What it costs to ask a colleague

Consider an engineer partway through a change who is unsure whether an approval step still applies to it. She could put the question to the colleague who wrote the step, or she could type it into a chat window and read the answer there. The first of those costs her a moment and leaves behind a small durable record that on this particular point she did not know.

Questions that used to go to a colleague or a manager now go to an assistant often enough that nobody is disputing the direction. The figures attached to it will not bear weight, since the ones in circulation were published by firms that sell assistants or sell training in using them, and none of the releases says who was asked or out of how many.

The convenience reading deserves more than a nod, because much of this is convenience, and nothing in it requires anybody to be frightened of their colleagues. An assistant answers at eleven at night, is never in a meeting, never needs the background explained again, and will take the same question again without comment. The gloomier telling, that junior staff stop learning the trade by asking the people who hold it, is plausible too, and is also a story about the assistant rather than about the route the traffic left, which is the question worth putting.

The condition on asking at all

Amy Edmondson’s 1999 study of work teams in a furniture manufacturer named the thing psychological safety. Put back in terms of the engineer and her approval step, whether she takes the question to a person turns on what she expects that to cost her. Edmondson is careful about what the study can carry. Her measure could not be cleanly told apart from ordinary trust, every reading came out of one instrument answered by the same people, and the design settles nothing about which way the cause runs. The study covers a single firm, fifty-one of its fifty-three teams, read at one point in time. A 2017 meta-analysis by Frazier and colleagues aggregates 136 independent samples covering more than 22,000 people. A 2024 preprint reviewing 217 studies of how the construct gets measured finds the instruments usable but too inconsistent across studies to support quoting one effect size. The direction of the relationship is well supported across settings and outcomes, and the size of it is still in dispute among the people who measure it.

Columbia and the imagery requests

The Columbia Accident Investigation Board left the fullest public account of what becomes of a question that nobody with standing converts. Foam struck the orbiter’s wing 82 seconds after launch, and engineers wanted Department of Defense imagery of the wing to see what the strike had done. The Board found three imagery requests made outside the official channels for such a request, along with several points at which a manager could have pursued the matter and none did. One of the three did reach Defense personnel who could have supplied the imagery, and NASA canceled it ninety minutes later. The reconstruction of why is the part that transfers: the chair of the Mission Management Team, having heard informally that there was a request for imagery, called three other members of the team and established that none of them knew of a requirement for it.

The Board’s report treats that unevenness as a cause of the accident rather than a detail of it, which is the part most retellings drop in favor of the foam. People without the standing to convert a request discounted their own concerns before anybody above them got the chance to. So a request and a requirement were different objects, and only the second obliged anybody. Seven people died with three imagery requests made and none of them converted, and the machinery available for acting on a concern only ever recognized the converted kind.

Assistant traffic draws on the same population of questions that people already had and never turned into anything anybody else could see, and it now has a route that converts nothing and asks nothing of the asker, so those questions leave no mark in the place where converting them would have had to happen.

What a manager can do with it

The diversion matters to a manager even though almost none of it can be read. Assistant conversations are private, often run on tools the employer has no telemetry over, and the privacy that makes them unreadable is the same property that moved the questions onto them. A manager who announces that assistant use will be monitored for insight into team concerns should expect the traffic to move again, since the announcement turns a free route into an expensive one.

What survives all of that is the fact of the move. A route has opened that asks nothing about anybody’s position, and the questions have gone to it, which is as close as anyone gets to a reading of what the older route had been charging for the same question.

Two things are available to somebody who wants the questions back in the room. The first is to ask in the open what people have been taking elsewhere. It is free to adopt, it fits inside the next team meeting, and it will under-report for exactly the reason the questions left, since the asking still happens in front of everyone. The second is to shift the cost off the individual and onto a date in the calendar, which is what a pre-mortem does: assume the project has already failed, then work backward through the causes. The device sits on this firm’s Insights page under Gary Klein’s name, and what it changes is that the pessimism is on the agenda rather than left to whoever is willing to voice it.

Of the two, the second is the one worth putting in place, and the argument for it concerns where the cost sits rather than anybody’s courage. An earlier piece on what an AI appointment leaves unanswered made the structural form of the same point: creating a capability says nothing about what was funded to carry its conclusions into the units that have to act on them. A standing pre-mortem is funded machinery with a date and an owner attached to it, and an instruction to be candid carries neither. On a calendar the difference is visible, in whether the recurring forum whose stated job is to produce bad news has a named owner and a date against it or rests on whoever in the room feels able to speak.