Interruption cost rises with how deep the task has gone
Aliases: resumption lag · engagement-dependent cost · interruption tax
What it is
Interruption cost is not the flash of annoyance when something rings. It is the time and error paid to reconstruct a goal after working memory has been displaced. The deeper the engagement—writing an argument, reconciling figures, coordinating a live step with someone else—the more the same banner costs. At shallow engagement (a glance at the home screen after unlock, waiting for an elevator) there is almost no goal to restore.
This entry is only about how cost tracks the work in progress. It does not assign an urgency tier to the incoming item, and it does not explain how false alarms later numb people to true ones.
Why it happens
While a goal is being executed, "what is next, how far along, which intermediate results are still needed" sits in working memory. A notification pulls attention; those representations decay. Coming back requires rereading the screen and recovering intent. That gap is resumption lag, often longer than looking at the notice. Deeper engagement packs denser, more interdependent representations, so reconstruction is dearer. Errors change too: a shallow interrupt is an extra glance; a deep one cuts away mid-step and leaves half-work and omissions.
So there is no stable answer to "is this notice annoying?" The answer is bound to task state at arrival. Senders cannot see that state and still assume idle.
Studying it
The classic method inserts a secondary task halfway through a primary one and records interruption lag (time to put the primary down) and resumption lag, plus post-resumption errors. Interface versions use a document being edited, a form being filled, or a video being watched as the primary task.
Independent variables: distance from a subgoal boundary, working-memory load of the primary task, whether the notice demands an immediate reply. Dependent variables: resumption lag, errors after return, discarded intermediate results, subjective interruptedness.
Lab insertions are scheduled and people know they will return; real notices often lead into another app, and the primary task may never be picked up. "Looked at the notice" is not the cost—the glance can take a second, rebuilding the goal tens of seconds. Ecological validity depends on the primary task actually having depth; feed-scrolling will not show an engagement effect.
Where it stops holding
Waiting, queues, and idle lock screens sit near zero engagement; interruption cost collapses, and "protect flow" over-constrains delivery. Driving, surgery, and control rooms combine extreme engagement with irreversible error; consumer thresholds for an acceptable banner do not apply. Some roles treat arrival as the job (support, dispatch); for them delay is the cost, and the engagement model flips. Short, auto-dismissing, non-focus-stealing cues have a flatter cost curve than modal dialogs.
Applying it
- During typing, full-screen, navigation, or a call, defer nonessential banners rather than flushing the send queue immediately.
- Treat "deep task" as a delivery-side signal: keyboard focus, unsaved draft, foreground media or navigation count as high engagement.
- If delivery cannot wait in that window, go silent into the shade instead of sound and screen-takeover, so pickup happens on the person's own cadence.
- Verify by inserting the same notice mid writing or form-fill and measuring time to the first useful action after return; if the high-engagement condition is substantially longer than idle home screen, the channel is not yielding to engagement.