Nobody Designed the Apprenticeship

Of everything in this literature, one finding deserves to end more arguments than it does. Emanuel, Harrington and Pallais studied software engineers at a Fortune 500 firm and were able to compare engineers who sat near their teammates with engineers who did not. Sitting near teammates raised the amount of online feedback junior engineers received on their code by 18.3 percent.

Note the direction of the effect and who it lands on. It is not that proximate engineers wrote more code. It is that proximate juniors were taught more.

And the paper's title names the trade honestly. Training for tomorrow, or productivity today. The feedback came from somewhere, and where it came from was a senior engineer's afternoon.

The second finding, which is about careers rather than code

Emanuel and Harrington studied a large call center through the period when its offices closed, which let them separate two things that are normally tangled together. Some of the gap between remote and on-site workers was selection, meaning that different people chose remote work. Some of it was treatment, meaning the effect of the arrangement itself.

Both were present. About a third of the initial gap was a genuine treatment effect. And the treatment effect was concentrated in two places worth naming. Remote work degraded call quality particularly for inexperienced workers, and it reduced workers' promotion rates.

Put the two studies together and a shape emerges that neither camp talks about in public. The people paying for the arrangement are early in their careers. The people deciding on it are not.

Almost everyone senior enough to set this policy learned their craft by proximity, and almost none of them will personally pay the cost of removing it.

Why it resists being rescheduled

The instinctive fix is a mentoring program. Pair people up, book a recurring thirty minutes, track attendance. Organizations do this, attendance is usually excellent, and the thing that was lost usually stays lost.

Nonaka's account of how organizations create knowledge explains why. He separates tacit from explicit knowledge and identifies four modes of conversion between them, one of which is socialization, the movement of tacit knowledge directly from person to person through shared experience. Socialization is the mode that does not run on documents, slides or agendas. It runs on watching somebody competent handle something they have not seen before.

A scheduled call is an excellent vehicle for explicit knowledge. It is a poor vehicle for judgment, because judgment shows up in the unplanned moment, and the unplanned moment does not know your calendar.

There is a second obstacle and it is about safety rather than logistics. Edmondson established that learning behavior in teams, meaning asking questions, seeking feedback, admitting error and experimenting, depends on psychological safety, and that learning behavior is what carries safety through to performance. Asking a question is an interpersonal risk. Asking it quietly to the person beside you is a small risk. Asking it with your face on a grid in front of eleven people, some of whom are recording, is a larger one.

Bartel, Wrzesniewski and Wiesenfeld found the compounding factor. More physically isolated employees perceived less respect within the organization, and that perception explained their lower identification with it. A person who is unsure of their standing asks fewer questions, which is precisely the behavior the development mechanism depends on.

The counterexample, because it is a real one

Choudhury, Foroughi and Larson studied patent examiners at the United States Patent and Trademark Office when the agency let them work from anywhere rather than merely from home, and found that output rose. That is a serious result from a serious setting, and it is not in tension with anything above.

Patent examination is individual, heavily codified work performed by people who have already been trained. The apprenticeship has already happened. Remove the constraint of location from an expert doing well-specified work and you get more of it, which is also what the broader meta-analytic picture suggested long before any of this was controversial. Gajendran and Harrison pooled 46 studies covering nearly 13,000 employees in 2007 and found telecommuting had small but mainly beneficial effects on outcomes such as perceived autonomy and lower work-family conflict.

So the honest statement is narrower and more useful than either slogan. Distance is close to free for an expert executing codified work, and expensive for a novice acquiring judgment. An organization that has both, which is every organization, cannot have one policy.

What building it deliberately actually looks like

The unfair part of this is that the office did it for free and doing it on purpose costs real money and real senior time. That is not an argument against doing it. It is an argument for admitting what the office was buying, which almost nobody defending the office has actually named.

• Measure feedback received, not sessions held. The study measured feedback on work product. That is the observable. A program with perfect attendance and no feedback flowing has reproduced the calendar entry and not the mechanism.
• Put the unplanned moment on purpose. Working sessions where a junior person watches a senior person handle something live beat any number of scheduled discussions about how one would handle it. Review a real decision as it is being made, not afterwards as a story.
• Give somebody the outcome. The apprenticeship failed because it belonged to nobody. Assign it the way a release is assigned, with a name attached and a report on how it is going.
• Concentrate proximity where it earns the most. If two days in the office is the budget, the research says spend them on the people who are learning rather than on the people who are executing. Most schedules currently do the opposite, because they are set by function.
• Lower the cost of asking. Small groups, a named person whose job includes answering, and visible senior people asking their own questions in public. Edmondson's finding is that the behavior follows the safety, so the lever is the safety.

The reason this one is different

Attention, alignment and authority all degrade in ways somebody notices within a quarter. A meeting load is visible. A missed handoff is visible. A decision stuck in an approval chain generates complaints.

A cohort of people who did not become as good as they could have generates nothing. No incident, no ticket, no line on a dashboard. It arrives five years later as a shortage of people ready for the next role, and by then nobody connects it to a seating policy.

Which is exactly why it needs to be a decision somebody makes rather than a thing that happens. It was one before, too. Nobody made it, and it worked anyway, and that is the only reason this is difficult now.