daylight saving come up before anyone opens the scheduler? That is really all you are checking. The reasoning you want is short: ask whether the job is pinned to an absolute instant (UTC or a fixed offset) or to a real time zone, because only the first one drifts when the clocks move. If nobody can say why the drift shows up for a week or so and then heals on its own, the knowledge of how this schedule works lives in exactly one head, and you have just found whose.