What Interruptions Cost an Engineering Team, and Where They Come From

Interruption cost is systematically underestimated because the visible part is the small part. Somebody asks a question, you answer in two minutes, and the two minutes is what both of you record. What neither of you records is the fifteen or twenty minutes it takes to rebuild the mental state you were holding before the question arrived.

"Every 'quick question' is 20 minutes."

The cost is the recovery, not the interruption

Engineering work involves holding a lot of state that exists nowhere but in your head: which invariant you are relying on, which three cases you have already ruled out, what you were about to check next. None of that is written down, because writing it down would take longer than the work.

An interruption drops the state. Answering takes two minutes; reconstructing takes considerably longer, and sometimes the reconstruction is imperfect, which is where a certain class of bug comes from. Microsoft's research on workplace interruptions puts the average worker's day at around 275 interruptions, which is roughly one every two minutes of a working day. For work that requires sustained state, that is not a productivity problem, it is a structural one.

The practical consequence is the pattern every engineering manager recognises: the real work happens before 09:00, after 18:00, or on the day someone takes as focus time. Those are the hours when the state survives long enough to be useful.

Why the load is so uneven

Interruptions flow toward whoever answers fastest and knows most. That is a compliment and a tax, and it compounds: the more reliably someone answers, the more they get asked, and the more of the team's knowledge appears to live with them, which justifies asking them next time.

So on a team of twelve, two or three people absorb most of the volume. They are usually the senior engineers, which means the team's most expensive capacity is spent on retrieval rather than on the work only they can do. And because each individual instance is helpful and brief, nobody experiences it as a problem until that person's output visibly drops or they resign. The dynamic is covered in reducing dependency on the one person everyone asks.

Where interruptions come from

Source	Typical share	Can it be removed?
Question with an answer that already exists	The majority	Yes, by making the answer reachable
Question needing judgement	Smaller	No, but it can be batched and routed
Genuine emergency	Rare	No, and it should not be
Notifications and channel noise	High volume, low cost each	Yes, individually configurable

The important row is the first. Most interruptions are questions with an existing answer, which means most interruption reduction is a retrieval problem rather than a politeness problem. Asking people to interrupt less does not work, because the interruption was rational: they needed something, and a person was the fastest route to it.

How to measure it in one week

You do not need instrumentation. Ask three or four of the most-interrupted people to keep a tally for five days, one line per interruption: who, what they wanted, and whether the answer already existed somewhere.

Two numbers come out of that. The count per person per day, which shows how uneven the load is, and the share where the answer already existed, which is the size of the fixable portion. Both are more persuasive than any argument, and the second number is usually high enough to surprise the people who were doing the answering.

Removing the largest category

For the questions that already have answers, the requirement is strict: the alternative to asking a person has to be faster than asking a person. Asking a colleague takes about eleven seconds. A wiki search that takes two minutes and might fail will lose every time, which is why documentation never displaces the direct message.

This is what StandIn is for. Each person confirms a ninety-second brief at the end of their day, mostly pre-drafted from the work they already did: current state, open questions, blockers, next actions. While they are off, and that includes a focus block or a meeting, their StandIn answers questions from that brief in their words, with a source under every answer, clearly labelled, and never guessing. When the answer is not there, it says so and names who to ask, so the judgement questions still reach a human, at a moment that can be batched.

The senior engineer's day stops being a support desk. See StandIn for leadership teams.

Common Questions

How much does an interruption really cost?

Far more than its duration, because the expensive part is rebuilding the mental state that was dropped. For work requiring sustained concentration, a two-minute question commonly costs fifteen to twenty minutes in total, which is why counting interruption minutes understates the problem so badly.

Is it better to batch interruptions?

Yes, substantially. Four questions answered in one block cost one recovery rather than four. Publishing a daily window when you answer things is a real improvement, though it moves the cost onto the asker, which only works if their question can wait.

Should we ask people to interrupt less?

On its own it does not work, because the interruption was the rational choice given the options available. Change the options: make the answer reachable without a person, and the behaviour changes without anyone being asked to be more considerate.

Which interruptions should we protect?

Genuine emergencies and real judgement calls, both of which are legitimate and should reach a person quickly. The goal is not fewer interruptions in the abstract, it is that the ones that land on people are the ones that needed a person.